Research ArticleCoal-Rock Recognition in Top Coal Caving Using Bimodal DeepLearning and Hilbert-Huang Transform

Guoxin Zhang12 Zengcai Wang12 Lei Zhao12 Yazhou Qi12 and JinshanWang12

1School of Mechanical Engineering Shandong University No 17923 Jingshi Road Jinan 250061 China2Key Laboratory of High Efficiency and Clean Mechanical Manufacture Shandong University Ministry of EducationNo 17923 Jingshi Road Jinan 250061 China

Correspondence should be addressed to Zengcai Wang wangzcsdueducn

Received 5 April 2017 Revised 11 June 2017 Accepted 19 June 2017 Published 27 July 2017

Academic Editor Matteo Filippi

Copyright copy 2017 Guoxin Zhang et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

This study employs the mechanical vibration and acoustic waves of a hydraulic support tail beam for an accurate and fast coal-rockrecognition The study proposes a diagnosis method based on bimodal deep learning and Hilbert-Huang transform The bimodaldeep neural networks (DNN) adopt bimodal learning and transfer learning The bimodal learning method attempts to learn jointrepresentation by considering acceleration and sound pressure modalities which both contribute to coal-rock recognition Thetransfer learning method solves the problem regarding DNN in which a large number of labeled training samples are necessary tooptimize the parameters while the labeled training sample is limited A suitable installation location for sensors is determined inrecognizing coal-rock The extraction features of acceleration and sound pressure signals are combined and effective combinationfeatures are selected Bimodal DNN consists of two deep belief networks (DBN) each DBN model is trained with related samplesand the parameters of the pretrained DBNs are transferred to the final recognition model Then the parameters of the proposedmodel are continuously optimized by pretraining and fine-tuning Finally the comparison of experimental results demonstratesthe superiority of the proposed method in terms of recognition accuracy

1 Introduction

Coal is an important source of energy accounting forapproximately 2921 of primary energy consumption of theworld in 2015 according to BP Statistical Review of WorldEnergy (June 2016) China both produces and consumeslarge amount of coal accounting for 4770 and 5001 ofglobal coal production and consumption in the past yearrespectively Approximately 1284 of coal reserves in theworld are distributed in China of which 448 are thick coalseam Therefore safe and efficient mining thick coal seamis considerably important Fully mechanized top coal cavinghas been widely applied in the mining of thick coal seam dueto its safety high efficiency high yield and low productioncost

However low-level automation and intelligence havealways been problems in fully mechanized technology ontop coal caving Particularly one of the key technologiesof caving degrees completely relies on peoplersquos judgment

Relying on artificial vision and auditory in determining thedegree of caving is prone to over- and less caving causedby harsh environment including poor light coal dust noiseand narrow space Overcaving and less caving can lead tolow recovery rate decline in coal quality and increase incost In addition the safety and health of operators areoften threatened because they are relatively close to coal-falling areas Therefore an accurate and rapid approach ofidentifying coal-rock is considerably important to controlthe coal-falling time precisely and improve the automationand intelligence of a fully mechanized top coal cavingThese objectives are important in ameliorating the workingenvironment of coal miners improving coal recovery rateand reducing production cost

Since 1960s more than 20 types of coal-rock interfaceidentification methods and coal-rock recognition meth-ods have been proposed by researchers The representativecoal-rock interface identification methods include detectionthrough artificial 120574-ray natural 120574-ray [1ndash3] and radar [4]The

HindawiShock and VibrationVolume 2017 Article ID 3809525 13 pageshttpsdoiorg10115520173809525

2 Shock and Vibration

artificial 120574-ray detection method uses 120574-ray backscatteringto detect the thickness of the top coal and identify coal-rock interface However artificial 120574-ray is harmful to humanbeings its penetration ability is limited and it has beengradually abandoned Natural 120574-ray detection measures theintensity of gamma ray in the roof passing through theremaining coal seam and determines the thickness of thetop seam according to its attention law to identify coal-rock interface Natural 120574-ray detection has been relativelymature for commercialized coal-rock interface recognitionHowever when the top coal contains no or low radioactiveelements or contains excessive gangue natural 120574-ray detec-tion is no longer applicable In addition the cost of thisdetection is high Radar detection utilizes the reflection of theelectromagnetic wave at the interface of coal gangue to detectthe thickness of the top coal Moreover the advantages ofradar detection include far resolution distance and extensiveapplication range However radar detection is no longerapplicable when the thickness of the top coal is extremelythick

The coal-rock identification method recognize coal androck to determine whether it reached the coal-rock inter-face The representative methods include detection throughcutting force response [5] image [6 7] vibration acoustic[8ndash15] and infrared [16] Cutting force response detectionusing drum picks will have different performance cuttingcoal or rock to identify coal and rock This method hasgood adaptability but it is inapplicable to top coal cavingImage detection uses coal and gangue with different colorshardness gloss and other information to identify coal androck However this method is sensitive to dust and lightand needs further research Infrared detection uses a cuttingmachine to produce different temperatures when it cuts tocoal and rock and determine whether the cut is coal or rockThis method has a quick response and is ideal for real-timeapplication However this detection is still immature becauseof its sensitivity to an ambient temperature

In recent years the coal-rock recognition method basedon vibration technology has been widely used Proposedby the US Mining Bureau the vibration method mainlycomprises acoustic method slot wave seismic method andmechanical vibration method Its principle is coal andgangue which have different frequencies In [10] the vibra-tion of till beam was analyzed to recognize coal and rock inthe top coal caving process In [11 12] vibration signals wereused to determine whether cutting machine is cutting coal orrock In [13] the cutting acoustic signal is used to recognizethe cutting pattern In [15] acoustic signals of coal and rock infully mechanized face were analyzed to rock-coal recognitionin top coal caving

Although significant progress has been made on coal-rock recognition in the past few decades challenges stillremain Majority of the previous works based on vibrationtechnology independently treated the acoustic or mechanicalvibration methods wherein only the acoustic or mechanicalvibration is employed for recognition Each single modalityhas been demonstrated to be useful for recognition Howeverone single modality alone cannot provide sufficient infor-mation on the differences between coal and rock Therefore

the method of integrating two modalities to improve therecognition accuracy of rock and coal still needs furtherinvestigation

Several algorithms were proposed to address the repre-sentation learning for multiple modalities In [17] video andaudio modality inputs were employed to learn bimodal deepbelief networks (DBN) In [18] multimodal deep neural net-works (DNN)were proposed to study the correlation betweentexture and landmark modalities for facial expression reor-ganization wherein several stacked autoencoders (AE) wereused In [19] bimodal DNN were used to determine driverfatigue expression In this study acoustic and mechanicalvibrations are integrated for the first time and feature statis-tics and classification are assembled together for coal-rockrecognition Acceleration and sound pressure sensors areemployed to detect the vibration and acoustic wave signalsA joint representation layer for recognition is learned fromthe acoustic and mechanical vibration modalities

In addition transfer learning method is adopted becausetransferring knowledge of a general object from classificationto recognition task has been found to be successful invisually similar categories [20] In the present study bimodaltransfer DNN (BT-DNN) is initially pretrained by simulatingsamples (coal and rock hitting the tail beam in the laboratoryenvironment) and irrelevant samples (hand knocking thetable)and solving the problem in which massive samplesare necessary to train DNN Then the proposed model ispretrained by real training samples Last supervised fine-tuning is performed With bimodal learning and transferlearning processes the proposed method not only has highrecognition accuracy but also has the advantages of low costextensive adaptability insensitivity to the environment andsimplified technical difficulty

Another important step in the recognition is extractingeffective features frommeasured signalsThemost commonlyapplied time-frequency analysis methods include Fouriertransform (FT) [21ndash23] wavelet transform (WT) [24ndash27]and Hilbert-Huang transform (HHT) [28ndash31] methodsThese methods have their own advantages and applicationareas In this work HHT is utilized to process signalsIn addition five statistics characteristics are computedwhich include four characteristics based on intrinsic modefunctions (IMFs) and one characteristic based on Hilbertmarginal energy spectrum However a few extracted featuresmay not provide several contributions for recognition Bycontrast combining different features as input yields anefficient combination of features

The rest of this paper is organized as follows Section 2introduces the used methods Section 3 describes the recog-nition system of coal-rock in top coal caving based onthe proposed method including signal acquisition featuresextraction and experiments Section 4 summarizes this paperand proposes future work

2 Methods

21 Deep Belief Network A deep belief network (DBN) isa neural network that contains several layers of restrictedBoltzmann machine (RBM) in which the input layer of the

Shock and Vibration 3

Visible layer

Hiddenlayer

middot middot middot

middot middot middot

b1 b2b3 b4

bnhℎ1 ℎ2 ℎ3 ℎ4 ℎnh

a1 a2 a3 anv1 2 3 nv

Figure 1 Schematic of RBM

next RBM is the output layer of the previous RBM RBMwas proposed by Smolensky in 1986 [32] it is a probabilisticgraphical model that can be explained by stochastic neuralnetwork RBM is a binary graph in which visible unitsare connected to hidden units however no connectionsexist between visible-visible units or hidden-hidden unitsThe visible layer represents observations and the hiddenlayer learns features RBM is applied to numerous machinelearningmethods due to its desirable properties In particularafter Hinton et al proposed DBN based on RBM as a basiccomponent [33] RBM has been successfully implementedin dimensionality reduction [34] feature learning [35] andclassification [36]The schematic of RBM is shown in Figure 1

RBM is an energy basedmodel and its energy function isas follows

119864120579 (V ℎ) = minus 119899Vsum119894=1119886119894V119894 minus 119899ℎsum

119895=1119887119895V119895 minus 119899Vsum

119894=1

119899ℎsum119895=1ℎ119895119908119894119895V119894 (1)

where 119899V and 119899ℎ denote the number of neuron units in thevisible and hidden layers respectively V119894 and ℎ119895 denote thestates of the 119894th neuron in the visible layer and the 119895th neuronin the hidden layer respectively 119886119894 and 119887119895 denote the bias ofthe 119894th neuron in the visible layer and the 119895th neuron in thehidden layer respectively 119908119894119895 is the weight associated withthe connection between units 119894 and 119895 in the hidden and visiblelayers respectively

Given the independence of the activations of hidden andvisible units the individual activation probabilities are asfollows

119875 (ℎ119895 = 1 | V) = 120590(119887119895 + 119899Vsum119894=1119908119894119895V119894)

119875 (V119894 = 1 | ℎ) = 120590(119886119894 + 119899ℎsum119895=1119908119894119895ℎ119895)

(2)

In this study Gaussian-Bernoulli RBM was used and itsenergy of joint configuration is

119864 (V120579 ℎ) = 119899Vsum119894=1

(V119894 minus 119887119894)22 minus 119899Vsum119894=1

119899ℎsum119895=1119908119894119895V119894ℎ119895 minus 119899ℎsum

119895=1119886119895ℎ119895 (3)

Meanwhile the conditional distribution is

119901120579 (V119894 | ℎ) = 119873(119887119894 + 119899ℎsum119895=1119908119894119895ℎ119895 1) (4)

Apart from the preceding differences Gaussian-BernoulliRBM is the same as binary RBM

DBN was pretrained layer by layer by RBM and back-propagation (BP) algorithm was adopted to fine-tune theentire network for the optimization of all network parame-ters

22 Bimodal Learning Multimodal learning was proposedby Ngiam et al in 2011 [17] to learn features over multiplemodalities (image and audio modalities)The authors provedthat if multiple modalities were present at feature learningthen one modality can be better learned Since the idea ofmultimodal learning has been proposed many researchershave applied this idea and achieved good results [18 19]

As shown in Figure 2 the proposed bimodal coal-rockrecognition system in this study included bimodal DBNs ajoint representation layer and an output layer Each DBN hadbimodal RBMs and the architecture of each DBN was deter-mined after testing 100 different architectures The details onthe determination of eachDBN and joint representation layerarchitecture are provided in Section 331 and Section 332respectively

The training process of this model could be performed infour steps

First two DBNs were trained for acceleration and soundpressure respectively

Second two DBNs and a joint representation layer werecombined as a joint RBM which was trained afterward

Third the two DBNs joint RBM and the output layerwere combined to form a DNN The DNN was subjected togreedily layer-wise training The training process was still anunsupervised training which is also called pretraining

Fourth this network was fine-tuned to strengthen therecognition capability this process was a supervised training

23 Transfer Learning Deep networks usually need a largenumber of training samples to optimize their parametersThe limited labeled sample is the weakness of deep learningTo overcome this difficulty transfer learning is employed totransfer knowledge from related data In transfer learningknowledge obtained from different but related works withsufficient samples is used Moreover this method has alreadyachieved some success in the identification field [17 34]

In this study whether this transfer learning property ofDNN could be generalized to coal-rock was explored Firstsamples of simulating coal and rock hitting the tail beamwere used to pretrain the DNN model Second samples of

4 Shock and Vibration

RBM RBM

RBM RBM

RBM Classifier

DBN1

DBN2InputJoint

representation Output

Acceleration

Soundpressure

Coal

Rock

Vibration signal30

20

10

0

minus10

minus20

minus30Acce

lera

tion

(mＭ

2)

0 02 04 06 08 1

Time (s)

Sound signal543210

minus1minus2minus3minus4So

und

pres

sure

(Pa)

0 05 1 15 2 25 3

Time (s)

Figure 2 Architecture of bimodal learning

hand knocking the table were used to pretrain the DNNmodel Because these samples are also acceleration and soundpressure which belong to the same category with trainingsamples they played a pretraining effect and optimized theparameters In the two preceding processes the bimodallearningmethod is used After the unsupervised training theparameters of DNN are transferred to BT-DNN (Figure 3)

24 Hilbert-Huang Transform At present the most com-monly applied time-frequency analysis methods are FT WTand Hilbert-Huang transform (HHT) HHT comprises twoparts ensemble empirical mode decomposition (EEMD) andHilbert spectral analysis (HSA) EEMDwas proposed in 2004[37] Different from EMD EEMD solves the mode mixingproblem by adding a certain amount of Gaussian white noiseto the original signal each time before decomposing [38]Moreover EEMD has better performance in nonlinear andnonstationary signals than FT andWT which are extensivelyapplied in many industries EEMD was selected in thisstudy due to the strong background noise of the cavingenvironment and the nonlinear and nonstationary measuredsignals By using EEMD several IMFs were obtained

119909 (119905) = 119899sum119894=1119888119894 (119905) + 119903119899 (119905) (5)

where 119909(119905) is the original signal 119888119894(119905) is 119894th IMF and 119903119899(119905) isthe trend item

The obtained IMFs were transformed by Hilbert andHilbert energy spectra were obtained

119911119894 (119905) = 119888119894 (119905) + 119895119867 (119888119894 (119905)) = 119886119894 (119905) 119890119895120579119894(119905) (6)

where 119867(119888119894(119905)) is the Hilbert transform of 119888119894(119905) 119886119894(119905) is theamplitude function of 119911119894(119905) and 120579119894(119905) is the phase functionof 119911119894(119905)

Hilbert time-frequency spectra of the signal can beobtained as

119867(120596 119905) = Re119899sum119894=1119886119894 (119905) 119890119895 int120596119894(119905)119889119905 (7)

where 120596119894(119905) = 119889120579119894(119905)119889119905Finally Hilbert marginal energy spectra were obtained

119864 (120596) = int11987901198672 (119908 119905) 119889119905 (8)

3 Recognition System for the Coal and Rock

Data acquisition feature extraction and state recognitioncombined the intelligent coal-rock recognition system asshown in Figure 4 Data acquisition comprised the exper-imental system design selected sensors installed sensorsand signal acquisition Feature extraction comprised signalprocessing statistics feature computation and selection ofeffective features combination State recognition consistedof model building parameter optimization and patternrecognition

Shock and Vibration 5

DBN1

DBN2

Transfer model BT-DNN

DBN1

DBN2

Simulating samples

Acceleration

Soundpressure

W b c

Figure 3 The architecture of transfer learning

Training samples(pretraining + fine-tuning)

BT-DNN

Recognize

Testingsamples

Coal or rock

Parameters optimize

State Recognition

Data acquisitionfront end

EEMD

Tail boomIMF components

Data acquisition Feature extraction

Select effective featurescombination

HSA

Energy kurtosisand so forth

Data preprocessing

Acceleration signals

Soundpressure signals

Sensors

Simulating samples(transfer training)

Vibration signal30

20

10

0

minus10

minus20

minus30

0 01 02 03 05 07 0904 06 08 1

Time (s)

Sound signal

05 1 15 2 25 3

Time (s)

AccelerationSensor

Sound PressureSensor

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

Sample point

0050

minus005signa

l

0010

minus001IMF1

IMF2

IMF3

IMF4

IMF5

IMF6

IMF7

IMF8

IMF9

0010

minus001

50

minus5

50

minus5

50

minus5

20

minus2

20

minus2

10

minus1

20

minus2

times10

times10

times10

times10

times10

times10

times10

Acce

lera

tion

(mＭ

)

Figure 4 The proposed recognition system

31 Signal Acquisition Top coal caving working site and theself-designed experimental system are shown in Figures 5(a)and 5(b) respectively The acceleration and sound pressurevaried when coal and rock hit the tail beam supported by ahydraulic supporter Thus identifying coal-rock by detectingthe tail beam acceleration and sound pressure is feasible Thesensors were installed on the back of the tail beam due tothe following reasons First themain impact locations of coaland rock falling are the tail chute and beam However if thesensors are installed on the tail chute then they will be easily

buried by falling coals and gangues In addition sensors maybe damaged by the continuous spraying of water which isused for dustproofingMoreover the running conveyor chutegenerates numerous noise that will add to the difficulty ofdata processing and analysis Therefore the ideal location ofsensors is at the back of the tail beam Two sensors wereemployed acceleration and sound pressure sensors theirinstallation method was magnetic base Sensors and specificinstallation location are shown in Figures 6(a) and 6(b)respectively Their parameters are shown in Table 1

6 Shock and Vibration

SensorsHydraulicsupporter

Rock seam

Coal seam

(a)

Tail beam

(b)

Figure 5 (a) Top coal caving working site (b) self-designed experimental system

(a)

Accelerationsensor

Sound pressuresensor

(b)

Figure 6 (a) Acceleration sensor and sound pressure sensor (b) specific installation location

Table 1 Parameters of acceleration sensors

Parameters Range Inherent noiseAcceleration sensor minus71 g to 71 g 2 120583gradicHzSound pressure sensor 165 to 134 dB (dynamic) 165 dB

Data acquisition system is composed of data acquisitionfront end wireless router and software system The datacollection process is as follows The acceleration and soundpressure of the tail beam were detected by sensors obtainedby the data acquisition front end and then transmitted to anotebook by a wireless router The data were obtained in No1306 working platform in Xinglongzhuang Coal Mine Thethickness of coal is 734m to 890m The caving method isone knife in one caving the distance of the caving step is08m The frequency of the sampling is set at 65536Hz thetime of each pattern sampling is 52 s The sampling time ofeach sample is determined in Section 336 The signals areshown in Figure 7

32 Feature Extraction Feature extraction is important inpattern recognition because selecting effective features sig-nificantly contributes to the improvement of recognitionaccuracy By using EEMD several IMFs were obtained asshown in Figure 8

The selection of statistics is significantly important forfeature extraction In this study five statistics characteristicswere computed The four characteristics based on IMFsincluded skewness kurtosis energy and variance and onecharacteristic based on Hilbert marginal energy spectrum isthe energy statistics of frequency divisionThe characteristicswere calculated as follows

(a) Energy based on IMFs is

119890119894 = int1198790

1003816100381610038161003816119909119894 (119905)10038161003816100381610038162 119889119905 (9)

(b) Kurtosis based on IMFs is

Kurtosis = 119864 (119909 minus 120583)41205904 (10)

(c) Skewness based on IMFs is

skewness = 119864 (119909 minus 120583)31205903 (11)

Shock and Vibration 7

0

0

10

20

30Vibration signal

minus10

minus20

minus30

Acce

lera

tion

(mＭ

2)

01 02 03 04 05 06 07 08 09 1

Time (s)

(a)

0

2

4

6

Time (s)

Acoustic signal

Soun

d pr

essu

re (P

a)

minus2

minus4

0 05 1 15 2 25 3

(b)

Figure 7 Measured signals (a) acceleration (b) sound pressure

0 10 20 30 40 50 60

0

1IM

F5

minus1

0 10 20 30 40 50 60

0

05

IMF6

minus05

0 10 20 30 40 50 60

0

05

IMF8

minus05

0 10 20 30 40 50 60

0

05

Time (ms)

RES

minus05

0 10 20 30 40 50 60

0

02

IMF7

minus02

0 10 20 30 40 50 60

0

20

Sign

al

minus20

0 10 20 30 40 50 60

0

20

IMF1

minus20

0 10 20 30 40 50 60

0

10

IMF2

minus10

0 10 20 30 40 50 60

0

2

IMF3

minus2

0 10 20 30 40 50 60

0

1

Time (ms)

IMF4

minus1

Figure 8 IMFs decomposed from measured signal

(d) Variance based on IMFs is

1205902 = 1119873 minus 1119873sum119894=1

1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (12)

(e) Segmented energy ratio based onHilbert marginal energyspectrum the frequency was divided into 10 bands and

the ratio of each frequency band to the total energy wascalculated

119864119873119894 = 119864119894sum10119895=1 119864119895 (13)

When using acceleration as the algorithmrsquos only inputwe chose 119894 statistics from the 5 statistics as feature input 119894varied from 1 to 5 After sum5119894=1 1198621198945 = 31 experiments the mosteffective combination of statistics for acceleration input is

8 Shock and Vibration

Table 2 Recognition rates with different combinations feature

Number of acceleration statistics Number of sound statistics0 1 2 3 4 5

0 6657 7927 8237 8173 81971 6963 6410 6800 7217 7590 55432 7817 7240 7477 7377 7233 68873 8890 9657 8093 8907 8230 81734 9073 9003 9793 8940 9243 66235 8330 9913 8803 9283 9213 7447

determined It is found that when the number of statisticsis set to 4 corresponding to kurtosis energy variance andenergy statistics of frequency division the recognition ratereached the maximum value (9073) as shown in Table 2In the same way we determined the feature when the soundpressure is the algorithmrsquos only input The correspondingcombination of statistics is kurtosis variance and energystatistics of frequency division (8237)

When the acceleration and sound pressure are set as inputtogether we chose 119894 statistics from5 acceleration statistics and119895 statistics from 5 sound pressure statistics as feature inputthe total number of combinations is sum5119894=1 1198621198945 times sum5119894=1 1198621198945 =961 After 961 experiments it was found that when thenumber of acceleration statistics is set to 5 and the numberof sound pressure statistics is set to 1 corresponding to 5all statistics of acceleration and skewness of sound pressurethe recognition rate reached the maximum value (9913)And other comparison experiments use all 5 statistics ofacceleration and skewness of sound pressure as feature input

33 Experiments

331 Comparison among Algorithms with Different DBNStructures The network structure which includes the unitnumber of the visible and hidden layers has a profound influ-ence on network performance A total of 100 combinationswere tested for each DBN to seek the optimal structure Thefirst and second layers of acceleration and sound pressureDBNs both varied from 50 to 500 at intervals of 50The resultof acceleration DBN is shown in Figure 9 Good networkperformance was observed when the first hidden layer wasset from 50 to 150 and the second hidden layer was set atmore than 150 The structure of acceleration DBN was set to[42 100 250] to build the best recognition deep networkThestructure of sound pressure DBN could also be determined inthe samemanner with its optical structure set to [8 150 100]332 Comparison among Algorithms with Different Numbersof Joint Representative Units The number of joint represen-tative RBM units is considerably important for the proposedmodel performance which belongs to the hyperparameter ofthe network The number of joint representative units variedfrom 25 to 1000 to obtain a better performance results shownin Figure 10

It shows that the performance gradually and rapidlybecomes better initially then stabilizes and finally decreases

because of the increasing number of joint representativeunits This result could be attributed to having more unitsresulting in more parameters for learning more informationregarding inputs However excessive parameters could not belearned because the number of training samples was limitedIf the number of units was too small then learning theinformation would be difficult The network performed bestwhen the unit number was set to 150

333 Comparison of the Algorithms Using Unimodality andBimodality One of the key ideas of this study is integratingacceleration and sound pressure for coal-rock recognitionThus algorithms with unimodality and bimodality werecompared Figure 11 is the results of comparison of algorithmsusing acceleration and sound pressure respectively It showsthat the method using acceleration (9073) has betterperformance thanusing soundpressure (8237) but inferiorto the combination of two modalities (9913)

Another issue is that the modality number may affectthe performance of network algorithms with different inputmethods were compared One input method was treatingacceleration and sound pressure as one vector input fed intothe recognition network The recognition result is shown asthe green bar in Figure 12The other input method was treat-ing acceleration and sound pressure features separately as twomodalities fed into the recognition network The recognitionresult is shown as the yellow bar Separately treating twomodalities evidently helped the algorithm obtain a betterperformance (363) than combining features together

334 Comparison among Algorithms with and without RBMsPretraining plays an important role in the recognition ofcoal and rock Figure 13 shows the comparison of algorithmswith and without pretraining The comparison shows thatthe model with unsupervised pretrained RBMs (9913)has 483 improvement compared to the model withoutpretraining (9430) This result is due to the BP algorithmthat used gradient descent to gain the optima and pretrainingwhich allowed the network to be fine-tuned at a relativelygood initial value rather than a few random initial pointsthereby potentially avoiding the risk of falling into poor localoptima Therefore pretraining is significantly necessary formodel training

335 Comparison among Algorithms with and without Trans-fer Learning Transfer learning plays the same role with

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 3

Visible layer

Hiddenlayer

middot middot middot

middot middot middot

b1 b2b3 b4

bnhℎ1 ℎ2 ℎ3 ℎ4 ℎnh

a1 a2 a3 anv1 2 3 nv

Figure 1 Schematic of RBM

next RBM is the output layer of the previous RBM RBMwas proposed by Smolensky in 1986 [32] it is a probabilisticgraphical model that can be explained by stochastic neuralnetwork RBM is a binary graph in which visible unitsare connected to hidden units however no connectionsexist between visible-visible units or hidden-hidden unitsThe visible layer represents observations and the hiddenlayer learns features RBM is applied to numerous machinelearningmethods due to its desirable properties In particularafter Hinton et al proposed DBN based on RBM as a basiccomponent [33] RBM has been successfully implementedin dimensionality reduction [34] feature learning [35] andclassification [36]The schematic of RBM is shown in Figure 1

RBM is an energy basedmodel and its energy function isas follows

119864120579 (V ℎ) = minus 119899Vsum119894=1119886119894V119894 minus 119899ℎsum

119895=1119887119895V119895 minus 119899Vsum

119894=1

119899ℎsum119895=1ℎ119895119908119894119895V119894 (1)

where 119899V and 119899ℎ denote the number of neuron units in thevisible and hidden layers respectively V119894 and ℎ119895 denote thestates of the 119894th neuron in the visible layer and the 119895th neuronin the hidden layer respectively 119886119894 and 119887119895 denote the bias ofthe 119894th neuron in the visible layer and the 119895th neuron in thehidden layer respectively 119908119894119895 is the weight associated withthe connection between units 119894 and 119895 in the hidden and visiblelayers respectively

Given the independence of the activations of hidden andvisible units the individual activation probabilities are asfollows

119875 (ℎ119895 = 1 | V) = 120590(119887119895 + 119899Vsum119894=1119908119894119895V119894)

119875 (V119894 = 1 | ℎ) = 120590(119886119894 + 119899ℎsum119895=1119908119894119895ℎ119895)

(2)

In this study Gaussian-Bernoulli RBM was used and itsenergy of joint configuration is

119864 (V120579 ℎ) = 119899Vsum119894=1

(V119894 minus 119887119894)22 minus 119899Vsum119894=1

119899ℎsum119895=1119908119894119895V119894ℎ119895 minus 119899ℎsum

119895=1119886119895ℎ119895 (3)

Meanwhile the conditional distribution is

119901120579 (V119894 | ℎ) = 119873(119887119894 + 119899ℎsum119895=1119908119894119895ℎ119895 1) (4)

Apart from the preceding differences Gaussian-BernoulliRBM is the same as binary RBM

DBN was pretrained layer by layer by RBM and back-propagation (BP) algorithm was adopted to fine-tune theentire network for the optimization of all network parame-ters

22 Bimodal Learning Multimodal learning was proposedby Ngiam et al in 2011 [17] to learn features over multiplemodalities (image and audio modalities)The authors provedthat if multiple modalities were present at feature learningthen one modality can be better learned Since the idea ofmultimodal learning has been proposed many researchershave applied this idea and achieved good results [18 19]

As shown in Figure 2 the proposed bimodal coal-rockrecognition system in this study included bimodal DBNs ajoint representation layer and an output layer Each DBN hadbimodal RBMs and the architecture of each DBN was deter-mined after testing 100 different architectures The details onthe determination of eachDBN and joint representation layerarchitecture are provided in Section 331 and Section 332respectively

The training process of this model could be performed infour steps

First two DBNs were trained for acceleration and soundpressure respectively

Second two DBNs and a joint representation layer werecombined as a joint RBM which was trained afterward

Third the two DBNs joint RBM and the output layerwere combined to form a DNN The DNN was subjected togreedily layer-wise training The training process was still anunsupervised training which is also called pretraining

Fourth this network was fine-tuned to strengthen therecognition capability this process was a supervised training

23 Transfer Learning Deep networks usually need a largenumber of training samples to optimize their parametersThe limited labeled sample is the weakness of deep learningTo overcome this difficulty transfer learning is employed totransfer knowledge from related data In transfer learningknowledge obtained from different but related works withsufficient samples is used Moreover this method has alreadyachieved some success in the identification field [17 34]

In this study whether this transfer learning property ofDNN could be generalized to coal-rock was explored Firstsamples of simulating coal and rock hitting the tail beamwere used to pretrain the DNN model Second samples of

4 Shock and Vibration

RBM RBM

RBM RBM

RBM Classifier

DBN1

DBN2InputJoint

representation Output

Acceleration

Soundpressure

Coal

Rock

Vibration signal30

20

10

0

minus10

minus20

minus30Acce

lera

tion

(mＭ

2)

0 02 04 06 08 1

Time (s)

Sound signal543210

minus1minus2minus3minus4So

und

pres

sure

(Pa)

0 05 1 15 2 25 3

Time (s)

Figure 2 Architecture of bimodal learning

hand knocking the table were used to pretrain the DNNmodel Because these samples are also acceleration and soundpressure which belong to the same category with trainingsamples they played a pretraining effect and optimized theparameters In the two preceding processes the bimodallearningmethod is used After the unsupervised training theparameters of DNN are transferred to BT-DNN (Figure 3)

24 Hilbert-Huang Transform At present the most com-monly applied time-frequency analysis methods are FT WTand Hilbert-Huang transform (HHT) HHT comprises twoparts ensemble empirical mode decomposition (EEMD) andHilbert spectral analysis (HSA) EEMDwas proposed in 2004[37] Different from EMD EEMD solves the mode mixingproblem by adding a certain amount of Gaussian white noiseto the original signal each time before decomposing [38]Moreover EEMD has better performance in nonlinear andnonstationary signals than FT andWT which are extensivelyapplied in many industries EEMD was selected in thisstudy due to the strong background noise of the cavingenvironment and the nonlinear and nonstationary measuredsignals By using EEMD several IMFs were obtained

119909 (119905) = 119899sum119894=1119888119894 (119905) + 119903119899 (119905) (5)

where 119909(119905) is the original signal 119888119894(119905) is 119894th IMF and 119903119899(119905) isthe trend item

The obtained IMFs were transformed by Hilbert andHilbert energy spectra were obtained

119911119894 (119905) = 119888119894 (119905) + 119895119867 (119888119894 (119905)) = 119886119894 (119905) 119890119895120579119894(119905) (6)

where 119867(119888119894(119905)) is the Hilbert transform of 119888119894(119905) 119886119894(119905) is theamplitude function of 119911119894(119905) and 120579119894(119905) is the phase functionof 119911119894(119905)

Hilbert time-frequency spectra of the signal can beobtained as

119867(120596 119905) = Re119899sum119894=1119886119894 (119905) 119890119895 int120596119894(119905)119889119905 (7)

where 120596119894(119905) = 119889120579119894(119905)119889119905Finally Hilbert marginal energy spectra were obtained

119864 (120596) = int11987901198672 (119908 119905) 119889119905 (8)

3 Recognition System for the Coal and Rock

Data acquisition feature extraction and state recognitioncombined the intelligent coal-rock recognition system asshown in Figure 4 Data acquisition comprised the exper-imental system design selected sensors installed sensorsand signal acquisition Feature extraction comprised signalprocessing statistics feature computation and selection ofeffective features combination State recognition consistedof model building parameter optimization and patternrecognition

Shock and Vibration 5

DBN1

DBN2

Transfer model BT-DNN

DBN1

DBN2

Simulating samples

Acceleration

Soundpressure

W b c

Figure 3 The architecture of transfer learning

Training samples(pretraining + fine-tuning)

BT-DNN

Recognize

Testingsamples

Coal or rock

Parameters optimize

State Recognition

Data acquisitionfront end

EEMD

Tail boomIMF components

Data acquisition Feature extraction

Select effective featurescombination

HSA

Energy kurtosisand so forth

Data preprocessing

Acceleration signals

Soundpressure signals

Sensors

Simulating samples(transfer training)

Vibration signal30

20

10

0

minus10

minus20

minus30

0 01 02 03 05 07 0904 06 08 1

Time (s)

Sound signal

05 1 15 2 25 3

Time (s)

AccelerationSensor

Sound PressureSensor

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

Sample point

0050

minus005signa

l

0010

minus001IMF1

IMF2

IMF3

IMF4

IMF5

IMF6

IMF7

IMF8

IMF9

0010

minus001

50

minus5

50

minus5

50

minus5

20

minus2

20

minus2

10

minus1

20

minus2

times10

times10

times10

times10

times10

times10

times10

Acce

lera

tion

(mＭ

)

Figure 4 The proposed recognition system

31 Signal Acquisition Top coal caving working site and theself-designed experimental system are shown in Figures 5(a)and 5(b) respectively The acceleration and sound pressurevaried when coal and rock hit the tail beam supported by ahydraulic supporter Thus identifying coal-rock by detectingthe tail beam acceleration and sound pressure is feasible Thesensors were installed on the back of the tail beam due tothe following reasons First themain impact locations of coaland rock falling are the tail chute and beam However if thesensors are installed on the tail chute then they will be easily

buried by falling coals and gangues In addition sensors maybe damaged by the continuous spraying of water which isused for dustproofingMoreover the running conveyor chutegenerates numerous noise that will add to the difficulty ofdata processing and analysis Therefore the ideal location ofsensors is at the back of the tail beam Two sensors wereemployed acceleration and sound pressure sensors theirinstallation method was magnetic base Sensors and specificinstallation location are shown in Figures 6(a) and 6(b)respectively Their parameters are shown in Table 1

6 Shock and Vibration

SensorsHydraulicsupporter

Rock seam

Coal seam

(a)

Tail beam

(b)

Figure 5 (a) Top coal caving working site (b) self-designed experimental system

(a)

Accelerationsensor

Sound pressuresensor

(b)

Figure 6 (a) Acceleration sensor and sound pressure sensor (b) specific installation location

Table 1 Parameters of acceleration sensors

Parameters Range Inherent noiseAcceleration sensor minus71 g to 71 g 2 120583gradicHzSound pressure sensor 165 to 134 dB (dynamic) 165 dB

Data acquisition system is composed of data acquisitionfront end wireless router and software system The datacollection process is as follows The acceleration and soundpressure of the tail beam were detected by sensors obtainedby the data acquisition front end and then transmitted to anotebook by a wireless router The data were obtained in No1306 working platform in Xinglongzhuang Coal Mine Thethickness of coal is 734m to 890m The caving method isone knife in one caving the distance of the caving step is08m The frequency of the sampling is set at 65536Hz thetime of each pattern sampling is 52 s The sampling time ofeach sample is determined in Section 336 The signals areshown in Figure 7

32 Feature Extraction Feature extraction is important inpattern recognition because selecting effective features sig-nificantly contributes to the improvement of recognitionaccuracy By using EEMD several IMFs were obtained asshown in Figure 8

The selection of statistics is significantly important forfeature extraction In this study five statistics characteristicswere computed The four characteristics based on IMFsincluded skewness kurtosis energy and variance and onecharacteristic based on Hilbert marginal energy spectrum isthe energy statistics of frequency divisionThe characteristicswere calculated as follows

(a) Energy based on IMFs is

119890119894 = int1198790

1003816100381610038161003816119909119894 (119905)10038161003816100381610038162 119889119905 (9)

(b) Kurtosis based on IMFs is

Kurtosis = 119864 (119909 minus 120583)41205904 (10)

(c) Skewness based on IMFs is

skewness = 119864 (119909 minus 120583)31205903 (11)

Shock and Vibration 7

0

0

10

20

30Vibration signal

minus10

minus20

minus30

Acce

lera

tion

(mＭ

2)

01 02 03 04 05 06 07 08 09 1

Time (s)

(a)

0

2

4

6

Time (s)

Acoustic signal

Soun

d pr

essu

re (P

a)

minus2

minus4

0 05 1 15 2 25 3

(b)

Figure 7 Measured signals (a) acceleration (b) sound pressure

0 10 20 30 40 50 60

0

1IM

F5

minus1

0 10 20 30 40 50 60

0

05

IMF6

minus05

0 10 20 30 40 50 60

0

05

IMF8

minus05

0 10 20 30 40 50 60

0

05

Time (ms)

RES

minus05

0 10 20 30 40 50 60

0

02

IMF7

minus02

0 10 20 30 40 50 60

0

20

Sign

al

minus20

0 10 20 30 40 50 60

0

20

IMF1

minus20

0 10 20 30 40 50 60

0

10

IMF2

minus10

0 10 20 30 40 50 60

0

2

IMF3

minus2

0 10 20 30 40 50 60

0

1

Time (ms)

IMF4

minus1

Figure 8 IMFs decomposed from measured signal

(d) Variance based on IMFs is

1205902 = 1119873 minus 1119873sum119894=1

1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (12)

(e) Segmented energy ratio based onHilbert marginal energyspectrum the frequency was divided into 10 bands and

the ratio of each frequency band to the total energy wascalculated

119864119873119894 = 119864119894sum10119895=1 119864119895 (13)

When using acceleration as the algorithmrsquos only inputwe chose 119894 statistics from the 5 statistics as feature input 119894varied from 1 to 5 After sum5119894=1 1198621198945 = 31 experiments the mosteffective combination of statistics for acceleration input is

8 Shock and Vibration

Table 2 Recognition rates with different combinations feature

Number of acceleration statistics Number of sound statistics0 1 2 3 4 5

0 6657 7927 8237 8173 81971 6963 6410 6800 7217 7590 55432 7817 7240 7477 7377 7233 68873 8890 9657 8093 8907 8230 81734 9073 9003 9793 8940 9243 66235 8330 9913 8803 9283 9213 7447

determined It is found that when the number of statisticsis set to 4 corresponding to kurtosis energy variance andenergy statistics of frequency division the recognition ratereached the maximum value (9073) as shown in Table 2In the same way we determined the feature when the soundpressure is the algorithmrsquos only input The correspondingcombination of statistics is kurtosis variance and energystatistics of frequency division (8237)

When the acceleration and sound pressure are set as inputtogether we chose 119894 statistics from5 acceleration statistics and119895 statistics from 5 sound pressure statistics as feature inputthe total number of combinations is sum5119894=1 1198621198945 times sum5119894=1 1198621198945 =961 After 961 experiments it was found that when thenumber of acceleration statistics is set to 5 and the numberof sound pressure statistics is set to 1 corresponding to 5all statistics of acceleration and skewness of sound pressurethe recognition rate reached the maximum value (9913)And other comparison experiments use all 5 statistics ofacceleration and skewness of sound pressure as feature input

33 Experiments

331 Comparison among Algorithms with Different DBNStructures The network structure which includes the unitnumber of the visible and hidden layers has a profound influ-ence on network performance A total of 100 combinationswere tested for each DBN to seek the optimal structure Thefirst and second layers of acceleration and sound pressureDBNs both varied from 50 to 500 at intervals of 50The resultof acceleration DBN is shown in Figure 9 Good networkperformance was observed when the first hidden layer wasset from 50 to 150 and the second hidden layer was set atmore than 150 The structure of acceleration DBN was set to[42 100 250] to build the best recognition deep networkThestructure of sound pressure DBN could also be determined inthe samemanner with its optical structure set to [8 150 100]332 Comparison among Algorithms with Different Numbersof Joint Representative Units The number of joint represen-tative RBM units is considerably important for the proposedmodel performance which belongs to the hyperparameter ofthe network The number of joint representative units variedfrom 25 to 1000 to obtain a better performance results shownin Figure 10

It shows that the performance gradually and rapidlybecomes better initially then stabilizes and finally decreases

because of the increasing number of joint representativeunits This result could be attributed to having more unitsresulting in more parameters for learning more informationregarding inputs However excessive parameters could not belearned because the number of training samples was limitedIf the number of units was too small then learning theinformation would be difficult The network performed bestwhen the unit number was set to 150

333 Comparison of the Algorithms Using Unimodality andBimodality One of the key ideas of this study is integratingacceleration and sound pressure for coal-rock recognitionThus algorithms with unimodality and bimodality werecompared Figure 11 is the results of comparison of algorithmsusing acceleration and sound pressure respectively It showsthat the method using acceleration (9073) has betterperformance thanusing soundpressure (8237) but inferiorto the combination of two modalities (9913)

Another issue is that the modality number may affectthe performance of network algorithms with different inputmethods were compared One input method was treatingacceleration and sound pressure as one vector input fed intothe recognition network The recognition result is shown asthe green bar in Figure 12The other input method was treat-ing acceleration and sound pressure features separately as twomodalities fed into the recognition network The recognitionresult is shown as the yellow bar Separately treating twomodalities evidently helped the algorithm obtain a betterperformance (363) than combining features together

334 Comparison among Algorithms with and without RBMsPretraining plays an important role in the recognition ofcoal and rock Figure 13 shows the comparison of algorithmswith and without pretraining The comparison shows thatthe model with unsupervised pretrained RBMs (9913)has 483 improvement compared to the model withoutpretraining (9430) This result is due to the BP algorithmthat used gradient descent to gain the optima and pretrainingwhich allowed the network to be fine-tuned at a relativelygood initial value rather than a few random initial pointsthereby potentially avoiding the risk of falling into poor localoptima Therefore pretraining is significantly necessary formodel training

335 Comparison among Algorithms with and without Trans-fer Learning Transfer learning plays the same role with

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

4 Shock and Vibration

RBM RBM

RBM RBM

RBM Classifier

DBN1

DBN2InputJoint

representation Output

Acceleration

Soundpressure

Coal

Rock

Vibration signal30

20

10

0

minus10

minus20

minus30Acce

lera

tion

(mＭ

2)

0 02 04 06 08 1

Time (s)

Sound signal543210

minus1minus2minus3minus4So

und

pres

sure

(Pa)

0 05 1 15 2 25 3

Time (s)

Figure 2 Architecture of bimodal learning

hand knocking the table were used to pretrain the DNNmodel Because these samples are also acceleration and soundpressure which belong to the same category with trainingsamples they played a pretraining effect and optimized theparameters In the two preceding processes the bimodallearningmethod is used After the unsupervised training theparameters of DNN are transferred to BT-DNN (Figure 3)

24 Hilbert-Huang Transform At present the most com-monly applied time-frequency analysis methods are FT WTand Hilbert-Huang transform (HHT) HHT comprises twoparts ensemble empirical mode decomposition (EEMD) andHilbert spectral analysis (HSA) EEMDwas proposed in 2004[37] Different from EMD EEMD solves the mode mixingproblem by adding a certain amount of Gaussian white noiseto the original signal each time before decomposing [38]Moreover EEMD has better performance in nonlinear andnonstationary signals than FT andWT which are extensivelyapplied in many industries EEMD was selected in thisstudy due to the strong background noise of the cavingenvironment and the nonlinear and nonstationary measuredsignals By using EEMD several IMFs were obtained

119909 (119905) = 119899sum119894=1119888119894 (119905) + 119903119899 (119905) (5)

where 119909(119905) is the original signal 119888119894(119905) is 119894th IMF and 119903119899(119905) isthe trend item

The obtained IMFs were transformed by Hilbert andHilbert energy spectra were obtained

119911119894 (119905) = 119888119894 (119905) + 119895119867 (119888119894 (119905)) = 119886119894 (119905) 119890119895120579119894(119905) (6)

where 119867(119888119894(119905)) is the Hilbert transform of 119888119894(119905) 119886119894(119905) is theamplitude function of 119911119894(119905) and 120579119894(119905) is the phase functionof 119911119894(119905)

Hilbert time-frequency spectra of the signal can beobtained as

119867(120596 119905) = Re119899sum119894=1119886119894 (119905) 119890119895 int120596119894(119905)119889119905 (7)

where 120596119894(119905) = 119889120579119894(119905)119889119905Finally Hilbert marginal energy spectra were obtained

119864 (120596) = int11987901198672 (119908 119905) 119889119905 (8)

3 Recognition System for the Coal and Rock

Data acquisition feature extraction and state recognitioncombined the intelligent coal-rock recognition system asshown in Figure 4 Data acquisition comprised the exper-imental system design selected sensors installed sensorsand signal acquisition Feature extraction comprised signalprocessing statistics feature computation and selection ofeffective features combination State recognition consistedof model building parameter optimization and patternrecognition

Shock and Vibration 5

DBN1

DBN2

Transfer model BT-DNN

DBN1

DBN2

Simulating samples

Acceleration

Soundpressure

W b c

Figure 3 The architecture of transfer learning

Training samples(pretraining + fine-tuning)

BT-DNN

Recognize

Testingsamples

Coal or rock

Parameters optimize

State Recognition

Data acquisitionfront end

EEMD

Tail boomIMF components

Data acquisition Feature extraction

Select effective featurescombination

HSA

Energy kurtosisand so forth

Data preprocessing

Acceleration signals

Soundpressure signals

Sensors

Simulating samples(transfer training)

Vibration signal30

20

10

0

minus10

minus20

minus30

0 01 02 03 05 07 0904 06 08 1

Time (s)

Sound signal

05 1 15 2 25 3

Time (s)

AccelerationSensor

Sound PressureSensor

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

Sample point

0050

minus005signa

l

0010

minus001IMF1

IMF2

IMF3

IMF4

IMF5

IMF6

IMF7

IMF8

IMF9

0010

minus001

50

minus5

50

minus5

50

minus5

20

minus2

20

minus2

10

minus1

20

minus2

times10

times10

times10

times10

times10

times10

times10

Acce

lera

tion

(mＭ

)

Figure 4 The proposed recognition system

31 Signal Acquisition Top coal caving working site and theself-designed experimental system are shown in Figures 5(a)and 5(b) respectively The acceleration and sound pressurevaried when coal and rock hit the tail beam supported by ahydraulic supporter Thus identifying coal-rock by detectingthe tail beam acceleration and sound pressure is feasible Thesensors were installed on the back of the tail beam due tothe following reasons First themain impact locations of coaland rock falling are the tail chute and beam However if thesensors are installed on the tail chute then they will be easily

buried by falling coals and gangues In addition sensors maybe damaged by the continuous spraying of water which isused for dustproofingMoreover the running conveyor chutegenerates numerous noise that will add to the difficulty ofdata processing and analysis Therefore the ideal location ofsensors is at the back of the tail beam Two sensors wereemployed acceleration and sound pressure sensors theirinstallation method was magnetic base Sensors and specificinstallation location are shown in Figures 6(a) and 6(b)respectively Their parameters are shown in Table 1

6 Shock and Vibration

SensorsHydraulicsupporter

Rock seam

Coal seam

(a)

Tail beam

(b)

Figure 5 (a) Top coal caving working site (b) self-designed experimental system

(a)

Accelerationsensor

Sound pressuresensor

(b)

Figure 6 (a) Acceleration sensor and sound pressure sensor (b) specific installation location

Table 1 Parameters of acceleration sensors

Parameters Range Inherent noiseAcceleration sensor minus71 g to 71 g 2 120583gradicHzSound pressure sensor 165 to 134 dB (dynamic) 165 dB

Data acquisition system is composed of data acquisitionfront end wireless router and software system The datacollection process is as follows The acceleration and soundpressure of the tail beam were detected by sensors obtainedby the data acquisition front end and then transmitted to anotebook by a wireless router The data were obtained in No1306 working platform in Xinglongzhuang Coal Mine Thethickness of coal is 734m to 890m The caving method isone knife in one caving the distance of the caving step is08m The frequency of the sampling is set at 65536Hz thetime of each pattern sampling is 52 s The sampling time ofeach sample is determined in Section 336 The signals areshown in Figure 7

32 Feature Extraction Feature extraction is important inpattern recognition because selecting effective features sig-nificantly contributes to the improvement of recognitionaccuracy By using EEMD several IMFs were obtained asshown in Figure 8

The selection of statistics is significantly important forfeature extraction In this study five statistics characteristicswere computed The four characteristics based on IMFsincluded skewness kurtosis energy and variance and onecharacteristic based on Hilbert marginal energy spectrum isthe energy statistics of frequency divisionThe characteristicswere calculated as follows

(a) Energy based on IMFs is

119890119894 = int1198790

1003816100381610038161003816119909119894 (119905)10038161003816100381610038162 119889119905 (9)

(b) Kurtosis based on IMFs is

Kurtosis = 119864 (119909 minus 120583)41205904 (10)

(c) Skewness based on IMFs is

skewness = 119864 (119909 minus 120583)31205903 (11)

Shock and Vibration 7

0

0

10

20

30Vibration signal

minus10

minus20

minus30

Acce

lera

tion

(mＭ

2)

01 02 03 04 05 06 07 08 09 1

Time (s)

(a)

0

2

4

6

Time (s)

Acoustic signal

Soun

d pr

essu

re (P

a)

minus2

minus4

0 05 1 15 2 25 3

(b)

Figure 7 Measured signals (a) acceleration (b) sound pressure

0 10 20 30 40 50 60

0

1IM

F5

minus1

0 10 20 30 40 50 60

0

05

IMF6

minus05

0 10 20 30 40 50 60

0

05

IMF8

minus05

0 10 20 30 40 50 60

0

05

Time (ms)

RES

minus05

0 10 20 30 40 50 60

0

02

IMF7

minus02

0 10 20 30 40 50 60

0

20

Sign

al

minus20

0 10 20 30 40 50 60

0

20

IMF1

minus20

0 10 20 30 40 50 60

0

10

IMF2

minus10

0 10 20 30 40 50 60

0

2

IMF3

minus2

0 10 20 30 40 50 60

0

1

Time (ms)

IMF4

minus1

Figure 8 IMFs decomposed from measured signal

(d) Variance based on IMFs is

1205902 = 1119873 minus 1119873sum119894=1

1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (12)

(e) Segmented energy ratio based onHilbert marginal energyspectrum the frequency was divided into 10 bands and

the ratio of each frequency band to the total energy wascalculated

119864119873119894 = 119864119894sum10119895=1 119864119895 (13)

When using acceleration as the algorithmrsquos only inputwe chose 119894 statistics from the 5 statistics as feature input 119894varied from 1 to 5 After sum5119894=1 1198621198945 = 31 experiments the mosteffective combination of statistics for acceleration input is

8 Shock and Vibration

Table 2 Recognition rates with different combinations feature

Number of acceleration statistics Number of sound statistics0 1 2 3 4 5

0 6657 7927 8237 8173 81971 6963 6410 6800 7217 7590 55432 7817 7240 7477 7377 7233 68873 8890 9657 8093 8907 8230 81734 9073 9003 9793 8940 9243 66235 8330 9913 8803 9283 9213 7447

determined It is found that when the number of statisticsis set to 4 corresponding to kurtosis energy variance andenergy statistics of frequency division the recognition ratereached the maximum value (9073) as shown in Table 2In the same way we determined the feature when the soundpressure is the algorithmrsquos only input The correspondingcombination of statistics is kurtosis variance and energystatistics of frequency division (8237)

When the acceleration and sound pressure are set as inputtogether we chose 119894 statistics from5 acceleration statistics and119895 statistics from 5 sound pressure statistics as feature inputthe total number of combinations is sum5119894=1 1198621198945 times sum5119894=1 1198621198945 =961 After 961 experiments it was found that when thenumber of acceleration statistics is set to 5 and the numberof sound pressure statistics is set to 1 corresponding to 5all statistics of acceleration and skewness of sound pressurethe recognition rate reached the maximum value (9913)And other comparison experiments use all 5 statistics ofacceleration and skewness of sound pressure as feature input

33 Experiments

331 Comparison among Algorithms with Different DBNStructures The network structure which includes the unitnumber of the visible and hidden layers has a profound influ-ence on network performance A total of 100 combinationswere tested for each DBN to seek the optimal structure Thefirst and second layers of acceleration and sound pressureDBNs both varied from 50 to 500 at intervals of 50The resultof acceleration DBN is shown in Figure 9 Good networkperformance was observed when the first hidden layer wasset from 50 to 150 and the second hidden layer was set atmore than 150 The structure of acceleration DBN was set to[42 100 250] to build the best recognition deep networkThestructure of sound pressure DBN could also be determined inthe samemanner with its optical structure set to [8 150 100]332 Comparison among Algorithms with Different Numbersof Joint Representative Units The number of joint represen-tative RBM units is considerably important for the proposedmodel performance which belongs to the hyperparameter ofthe network The number of joint representative units variedfrom 25 to 1000 to obtain a better performance results shownin Figure 10

It shows that the performance gradually and rapidlybecomes better initially then stabilizes and finally decreases

because of the increasing number of joint representativeunits This result could be attributed to having more unitsresulting in more parameters for learning more informationregarding inputs However excessive parameters could not belearned because the number of training samples was limitedIf the number of units was too small then learning theinformation would be difficult The network performed bestwhen the unit number was set to 150

333 Comparison of the Algorithms Using Unimodality andBimodality One of the key ideas of this study is integratingacceleration and sound pressure for coal-rock recognitionThus algorithms with unimodality and bimodality werecompared Figure 11 is the results of comparison of algorithmsusing acceleration and sound pressure respectively It showsthat the method using acceleration (9073) has betterperformance thanusing soundpressure (8237) but inferiorto the combination of two modalities (9913)

Another issue is that the modality number may affectthe performance of network algorithms with different inputmethods were compared One input method was treatingacceleration and sound pressure as one vector input fed intothe recognition network The recognition result is shown asthe green bar in Figure 12The other input method was treat-ing acceleration and sound pressure features separately as twomodalities fed into the recognition network The recognitionresult is shown as the yellow bar Separately treating twomodalities evidently helped the algorithm obtain a betterperformance (363) than combining features together

334 Comparison among Algorithms with and without RBMsPretraining plays an important role in the recognition ofcoal and rock Figure 13 shows the comparison of algorithmswith and without pretraining The comparison shows thatthe model with unsupervised pretrained RBMs (9913)has 483 improvement compared to the model withoutpretraining (9430) This result is due to the BP algorithmthat used gradient descent to gain the optima and pretrainingwhich allowed the network to be fine-tuned at a relativelygood initial value rather than a few random initial pointsthereby potentially avoiding the risk of falling into poor localoptima Therefore pretraining is significantly necessary formodel training

335 Comparison among Algorithms with and without Trans-fer Learning Transfer learning plays the same role with

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 5

DBN1

DBN2

Transfer model BT-DNN

DBN1

DBN2

Simulating samples

Acceleration

Soundpressure

W b c

Figure 3 The architecture of transfer learning

Training samples(pretraining + fine-tuning)

BT-DNN

Recognize

Testingsamples

Coal or rock

Parameters optimize

State Recognition

Data acquisitionfront end

EEMD

Tail boomIMF components

Data acquisition Feature extraction

Select effective featurescombination

HSA

Energy kurtosisand so forth

Data preprocessing

Acceleration signals

Soundpressure signals

Sensors

Simulating samples(transfer training)

Vibration signal30

20

10

0

minus10

minus20

minus30

0 01 02 03 05 07 0904 06 08 1

Time (s)

Sound signal

05 1 15 2 25 3

Time (s)

AccelerationSensor

Sound PressureSensor

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

0 1000 2000 3000 4000 5000 6000 7000 8000

Sample point

0050

minus005signa

l

0010

minus001IMF1

IMF2

IMF3

IMF4

IMF5

IMF6

IMF7

IMF8

IMF9

0010

minus001

50

minus5

50

minus5

50

minus5

20

minus2

20

minus2

10

minus1

20

minus2

times10

times10

times10

times10

times10

times10

times10

Acce

lera

tion

(mＭ

)

Figure 4 The proposed recognition system

31 Signal Acquisition Top coal caving working site and theself-designed experimental system are shown in Figures 5(a)and 5(b) respectively The acceleration and sound pressurevaried when coal and rock hit the tail beam supported by ahydraulic supporter Thus identifying coal-rock by detectingthe tail beam acceleration and sound pressure is feasible Thesensors were installed on the back of the tail beam due tothe following reasons First themain impact locations of coaland rock falling are the tail chute and beam However if thesensors are installed on the tail chute then they will be easily

buried by falling coals and gangues In addition sensors maybe damaged by the continuous spraying of water which isused for dustproofingMoreover the running conveyor chutegenerates numerous noise that will add to the difficulty ofdata processing and analysis Therefore the ideal location ofsensors is at the back of the tail beam Two sensors wereemployed acceleration and sound pressure sensors theirinstallation method was magnetic base Sensors and specificinstallation location are shown in Figures 6(a) and 6(b)respectively Their parameters are shown in Table 1

6 Shock and Vibration

SensorsHydraulicsupporter

Rock seam

Coal seam

(a)

Tail beam

(b)

Figure 5 (a) Top coal caving working site (b) self-designed experimental system

(a)

Accelerationsensor

Sound pressuresensor

(b)

Figure 6 (a) Acceleration sensor and sound pressure sensor (b) specific installation location

Table 1 Parameters of acceleration sensors

Parameters Range Inherent noiseAcceleration sensor minus71 g to 71 g 2 120583gradicHzSound pressure sensor 165 to 134 dB (dynamic) 165 dB

Data acquisition system is composed of data acquisitionfront end wireless router and software system The datacollection process is as follows The acceleration and soundpressure of the tail beam were detected by sensors obtainedby the data acquisition front end and then transmitted to anotebook by a wireless router The data were obtained in No1306 working platform in Xinglongzhuang Coal Mine Thethickness of coal is 734m to 890m The caving method isone knife in one caving the distance of the caving step is08m The frequency of the sampling is set at 65536Hz thetime of each pattern sampling is 52 s The sampling time ofeach sample is determined in Section 336 The signals areshown in Figure 7

32 Feature Extraction Feature extraction is important inpattern recognition because selecting effective features sig-nificantly contributes to the improvement of recognitionaccuracy By using EEMD several IMFs were obtained asshown in Figure 8

The selection of statistics is significantly important forfeature extraction In this study five statistics characteristicswere computed The four characteristics based on IMFsincluded skewness kurtosis energy and variance and onecharacteristic based on Hilbert marginal energy spectrum isthe energy statistics of frequency divisionThe characteristicswere calculated as follows

(a) Energy based on IMFs is

119890119894 = int1198790

1003816100381610038161003816119909119894 (119905)10038161003816100381610038162 119889119905 (9)

(b) Kurtosis based on IMFs is

Kurtosis = 119864 (119909 minus 120583)41205904 (10)

(c) Skewness based on IMFs is

skewness = 119864 (119909 minus 120583)31205903 (11)

Shock and Vibration 7

0

0

10

20

30Vibration signal

minus10

minus20

minus30

Acce

lera

tion

(mＭ

2)

01 02 03 04 05 06 07 08 09 1

Time (s)

(a)

0

2

4

6

Time (s)

Acoustic signal

Soun

d pr

essu

re (P

a)

minus2

minus4

0 05 1 15 2 25 3

(b)

Figure 7 Measured signals (a) acceleration (b) sound pressure

0 10 20 30 40 50 60

0

1IM

F5

minus1

0 10 20 30 40 50 60

0

05

IMF6

minus05

0 10 20 30 40 50 60

0

05

IMF8

minus05

0 10 20 30 40 50 60

0

05

Time (ms)

RES

minus05

0 10 20 30 40 50 60

0

02

IMF7

minus02

0 10 20 30 40 50 60

0

20

Sign

al

minus20

0 10 20 30 40 50 60

0

20

IMF1

minus20

0 10 20 30 40 50 60

0

10

IMF2

minus10

0 10 20 30 40 50 60

0

2

IMF3

minus2

0 10 20 30 40 50 60

0

1

Time (ms)

IMF4

minus1

Figure 8 IMFs decomposed from measured signal

(d) Variance based on IMFs is

1205902 = 1119873 minus 1119873sum119894=1

1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (12)

(e) Segmented energy ratio based onHilbert marginal energyspectrum the frequency was divided into 10 bands and

the ratio of each frequency band to the total energy wascalculated

119864119873119894 = 119864119894sum10119895=1 119864119895 (13)

When using acceleration as the algorithmrsquos only inputwe chose 119894 statistics from the 5 statistics as feature input 119894varied from 1 to 5 After sum5119894=1 1198621198945 = 31 experiments the mosteffective combination of statistics for acceleration input is

8 Shock and Vibration

Table 2 Recognition rates with different combinations feature

Number of acceleration statistics Number of sound statistics0 1 2 3 4 5

0 6657 7927 8237 8173 81971 6963 6410 6800 7217 7590 55432 7817 7240 7477 7377 7233 68873 8890 9657 8093 8907 8230 81734 9073 9003 9793 8940 9243 66235 8330 9913 8803 9283 9213 7447

determined It is found that when the number of statisticsis set to 4 corresponding to kurtosis energy variance andenergy statistics of frequency division the recognition ratereached the maximum value (9073) as shown in Table 2In the same way we determined the feature when the soundpressure is the algorithmrsquos only input The correspondingcombination of statistics is kurtosis variance and energystatistics of frequency division (8237)

When the acceleration and sound pressure are set as inputtogether we chose 119894 statistics from5 acceleration statistics and119895 statistics from 5 sound pressure statistics as feature inputthe total number of combinations is sum5119894=1 1198621198945 times sum5119894=1 1198621198945 =961 After 961 experiments it was found that when thenumber of acceleration statistics is set to 5 and the numberof sound pressure statistics is set to 1 corresponding to 5all statistics of acceleration and skewness of sound pressurethe recognition rate reached the maximum value (9913)And other comparison experiments use all 5 statistics ofacceleration and skewness of sound pressure as feature input

33 Experiments

331 Comparison among Algorithms with Different DBNStructures The network structure which includes the unitnumber of the visible and hidden layers has a profound influ-ence on network performance A total of 100 combinationswere tested for each DBN to seek the optimal structure Thefirst and second layers of acceleration and sound pressureDBNs both varied from 50 to 500 at intervals of 50The resultof acceleration DBN is shown in Figure 9 Good networkperformance was observed when the first hidden layer wasset from 50 to 150 and the second hidden layer was set atmore than 150 The structure of acceleration DBN was set to[42 100 250] to build the best recognition deep networkThestructure of sound pressure DBN could also be determined inthe samemanner with its optical structure set to [8 150 100]332 Comparison among Algorithms with Different Numbersof Joint Representative Units The number of joint represen-tative RBM units is considerably important for the proposedmodel performance which belongs to the hyperparameter ofthe network The number of joint representative units variedfrom 25 to 1000 to obtain a better performance results shownin Figure 10

It shows that the performance gradually and rapidlybecomes better initially then stabilizes and finally decreases

because of the increasing number of joint representativeunits This result could be attributed to having more unitsresulting in more parameters for learning more informationregarding inputs However excessive parameters could not belearned because the number of training samples was limitedIf the number of units was too small then learning theinformation would be difficult The network performed bestwhen the unit number was set to 150

333 Comparison of the Algorithms Using Unimodality andBimodality One of the key ideas of this study is integratingacceleration and sound pressure for coal-rock recognitionThus algorithms with unimodality and bimodality werecompared Figure 11 is the results of comparison of algorithmsusing acceleration and sound pressure respectively It showsthat the method using acceleration (9073) has betterperformance thanusing soundpressure (8237) but inferiorto the combination of two modalities (9913)

Another issue is that the modality number may affectthe performance of network algorithms with different inputmethods were compared One input method was treatingacceleration and sound pressure as one vector input fed intothe recognition network The recognition result is shown asthe green bar in Figure 12The other input method was treat-ing acceleration and sound pressure features separately as twomodalities fed into the recognition network The recognitionresult is shown as the yellow bar Separately treating twomodalities evidently helped the algorithm obtain a betterperformance (363) than combining features together

334 Comparison among Algorithms with and without RBMsPretraining plays an important role in the recognition ofcoal and rock Figure 13 shows the comparison of algorithmswith and without pretraining The comparison shows thatthe model with unsupervised pretrained RBMs (9913)has 483 improvement compared to the model withoutpretraining (9430) This result is due to the BP algorithmthat used gradient descent to gain the optima and pretrainingwhich allowed the network to be fine-tuned at a relativelygood initial value rather than a few random initial pointsthereby potentially avoiding the risk of falling into poor localoptima Therefore pretraining is significantly necessary formodel training

335 Comparison among Algorithms with and without Trans-fer Learning Transfer learning plays the same role with

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

6 Shock and Vibration

SensorsHydraulicsupporter

Rock seam

Coal seam

(a)

Tail beam

(b)

Figure 5 (a) Top coal caving working site (b) self-designed experimental system

(a)

Accelerationsensor

Sound pressuresensor

(b)

Figure 6 (a) Acceleration sensor and sound pressure sensor (b) specific installation location

Table 1 Parameters of acceleration sensors

Parameters Range Inherent noiseAcceleration sensor minus71 g to 71 g 2 120583gradicHzSound pressure sensor 165 to 134 dB (dynamic) 165 dB

Data acquisition system is composed of data acquisitionfront end wireless router and software system The datacollection process is as follows The acceleration and soundpressure of the tail beam were detected by sensors obtainedby the data acquisition front end and then transmitted to anotebook by a wireless router The data were obtained in No1306 working platform in Xinglongzhuang Coal Mine Thethickness of coal is 734m to 890m The caving method isone knife in one caving the distance of the caving step is08m The frequency of the sampling is set at 65536Hz thetime of each pattern sampling is 52 s The sampling time ofeach sample is determined in Section 336 The signals areshown in Figure 7

32 Feature Extraction Feature extraction is important inpattern recognition because selecting effective features sig-nificantly contributes to the improvement of recognitionaccuracy By using EEMD several IMFs were obtained asshown in Figure 8

The selection of statistics is significantly important forfeature extraction In this study five statistics characteristicswere computed The four characteristics based on IMFsincluded skewness kurtosis energy and variance and onecharacteristic based on Hilbert marginal energy spectrum isthe energy statistics of frequency divisionThe characteristicswere calculated as follows

(a) Energy based on IMFs is

119890119894 = int1198790

1003816100381610038161003816119909119894 (119905)10038161003816100381610038162 119889119905 (9)

(b) Kurtosis based on IMFs is

Kurtosis = 119864 (119909 minus 120583)41205904 (10)

(c) Skewness based on IMFs is

skewness = 119864 (119909 minus 120583)31205903 (11)

Shock and Vibration 7

0

0

10

20

30Vibration signal

minus10

minus20

minus30

Acce

lera

tion

(mＭ

2)

01 02 03 04 05 06 07 08 09 1

Time (s)

(a)

0

2

4

6

Time (s)

Acoustic signal

Soun

d pr

essu

re (P

a)

minus2

minus4

0 05 1 15 2 25 3

(b)

Figure 7 Measured signals (a) acceleration (b) sound pressure

0 10 20 30 40 50 60

0

1IM

F5

minus1

0 10 20 30 40 50 60

0

05

IMF6

minus05

0 10 20 30 40 50 60

0

05

IMF8

minus05

0 10 20 30 40 50 60

0

05

Time (ms)

RES

minus05

0 10 20 30 40 50 60

0

02

IMF7

minus02

0 10 20 30 40 50 60

0

20

Sign

al

minus20

0 10 20 30 40 50 60

0

20

IMF1

minus20

0 10 20 30 40 50 60

0

10

IMF2

minus10

0 10 20 30 40 50 60

0

2

IMF3

minus2

0 10 20 30 40 50 60

0

1

Time (ms)

IMF4

minus1

Figure 8 IMFs decomposed from measured signal

(d) Variance based on IMFs is

1205902 = 1119873 minus 1119873sum119894=1

1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (12)

(e) Segmented energy ratio based onHilbert marginal energyspectrum the frequency was divided into 10 bands and

the ratio of each frequency band to the total energy wascalculated

119864119873119894 = 119864119894sum10119895=1 119864119895 (13)

When using acceleration as the algorithmrsquos only inputwe chose 119894 statistics from the 5 statistics as feature input 119894varied from 1 to 5 After sum5119894=1 1198621198945 = 31 experiments the mosteffective combination of statistics for acceleration input is

8 Shock and Vibration

Table 2 Recognition rates with different combinations feature

Number of acceleration statistics Number of sound statistics0 1 2 3 4 5

0 6657 7927 8237 8173 81971 6963 6410 6800 7217 7590 55432 7817 7240 7477 7377 7233 68873 8890 9657 8093 8907 8230 81734 9073 9003 9793 8940 9243 66235 8330 9913 8803 9283 9213 7447

determined It is found that when the number of statisticsis set to 4 corresponding to kurtosis energy variance andenergy statistics of frequency division the recognition ratereached the maximum value (9073) as shown in Table 2In the same way we determined the feature when the soundpressure is the algorithmrsquos only input The correspondingcombination of statistics is kurtosis variance and energystatistics of frequency division (8237)

When the acceleration and sound pressure are set as inputtogether we chose 119894 statistics from5 acceleration statistics and119895 statistics from 5 sound pressure statistics as feature inputthe total number of combinations is sum5119894=1 1198621198945 times sum5119894=1 1198621198945 =961 After 961 experiments it was found that when thenumber of acceleration statistics is set to 5 and the numberof sound pressure statistics is set to 1 corresponding to 5all statistics of acceleration and skewness of sound pressurethe recognition rate reached the maximum value (9913)And other comparison experiments use all 5 statistics ofacceleration and skewness of sound pressure as feature input

33 Experiments

331 Comparison among Algorithms with Different DBNStructures The network structure which includes the unitnumber of the visible and hidden layers has a profound influ-ence on network performance A total of 100 combinationswere tested for each DBN to seek the optimal structure Thefirst and second layers of acceleration and sound pressureDBNs both varied from 50 to 500 at intervals of 50The resultof acceleration DBN is shown in Figure 9 Good networkperformance was observed when the first hidden layer wasset from 50 to 150 and the second hidden layer was set atmore than 150 The structure of acceleration DBN was set to[42 100 250] to build the best recognition deep networkThestructure of sound pressure DBN could also be determined inthe samemanner with its optical structure set to [8 150 100]332 Comparison among Algorithms with Different Numbersof Joint Representative Units The number of joint represen-tative RBM units is considerably important for the proposedmodel performance which belongs to the hyperparameter ofthe network The number of joint representative units variedfrom 25 to 1000 to obtain a better performance results shownin Figure 10

It shows that the performance gradually and rapidlybecomes better initially then stabilizes and finally decreases

because of the increasing number of joint representativeunits This result could be attributed to having more unitsresulting in more parameters for learning more informationregarding inputs However excessive parameters could not belearned because the number of training samples was limitedIf the number of units was too small then learning theinformation would be difficult The network performed bestwhen the unit number was set to 150

333 Comparison of the Algorithms Using Unimodality andBimodality One of the key ideas of this study is integratingacceleration and sound pressure for coal-rock recognitionThus algorithms with unimodality and bimodality werecompared Figure 11 is the results of comparison of algorithmsusing acceleration and sound pressure respectively It showsthat the method using acceleration (9073) has betterperformance thanusing soundpressure (8237) but inferiorto the combination of two modalities (9913)

Another issue is that the modality number may affectthe performance of network algorithms with different inputmethods were compared One input method was treatingacceleration and sound pressure as one vector input fed intothe recognition network The recognition result is shown asthe green bar in Figure 12The other input method was treat-ing acceleration and sound pressure features separately as twomodalities fed into the recognition network The recognitionresult is shown as the yellow bar Separately treating twomodalities evidently helped the algorithm obtain a betterperformance (363) than combining features together

334 Comparison among Algorithms with and without RBMsPretraining plays an important role in the recognition ofcoal and rock Figure 13 shows the comparison of algorithmswith and without pretraining The comparison shows thatthe model with unsupervised pretrained RBMs (9913)has 483 improvement compared to the model withoutpretraining (9430) This result is due to the BP algorithmthat used gradient descent to gain the optima and pretrainingwhich allowed the network to be fine-tuned at a relativelygood initial value rather than a few random initial pointsthereby potentially avoiding the risk of falling into poor localoptima Therefore pretraining is significantly necessary formodel training

335 Comparison among Algorithms with and without Trans-fer Learning Transfer learning plays the same role with

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 7

0

0

10

20

30Vibration signal

minus10

minus20

minus30

Acce

lera

tion

(mＭ

2)

01 02 03 04 05 06 07 08 09 1

Time (s)

(a)

0

2

4

6

Time (s)

Acoustic signal

Soun

d pr

essu

re (P

a)

minus2

minus4

0 05 1 15 2 25 3

(b)

Figure 7 Measured signals (a) acceleration (b) sound pressure

0 10 20 30 40 50 60

0

1IM

F5

minus1

0 10 20 30 40 50 60

0

05

IMF6

minus05

0 10 20 30 40 50 60

0

05

IMF8

minus05

0 10 20 30 40 50 60

0

05

Time (ms)

RES

minus05

0 10 20 30 40 50 60

0

02

IMF7

minus02

0 10 20 30 40 50 60

0

20

Sign

al

minus20

0 10 20 30 40 50 60

0

20

IMF1

minus20

0 10 20 30 40 50 60

0

10

IMF2

minus10

0 10 20 30 40 50 60

0

2

IMF3

minus2

0 10 20 30 40 50 60

0

1

Time (ms)

IMF4

minus1

Figure 8 IMFs decomposed from measured signal

(d) Variance based on IMFs is

1205902 = 1119873 minus 1119873sum119894=1

1003816100381610038161003816119909119894 minus 12058310038161003816100381610038162 (12)

(e) Segmented energy ratio based onHilbert marginal energyspectrum the frequency was divided into 10 bands and

the ratio of each frequency band to the total energy wascalculated

119864119873119894 = 119864119894sum10119895=1 119864119895 (13)

When using acceleration as the algorithmrsquos only inputwe chose 119894 statistics from the 5 statistics as feature input 119894varied from 1 to 5 After sum5119894=1 1198621198945 = 31 experiments the mosteffective combination of statistics for acceleration input is

8 Shock and Vibration

Table 2 Recognition rates with different combinations feature

Number of acceleration statistics Number of sound statistics0 1 2 3 4 5

0 6657 7927 8237 8173 81971 6963 6410 6800 7217 7590 55432 7817 7240 7477 7377 7233 68873 8890 9657 8093 8907 8230 81734 9073 9003 9793 8940 9243 66235 8330 9913 8803 9283 9213 7447

determined It is found that when the number of statisticsis set to 4 corresponding to kurtosis energy variance andenergy statistics of frequency division the recognition ratereached the maximum value (9073) as shown in Table 2In the same way we determined the feature when the soundpressure is the algorithmrsquos only input The correspondingcombination of statistics is kurtosis variance and energystatistics of frequency division (8237)

When the acceleration and sound pressure are set as inputtogether we chose 119894 statistics from5 acceleration statistics and119895 statistics from 5 sound pressure statistics as feature inputthe total number of combinations is sum5119894=1 1198621198945 times sum5119894=1 1198621198945 =961 After 961 experiments it was found that when thenumber of acceleration statistics is set to 5 and the numberof sound pressure statistics is set to 1 corresponding to 5all statistics of acceleration and skewness of sound pressurethe recognition rate reached the maximum value (9913)And other comparison experiments use all 5 statistics ofacceleration and skewness of sound pressure as feature input

33 Experiments

331 Comparison among Algorithms with Different DBNStructures The network structure which includes the unitnumber of the visible and hidden layers has a profound influ-ence on network performance A total of 100 combinationswere tested for each DBN to seek the optimal structure Thefirst and second layers of acceleration and sound pressureDBNs both varied from 50 to 500 at intervals of 50The resultof acceleration DBN is shown in Figure 9 Good networkperformance was observed when the first hidden layer wasset from 50 to 150 and the second hidden layer was set atmore than 150 The structure of acceleration DBN was set to[42 100 250] to build the best recognition deep networkThestructure of sound pressure DBN could also be determined inthe samemanner with its optical structure set to [8 150 100]332 Comparison among Algorithms with Different Numbersof Joint Representative Units The number of joint represen-tative RBM units is considerably important for the proposedmodel performance which belongs to the hyperparameter ofthe network The number of joint representative units variedfrom 25 to 1000 to obtain a better performance results shownin Figure 10

It shows that the performance gradually and rapidlybecomes better initially then stabilizes and finally decreases

because of the increasing number of joint representativeunits This result could be attributed to having more unitsresulting in more parameters for learning more informationregarding inputs However excessive parameters could not belearned because the number of training samples was limitedIf the number of units was too small then learning theinformation would be difficult The network performed bestwhen the unit number was set to 150

333 Comparison of the Algorithms Using Unimodality andBimodality One of the key ideas of this study is integratingacceleration and sound pressure for coal-rock recognitionThus algorithms with unimodality and bimodality werecompared Figure 11 is the results of comparison of algorithmsusing acceleration and sound pressure respectively It showsthat the method using acceleration (9073) has betterperformance thanusing soundpressure (8237) but inferiorto the combination of two modalities (9913)

Another issue is that the modality number may affectthe performance of network algorithms with different inputmethods were compared One input method was treatingacceleration and sound pressure as one vector input fed intothe recognition network The recognition result is shown asthe green bar in Figure 12The other input method was treat-ing acceleration and sound pressure features separately as twomodalities fed into the recognition network The recognitionresult is shown as the yellow bar Separately treating twomodalities evidently helped the algorithm obtain a betterperformance (363) than combining features together

334 Comparison among Algorithms with and without RBMsPretraining plays an important role in the recognition ofcoal and rock Figure 13 shows the comparison of algorithmswith and without pretraining The comparison shows thatthe model with unsupervised pretrained RBMs (9913)has 483 improvement compared to the model withoutpretraining (9430) This result is due to the BP algorithmthat used gradient descent to gain the optima and pretrainingwhich allowed the network to be fine-tuned at a relativelygood initial value rather than a few random initial pointsthereby potentially avoiding the risk of falling into poor localoptima Therefore pretraining is significantly necessary formodel training

335 Comparison among Algorithms with and without Trans-fer Learning Transfer learning plays the same role with

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

8 Shock and Vibration

Table 2 Recognition rates with different combinations feature

Number of acceleration statistics Number of sound statistics0 1 2 3 4 5

0 6657 7927 8237 8173 81971 6963 6410 6800 7217 7590 55432 7817 7240 7477 7377 7233 68873 8890 9657 8093 8907 8230 81734 9073 9003 9793 8940 9243 66235 8330 9913 8803 9283 9213 7447

determined It is found that when the number of statisticsis set to 4 corresponding to kurtosis energy variance andenergy statistics of frequency division the recognition ratereached the maximum value (9073) as shown in Table 2In the same way we determined the feature when the soundpressure is the algorithmrsquos only input The correspondingcombination of statistics is kurtosis variance and energystatistics of frequency division (8237)

When the acceleration and sound pressure are set as inputtogether we chose 119894 statistics from5 acceleration statistics and119895 statistics from 5 sound pressure statistics as feature inputthe total number of combinations is sum5119894=1 1198621198945 times sum5119894=1 1198621198945 =961 After 961 experiments it was found that when thenumber of acceleration statistics is set to 5 and the numberof sound pressure statistics is set to 1 corresponding to 5all statistics of acceleration and skewness of sound pressurethe recognition rate reached the maximum value (9913)And other comparison experiments use all 5 statistics ofacceleration and skewness of sound pressure as feature input

33 Experiments

331 Comparison among Algorithms with Different DBNStructures The network structure which includes the unitnumber of the visible and hidden layers has a profound influ-ence on network performance A total of 100 combinationswere tested for each DBN to seek the optimal structure Thefirst and second layers of acceleration and sound pressureDBNs both varied from 50 to 500 at intervals of 50The resultof acceleration DBN is shown in Figure 9 Good networkperformance was observed when the first hidden layer wasset from 50 to 150 and the second hidden layer was set atmore than 150 The structure of acceleration DBN was set to[42 100 250] to build the best recognition deep networkThestructure of sound pressure DBN could also be determined inthe samemanner with its optical structure set to [8 150 100]332 Comparison among Algorithms with Different Numbersof Joint Representative Units The number of joint represen-tative RBM units is considerably important for the proposedmodel performance which belongs to the hyperparameter ofthe network The number of joint representative units variedfrom 25 to 1000 to obtain a better performance results shownin Figure 10

It shows that the performance gradually and rapidlybecomes better initially then stabilizes and finally decreases

because of the increasing number of joint representativeunits This result could be attributed to having more unitsresulting in more parameters for learning more informationregarding inputs However excessive parameters could not belearned because the number of training samples was limitedIf the number of units was too small then learning theinformation would be difficult The network performed bestwhen the unit number was set to 150

333 Comparison of the Algorithms Using Unimodality andBimodality One of the key ideas of this study is integratingacceleration and sound pressure for coal-rock recognitionThus algorithms with unimodality and bimodality werecompared Figure 11 is the results of comparison of algorithmsusing acceleration and sound pressure respectively It showsthat the method using acceleration (9073) has betterperformance thanusing soundpressure (8237) but inferiorto the combination of two modalities (9913)

Another issue is that the modality number may affectthe performance of network algorithms with different inputmethods were compared One input method was treatingacceleration and sound pressure as one vector input fed intothe recognition network The recognition result is shown asthe green bar in Figure 12The other input method was treat-ing acceleration and sound pressure features separately as twomodalities fed into the recognition network The recognitionresult is shown as the yellow bar Separately treating twomodalities evidently helped the algorithm obtain a betterperformance (363) than combining features together

334 Comparison among Algorithms with and without RBMsPretraining plays an important role in the recognition ofcoal and rock Figure 13 shows the comparison of algorithmswith and without pretraining The comparison shows thatthe model with unsupervised pretrained RBMs (9913)has 483 improvement compared to the model withoutpretraining (9430) This result is due to the BP algorithmthat used gradient descent to gain the optima and pretrainingwhich allowed the network to be fine-tuned at a relativelygood initial value rather than a few random initial pointsthereby potentially avoiding the risk of falling into poor localoptima Therefore pretraining is significantly necessary formodel training

335 Comparison among Algorithms with and without Trans-fer Learning Transfer learning plays the same role with

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 9

0 100 200 300 400 500

0100200300400500

20

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

Reco

gniti

on ac

cura

cy(

)

(a)

Reco

gniti

on ac

cura

cy(

)

0 100 200 300 400 500

010020030040050020

40

60

80

100

The number of first hidden

layer units

The number of second hidden layer units

(b)

0 100 200 300 400 500

0100200300400500

5060708090

100

Reco

gniti

on ac

cura

cy(

)

The number of first hidden

layer units

The number of second hidden layer units

(c)

Figure 9 Recognition results for different DBN1 structures (a) coal (b) rock (c) Ave

0 100 200 300 400 500 600 700 800 900 100050556065707580859095

100105

5383

6646

9606

9446

9913 9897

98999699

94199129

9292

8286

8756

Number of joint representative units

Reco

gniti

on ac

cura

cy (

)

Figure 10 The recognition results with different numbers of joint units

pretraining in the recognition of coal and rock (Table 3)In this experiment algorithms with and without transferlearning were compared Herein the algorithm is pretrainedfirst by the data obtained from the laboratory environmentsimulating coal and rock hitting the tail beam then pre-trained by the irrelevant data (hand knocking the table) Theresult shows that the model with transfer learning has 070improvement compared with the model without transferlearningThis result is due to the transfer learning optimizingthe initial value of formal training to enable the networkto easily obtain the global optimum which is similar topretraining

336 Comparison of Algorithms Using Different Lengths ofTime In this experiment algorithms with different lengthsof time per sample from 78125 (each sample contains512 sampling points) to 125ms (each sample contains 8192sampling points) were compared Figure 14 shows thatgiven the increasing length of sample time recognitionaccuracy rapidly improves initially then stabilizes and finallydecreases slightly The algorithm obtained the best perfor-mance when the sample time length was set to 625ms(each sample comprised 4096 sampling points) A longersample time resulted in more included information andperformance gradually improved with the increasing length

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

10 Shock and Vibration

Coal Rock Average0

20

40

60

80

100

Reco

gniti

on ac

cura

cy (

) 99139073

8237

99079616

7760

9920

85298713

Sound pressureAccelerationBimodality

Figure 11 Comparison of algorithms with unimodality and bimodality

Coal Rock Ave0

20

40

60

80

100

120

91739920 9927 9907 9550 9913

Reco

gniti

on ac

cura

cy (

)

Input togetherInput separately

Figure 12 Comparison of algorithms with different input methods

Coal Rock Ave0

20

40

60

80

100

120

9680 99209180

9907 94309913

Reco

gniti

on ac

cura

cy (

)

BPBP + pretraining

Figure 13 Comparison of algorithms with and without pretraining

of time However when the time length of the sample wasextremely long a few samples may contain two states and alonger time would increase the data processing time therebyaffecting the real-time performance

337 Comparison amongAlgorithmswithDifferent ClassifiersIn this experiment to prove the superiority of the proposed

method other five algorithms were compared including 119896-nearest neighbor (KNN) support vector machine (SVM)Naıve Bayes decision tree and DBN As shown in Figure 15decision tree is inferior to the other algorithms in this experi-ment Naıve Bayes performs better than the other methods inrock recognition except for DBN and the proposed methodSVM and KNN fairly perform and KNN performs better

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 11

78125 15625 3125 625 125

50

60

70

80

90

100

110

50

9563 9883 9913 99

The length of time for each sample (ms)

Reco

gniti

on ac

cura

cy (

)

Figure 14 Comparison among algorithms with different frame rates

Decisiontree

NaiveByes

SVM

CoalRockAve

KNN DBN Proposed50

60

70

80

90

100

110

Reco

gniti

on ac

cura

cy (

)

9913

9907

9550

8663

8429

8120

6933

9927

8787

7962

9547

6760

9920

9173

854

8895

6693

7107

Figure 15 Proposed method versus other methods

Table 3 Comparison among algorithms with and without transferlearning

Method Coal Rock AveWithout transfer learning 9696 9990 9843With transfer learning 9920 9907 9913

than SVM especially in rock recognition DBNperformswellin rock recognition however in the coal recognition DBNis inferior to the proposed method Evidently the proposedmethod performs best in this comparison experiment

338 Real-Time Comparison of Algorithms In this exper-iment the real time of methods was compared becausefast and real-time identification of coal or rock during theprocess of top coal caving is very important to achieve theautomated mining The processing time measured using thecomputer system clock is estimated using Matlab R2017aon an Intel(R) Core (TM) i7-6700 CPU 340GHz with8GB RAM running on Windows 10 operating system Theaverage data processing time and recognition time of eachalgorithm per sample are shown in Table 4 It can be seen

that data processing time accounts for the vast majority ofthe total time (more than 999) Because the total time isless than 343ms in the actual application it can be identifiedtwice per second all these methods can meet the real-timerequirement

4 Conclusions

A novel method is proposed in this study based on bimodaldeep learning and Hilbert-Huang transform for the identifi-cation of coal-rock in top coal caving Four main innovationsare present in this study First the bimodal learning methodis adopted in enabling DNN to study the characteristicsof coal and rock completely Second the transfer learningmethod is adopted to solve the problem wherein a largenumber of samples are necessary to train the DNN Thirdthe extracted features of acceleration and sound pressuresignals are combined to extract the most efficient featuresFourth the most suitable installation location for the sensorsis selected

For future works the authors plan to obtain substantialcoal-rock data in different coal mines to improve the appli-cability of the proposed method use more effective feature

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

12 Shock and Vibration

Table 4 Comparison of the recognition time

Method Data processingms Recognitionms TotalmsDecision tree 34202 277 times 10minus3 34202Naıve Byes 34202 196 times 10minus3 34202SVM 34202 443 times 10minus2 34206KNN 34202 326 times 10minus1 34234DBN 34202 502 times 10minus3 34203Proposed 34202 193 times 10minus2 34204

extraction methods improve the real-time performance andstrengthen the stability and robustness Moreover producingan intelligent identification with high precision and speedadaptability and better practical products is the aim for futureworks

Conflicts of Interest

The authors declare no conflicts of interest

Authorsrsquo Contributions

Zengcai Wang designed the experimental system GuoxinZhang and Lei Zhao formulated the proposed algorithmYazhou Qi and Jinshan Wang performed the experimentsGuoxin Zhang analyzed the data and wrote the paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Grant no 51174126)

References

[1] J Qin J Zheng X Zhu and D Shi ldquoEstablishment of atheoretical model of sensor for identification of coal and rockinterface by natural 120574-ray and underground trialsrdquo Journal ofChina Coal Society vol 21 pp 513ndash516 1996

[2] S L Bessinger andM G Nelson ldquoRemnant roof coal thicknessmeasurement with passive gamma ray instruments in coalminesrdquo IEEE Transactions on Industry Applications vol 29 no3 pp 562ndash565 1993

[3] Z C Wang R L Wang and J H Xu ldquoResearch on coal seamthickness detection by natural Gamma ray in shearer horizoncontrolrdquo Journal of China Coal Society vol 27 pp 425ndash4292002

[4] A D Strange 2007 Robust thin layer coal thickness estimationusing ground penetrating radar

[5] F Ren Z Liu and Z Yang ldquoHarmonic response analysis oncutting part of shearer physical simulation system paper titlerdquoin Proceedings of the 2010 IEEE 10th International Conference onSignal Processing ICSP2010 pp 2509ndash2513 October 2010

[6] J Sun and B Su ldquoCoal-rock interface detection on the basis ofimage texture featuresrdquo International Journal of Mining Scienceand Technology vol 23 no 5 pp 681ndash687 2013

[7] W Hou ldquoIdentification of coal and gangue by feed-forwardneural network based on data analysisrdquo International Journal ofCoal Preparation and Utilization pp 1ndash11 2017

[8] G Zhang Z Wang and L Zhao ldquoRecognition of rockndashcoalinterface in top coal caving through tail beam vibrations byusing stacked sparse autoencodersrdquo Journal of Vibroengineeringvol 18 no 7 pp 4261ndash4275 2016

[9] L Si ZWang X Liu C Tan and L Zhang ldquoCutting state diag-nosis for shearer through the vibration of rocker transmissionpart with an improved probabilistic neural networkrdquo Sensors(Switzerland) vol 16 no 4 2016

[10] BWang ZWang and S Zhu ldquoCoal-rock interface recognitionbased on time series analysisrdquo in Proceedings of the 2010International Conference on Computer Application and SystemModeling ICCASM 2010 pp V8356ndashV8359 October 2010

[11] F Ren Z Liu and Z Yang ldquoDynamics analysis for cuttingpart of shearer physical simulation systemrdquo in Proceedings of theIEEE International Conference on Information and Automation(ICIA rsquo10) pp 260ndash264 IEEE Harbin China June 2010

[12] L Si Z Wang C Tan and X Liu ldquoVibration-based signalanalysis for shearer cutting status recognition based on localmean decomposition and fuzzy C-means clusteringrdquo AppliedSciences vol 7 no 2 p 164 2017

[13] J Xu ZWang JWang C Tan L Zhang and X Liu ldquoAcoustic-based cutting pattern recognition for shearer through fuzzy C-means and a hybrid optimization algorithmrdquo Applied Sciences(Switzerland) vol 6 no 10 article no 294 2016

[14] Y-w Liang and S-b Xiong ldquoForecast of coal-rock interfacebased on neural network and Dempster-Shafter theoryrdquoMeitanXuebaoJournal of the China Coal Society vol 1 p 17 2003

[15] Y-l Zhang and S-x Zhang ldquoAnalysis of coal and gangue acous-tic signals based on Hilbert-Huang transformationrdquo Journal ofChina Coal Society vol 1 p 44 2010

[16] J R Markham P R Solomon and P E Best ldquoAn FT-IR basedinstrument formeasuring spectral emittance of material at hightemperaturerdquo Review of Scientific Instruments vol 61 no 12 pp3700ndash3708 1990

[17] J Ngiam A Khosla M Kim J Nam H Lee and A YNg ldquoMultimodal deep learningrdquo in Proceedings of the 28thInternational Conference on Machine Learning (ICML rsquo11) pp689ndash696 Omnipress Bellevue Wash USA July 2011

[18] W Zhang Y Zhang L Ma J Guan and S Gong ldquoMultimodallearning for facial expression recognitionrdquo Pattern Recognitionvol 48 no 10 pp 3191ndash3202 2015

[19] L Zhao Z Wang X Wang Y Qi Q Liu and G ZhangldquoHuman fatigue expression recognition through image-baseddynamic multi-information and bimodal deep learningrdquo Jour-nal of Electronic Imaging vol 25 no 5 Article ID 053024 2016

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 13

[20] B Q Huynh H Li and M L Giger ldquoDigital mammographictumor classification using transfer learning from deep convolu-tional neural networksrdquo Journal of Medical Imaging vol 3 no3 Article ID 034501 2016

[21] G Prudhomme L Berthe J Benier O Bozier and P MercierldquoRadiometric short-term fourier transform analysis of pho-tonic doppler velocimetry recordings and detectivity limitrdquo inProceedings of the Conference of the American Physical SocietyTopical Group on Shock Compression of Condensed Matter p160007 AIP Publishing Tampa Bay Fla USA

[22] P Neri ldquoBladed wheels damage detection through non-harmonic fourier analysis improved algorithmrdquo MechanicalSystems and Signal Processing vol 88 pp 1ndash8 2017

[23] W Zhao Z Wang J Ma and L Li ldquoFault diagnosis of ahydraulic pump based on the CEEMD-STFT time-frequencyentropy method and multiclass SVM classifierrdquo Shock andVibration vol 2016 Article ID 2609856 8 pages 2016

[24] M Zhang G Wang and G M Evans ldquoFlow visualizationsaround a bubble detaching from a particle in turbulent flowsrdquoMinerals Engineering vol 92 pp 176ndash188 2016

[25] T Putra S Abdullah D Schramm M Nuawi and T Bruck-mann ldquoReducing cyclic testing time for components of auto-motive suspension system utilising the wavelet transform andthe fuzzy C-Meansrdquo Mechanical Systems and Signal Processingvol 90 pp 1ndash14 2017

[26] M A Brdys M T Brdys and S M Maciejewski ldquoAdaptivepredictions of the eurozłoty currency exchange rate using statespace wavelet networks and forecast combinationsrdquo Interna-tional Journal of Applied Mathematics and Computer Sciencevol 26 no 1 pp 161ndash173 2016

[27] D Sun andQ Ren ldquoSeismic damage analysis of concrete gravitydam based on wavelet transformrdquo Shock and Vibration vol2016 Article ID 6841836 8 pages 2016

[28] N E Huang and ZWu ldquoA review onHilbert-Huang transformmethod and its applications to geophysical studiesrdquo Reviews ofGeophysics vol 46 no 2 2008

[29] M Lozano J A Fiz andR Jane ldquoPerformance evaluation of theHilbert-Huang transform for respiratory sound analysis and itsapplication to continuous adventitious sound characterizationrdquoSignal Processing vol 120 pp 99ndash116 2016

[30] H Xu J Liu H Hu and Y Zhang ldquoWearable sensor-based human activity recognition method with multi-featuresextracted from Hilbert-Huang transformrdquo Sensors (Switzer-land) vol 16 no 12 article no 2048 2016

[31] X Yu E Ding C Chen X Liu and L Li ldquoA novel characteristicfrequency bands extraction method for automatic bearingfault diagnosis based on Hilbert Huang transformrdquo Sensors(Switzerland) vol 15 no 11 pp 27869ndash27893 2015

[32] P Smolensky Information Processing in Dynamical SystemsFoundations of Harmony Theory DTIC Document 1986

[33] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006

[34] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo American Associationfor the Advancement of Science Science vol 313 no 5786 pp504ndash507 2006

[35] A Coates H Lee and A Y Ng An Analysis of Single-LayerNetworks in Unsupervised Feature Learning vol 1001 of 2 2010

[36] H Larochelle and Y Bengio ldquoClassification using discrimina-tive restricted boltzmann machinesrdquo in Proceedings of the 25th

International Conference on Machine Learning (ICML rsquo08) pp536ndash543 July 2008

[37] Z H Wu and N E Huang ldquoA study of the characteristics ofwhite noise using the empirical mode decomposition methodrdquoProceedings of the Royal Society A Mathematical Physical andEngineering Sciences vol 460 no 2046 pp 1597ndash1611 2004

[38] R Ditommaso M Mucciarelli S Parolai and M PicozzildquoMonitoring the structural dynamic response of a masonrytower comparing classical and time-frequency analysesrdquo Bul-letin of Earthquake Engineering vol 10 no 4 pp 1221ndash12352012

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 201

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of