a survey on deep learning techniques in wireless signal

Review ArticleA Survey on Deep Learning Techniques in WirelessSignal Recognition

Xiaofan Li 12 Fangwei Dong 2 Sha Zhang12 and Weibin Guo2

1State Radio Monitoring Center and Testing Center Beijing China2Shenzhen Institute of Radio Testing and Tech Shenzhen China

Correspondence should be addressed to Fangwei Dong dongfangweisrtcorgcn

Received 16 November 2018 Accepted 30 January 2019 Published 17 February 2019

Guest Editor Huaming Wu

Copyright copy 2019 Xiaofan Li 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

Wireless signal recognition plays an important role in cognitive radio which promises a broad prospect in spectrum monitoringand management with the coming applications for the 5G and Internet of Things networks Therefore a great deal of research andexploration on signal recognition has been done and a series of effective schemes has been developed In this paper a brief overviewof signal recognition approaches is presented More specifically classical methods emerging machine learning and deep leaningschemes are extended from modulation recognition to wireless technology recognition with the continuous evolution of wirelesscommunication system In addition the opening problems and new challenges in practice are discussed Finally a conclusion ofexisting methods and future trends on signal recognition is given

1 Introduction

With the increasing innovation in wireless communicationsystem numerous wireless terminals and equipment areconstantly emerging which has brought profound changes toour daily life Unfortunately the limited spectrum resourcecan hardly meet the ever-changing demand of the coming5G [1] and Internet of Things (IoT) networks [2] whichposes a significant challenge to the spectrum utilization andmanagement The Federal Communications Commission(FCC) and European Union (EU) authorities have attachedhigh priority to spectrum policy and committed to fur-ther improve the performance of spectrum sensing as wellas signal recognition algorithms to satisfy the demand ofspectrum management Some concepts including dynamicspectrumaccess (DSA) and cognitive spectrum-sharing tech-niques have aroused widespread discussions in academiaHowever much work is limited to specific scenarios andpoor in adaptability to the different channel conditions anddevice types Therefore new spectrum sensing schemes andnovel signal recognition mechanisms have attracted moreand more attention which pave the way to cognitive radio(CR) [3]

Wireless signal recognition (WSR) has great promise onmilitary and civilian applications [4] which may includesignal reconnaissance and interception antijamming anddevices identification Generally WSRmainly includes mod-ulation recognition (MR) and wireless technology recog-nition (WTR) MR also known as automatic modulationclassification (AMC) is first widely used in military field andlater extended to civilian field MR classifies radio signals byidentifying modulation modes which facilitates to evaluatewireless transmission schemes and device types What ismore MR is capable of extracting digital baseband informa-tion even under the condition of limited prior informationRecently WTR attracts much attention with the variouswireless communication technologies emerging and develop-ing Wireless technology usually contains various technicalstandards not just a single modulation mode Moreoverdifferent technical standards may adopt the same modulationmode WTR is able to leverage more comprehensive featuresto identify technical standards which has important practicalsignificance in current increasingly tense electromagneticenvironment In addition the estimation of some parameterssuch as frequency bandwidth and symbol rate may alsocontribute to WSR Due to the variety and complexity of

HindawiWireless Communications and Mobile ComputingVolume 2019 Article ID 5629572 12 pageshttpsdoiorg10115520195629572

2 Wireless Communications and Mobile Computing

signal transmission mechanisms it is difficult to recognizecomplex signals by only estimating single or few parametersTherefore the MR and WTR technologies are focused in thispaper

Traditional algorithms of MR could mainly be separatedinto two groups likelihood-based (LB) and feature-based(FB) approaches [5] LB approaches are based on hypoth-esis testing theory the performance based on decision-theoretic is optimal but suffers high computation complexityTherefore feature-based approaches as suboptimal classifierswere developed for application in practice In particularthe feature-based approaches usually extracted features forthe preprocessing and then employed classifiers to real-ized modulation classification Conventional FB approachesheavily rely on the expertrsquos knowledge which may per-form well on specialized solutions but poor in generalityand suffer high complexity and time-consuming To tacklethese problems machine learning (ML) classifiers have beenadopted and shown great advantages for example supportvector machine (SVM) in [6] Although ML methods havethe advantage of classification efficiency and performancethe feature engineering to some extent still depends onexpert experience resulting in degradation of accuracy rateTherefore the self-learning ability is very important whenconfronted with unknown environment A new dawn seemsto have arrived since the DL performs very well in computervision (CV) [7] machine translation (MT) [8] and naturallanguage processing (NLP) [9] DL architecture has alsobeen introduced to MR for instance the convolutionalneural networks (CNN) model is employed in modulationclassification without expert feature extraction [10] whichdemonstrates excellent performance both on efficiency andaccuracy

On the other hand current communication systems tendto be complex and diverse various new wireless technologiesare constantly updated and the coexistence of homogeneousand heterogeneous signals is entering a new norm As aresult the detection and recognition of complex signals willbe confronted with new dilemma and the methods forsignal recognition are needed to keep up with the paceFortunately some literature on WTR has emerged in whichML and DL model with explicit features have also been doneto realize the recognition of specific wireless technologiesMeanwhile some emerging applications are been exploredsuch as base station detection interfere signals recognitionand localization in mobile communication network [11]

In this paper the need of deep learning in signal recog-nition is reviewed in Section 2 Section 3 introduces themodulation recognition with various approaches In Sec-tion 4 wireless technology recognition is presented Openingproblems are discussed in Section 5 Section 6 concludes thepaper

2 Need of Deep Learning in WirelessSignal Recognition

21 Wireless Signal Recognition With the continuous expan-sion of military and civilian needs communication systems

Wireless signal recognition

Modulation recognitionWireless technology recognition

Frequency and bandwidthestimation

Symbol rate evaluation

middot middot middot

Figure 1 Wireless signal recognition

Input Layer Hidden Layer Output Layer

Figure 2 Neural network framework

and electromagnetic environments have greatly changed overthe past decades the capability of detection and recognitionof communication signals has also made significant progressand gradually become maturity Generally communicationsignal recognition takes advantage of some signal parametersto classify or identify the types of signals These techniquesmay include frequency and bandwidth estimation symbolrate evaluation modulation type classification and wire-less technology identification which could be collectivelyreferred to as wireless signal recognition as shown in Figure 1In this paper we are concerned with the current two maintechnologies namely MR and WTR MR commits to realizethe modulation type recognition so as to evaluate wirelesstransmission schemes and device types while WTR takeswireless technology identification as object for improvinginterference management and electromagnetic environmen-tal assessment

22 Definition of DL Problem The concept of DL originatesfrom the research on artificial neural network [12] and thegoal is to understand data by mimicking the mechanism ofthe human brain [13] The basic neural network frameworkconsists of three parts input hidden and output layer andis shown in Figure 2 Hidden layers maybe one layer ormultilayer and each layer consists of several nodes The

Wireless Communications and Mobile Computing 3

Table 1 ML VS DL in wireless signal recognition

Leaning model Machine learning Deep learning

Application scenarios (i) small signal data (i) high-dimensional signal data(ii) signal under relatively ideal conditions (ii) good feasibility in real field environment

Algorithms

(i) ANN [26 37](ii) KNN [38 91](iii) SVM [6 27 47 48 92](iv) Naıve Bayes [39](v) HMM [46](vi) Fuzzy classifier [93](vii) Polynomial classifier [40 94]

(i) DNN [24 30 31 61](ii) DBN [49 63](iii) CNN [17 19ndash21 54 64 65 70 73ndash76 79 81 82 95 96](iv) LSTM [29 69](v) CRBM [53](vi) Autoencoder network [50 62](vii) Generative adversarial networks [66 67](viii) HDMF [71 72](ix) NFSC [78]

Pros (i) works better on small data(ii) low implementation cost

(i) simple pre-processing(ii) high accuracy and efficiency(iii) adaptive to different applications

Cons

(i) time demanding(ii) complex feature engineering(iii) depends heavily on the representation of the data(iv) prone to curse of dimensionality

(i) demanding large amounts of data(ii) high hardware cost

hellip

b

Outputhellip

1

Sum

Ｒ1

Ｒ2

ＲＧ

Ｑ1

Ｑ2

ＱＧ

f(∙)

Figure 3 Hidden layer node

node presented in Figure 3 is the basic operational unit inwhich the input vector is multiplied by a series of weightsand the sum value is fed into the activation function 119891These operational units contribute to a powerful networkwhich could realize complex functions such as regressionand classification In fact the study of DL makes substantialprogress by Hinton in 2006 [14] which causes a greatsensation DL architecture is comprised of many stackedlayers of neural networks emphasizing the learning fromsuccessive layers to obtain more meaningful information orhigh-level representations The fundamental idea is to utilizefeedback information to iteratively optimize weight valuein multilayer neural networks Moreover the layers with arelatively large scale will bring about amazing effects Due tothe superior performance DL has been employed on a widevariety of tasks including CV MT and NLP

23 FromML to DL AlthoughML is not a new field with theexplosive growth of the amount of data and the developmentof hardware technologyMLhas been one of research hotspots

in both academe and industry [15 16] Recently ML has alsomade great strides in signal recognition ML works well onautomatic signal classification for small signal dataset underrelatively ideal conditions [17] Most algorithms are easy tointerpret and have the advantage of low implementation costWhile for a high-dimensional dataset machine learning isprone to curse of dimensionality and the complex featureengineering is time demanding In terms of algorithm theoryML has a risk of falling into local optimum which may leadto great performance degradation Furthermore ML modelsare trained for specific solutions and lack of generality indifferent real field environments To overcome these issuesDL is developed and achieves high accuracy and efficiency[18] Owing to multilayered artificial neural networks DLneeds few data preprocessing and shows adaptive to differentapplication scenarios A comparison between ML and DL issummarized in Table 1

24 Advantage of DL Applied in Wireless Signal Recogni-tion The perfect combination of DL and signal recognitionpossesses notable superiority On the one hand large-scaledata is essential for the training process in DL modelwhich is accessible to get for various communication piecesof equipment in daily life On the other hand featureengineering can be left out in DL architecture which isan indispensable part for conventional recognition schemesFeature selection from the received signal is usually difficultin practical applications For example the prior informationis essential in the estimation of many parameters whichmay be impractical or inaccurate [19] Although some FBapproaches perform well in certain solutions the featureselection may suffer high complexity and time-consumingso the feature self-learning model is of great significance forrealistic scenarios to free from expert experience [20] Inaddition with the further improvement ofDL algorithms and


Modulationrecognition

Likelihood-basedapproaches

Feature-basedapproaches

Instantaneous timefeatures

Statistical features

Transform features

Other features

Feature learning

Moments andCumulants

Cyclostationarity

Fourier transform

Wavelet transform

S transform

Constellation shape

Zero crossing

Figure 4 Modulation recognition

theory research more application prospects will be excavatedfor signal recognition in future communications systems [21]

3 Modulation Recognition

Modulation recognition has attracted much attention in lastdecades quite a few scholars have presented a variety ofexcellent approaches which can be approximately separatedinto two categories LB [22 23] and FB approaches as shownin Figure 4 Theoretically the LB approaches are capable ofobtain optimal performance by considering all the unknownquantities for the probability density function (PDF) andare usually employed as the theoretical upper bound forthe performance comparison of modulation recognitionHowever such approaches are burdened with high com-putational complexity and are prone to mismatching whenapplying the theoretical system model to the actual scene Toconsider the effects of frequency offset and time drift a largenumber of computational operations bring about extremelyhigh complexity in maximum likelihood-based detector[24]

To tackle the problem of high complexity in practicea large number of suboptimal approaches come into beingThe FB approaches usually extract certain features from thereceived signal then reasonable classifiers are used to classifydifferent modulation signals Since the training samplesare employed to train a good classifier the robustness ofFB approaches is significantly improved for various systemmodels and channel conditions In addition effective featuresand enough training data are also significant for improvingclassification accuracy

To reduce the difficulty in expert-feature extraction andenhance the flexibility for modulation classification appliedin different system and channel fading environment DL is

applied for self-feature learning based on the IQ raw data orsampling data

In the following we introduce the feature-basedapproaches and feature learning approaches in modulationrecognition

31 Feature-Based Approaches These features can be sum-marized as follows instantaneous time features statisticalfeatures transform features other features including con-stellation shape etc The feature-based approaches are morerobust and general for various signals When combined withdeep learning methods the feature-based approaches willprovide a significant improvement in the performance withhigh efficiency and robustness

311 Instantaneous Time Features Generally instantaneoustime features consist of a series of parameters namely carrieramplitude phase and frequency These parameters and thevariation of them are developed to modulation classificationNine key features are summarized in the Table 2

In [25] the authors use eight key features in the recog-nition procedure and compare with some manually selectedsuitable thresholds to classify both analogue and digitalmodulation signals

Unlike the previous approaches with manual thresh-old machine learning has been developed to improve theperformance and reduce tedious threshold operations Tochoose the threshold automatically and adaptively artificialneural networks (ANN) is employed in some literatureIn [26] a universal ANN was proposed for analogue aswell as digital modulation types classification with nine keyfeatures employed Although ANN has achieved successin modulation recognition its overdependence on trainingsample data and easily getting into a local optimum restrict


Table 2 Instantaneous time features [26]

Feature Definition description120574119898119886119909 Maximum value of the power spectral density of the normalized-centered instantaneous amplitude120590119886119901 Standard deviation of the absolute value of the nonlinear component of the instantaneous phase120590119889119901 Standard deviation of the direct value of the nonlinear component of the instantaneous phase120590119886119886 Standard deviation of the absolute value of the normalized-centered instantaneous amplitude120590119886119891 Standard deviation of the absolute value of the normalized instantaneous frequency120590119886 Standard deviation of the normalized-centered instantaneous amplitude119875 Spectrum symmetry12058311988642 Kurtosis of the normalized instantaneous amplitude12058311989142 Kurtosis of the normalized instantaneous frequency

the performance and application while the support vectormachines (SVM) can effectively alleviate these problems In[27] the straightforward ordered magnitude and phase wereadopted by SVM classifier The proposed method is close tothe theoretical upper bound and has an advantage of easyimplementation

Recently deep learning as an extension of ANN isshowing its great potential in modulation recognition [28]Complex features are learned by multiple hidden layersso the processing of input data could be simpler in deeplearning schemes A new model Long Short Term Memory(LSTM) has been applied to modulation recognition whichis suitable for processing the time series data with relativelylong intervals or delays [29]The time domain amplitude andphase data were employed in the model without complexfeature extraction Simulations show that the method couldachieve high classification accuracy at varying SNR condi-tionsThe flexibility of the proposedmodel was also validatedin scenarios with variable symbol rates Unlike the existingmethodswhich only focus onmodulation types classificationin [30 31] they propose a novel scheme for classifyingmodulation format and estimating SNR simultaneously Theasynchronous delay-tap plots are extracted as the trainingdata and two multilayer perceptron (MLP) architectures areadopted for real-world signal recognition tasks

32 Statistical Features Moments cumulants and cyclosta-tionarity will be introduced as following

321 Moments and Cumulants In mathematics momentsare employed to describe the probability distribution of afunction The p-th order and q-th conjugations moment [32]for a received signal 119910(119899) can be given as

119872119901119902 = E [119910 (119899)119901minus119902 (119910lowast (119899))119902] (1)

where E[∙] is expectation operator and (∙)lowast is complexconjugate Although moments have been widely used inthe field of signal process for the advanced ability on noisesuppression most practical applications tend to cumulantsfor its superiority on non-Gaussian random processes

Table 3 Cumulants formulas

Cumulants Formulas11986240 11987240 minus 311987222011986241 11987241 minus 3119872211198722011986242 11987242 minus 10038161003816100381610038161198722010038161003816100381610038162 minus 211987222111986260 11987260 minus 151198722011987240 + 3011987232011986263 11987263 minus 61198722011987241 minus 91198724211987241 + 1811987222011987221 + 1211987232111986280 11987280 minus 35119872240 minus 281198726011987220 + 42011987240119872220 minus 630119872420

In statistical theory cumulants are made up of momentsand can be regarded as an alternative to the momentsCommonly used cumulants are expressed in Table 3

Cumulants are often employed for modulation classifica-tion to against the carrier offsets and non-Gaussian noiseIn [33] fourth-order cumulants were utilized in the taskof signal classification Higher-order cumulants (HOC) areconducive to further enhance the performance and extendthe range of recognition signals In [34] sixth-order cumu-lants show significantly performance improvement than thefourth-order cumulants In [35] eighth-order cumulants areemployed to distinguish between 8PSK and 16PSK signalAdditionally higher-order cyclic cumulants are developed tofurther expand the signal types including 256-QAM in [36]

Somemethods combined cumulants withmachine learn-ing for modulation classification have been proposed inrecent works such as ANN K-Nearest Neighbor (KNN)SVM Naıve Bayes (NB) and polynomial To improve theperformance two updated ANN training algorithms are pre-sented in [37] KNN is attractive due to its easy implement In[38] the authors propose a KNN classifier using cumulants-based features of the intercepted signal Satisfactory perfor-mance was verified by numerous simulations SVM has greatpotential on the recognition of small-scale and high dimen-sional dataset Another interesting scheme combined HOCwith SVM is presented for digital modulation classificationin NB model originates from classical mathematical theoryand the major advantages lie in simplicity and less sensitiveto missing data In [39] the authors proposed a new schemecombined HOCwith NB classifier Simulations indicated thatthe NB classifier was superior to both Maximum Likelihood


and SVM model on computational complexity and close toSVM on performance To further reduce the complexity anovel AMC system based on polynomial classifier was pre-sented in [40] Second fourth and sixth HOCwere utilized toclassifyM-PSK andM-QAM signals with low computationalcomplexity In addition a comprehensive comparison amongKNN SVM NB and the proposed polynomial classifiercombined with extraction of higher order cumulants featuresis carried out in [41] The proposed method achieves goodtradeoff in terms of accuracy and structure simplicity

322 Cyclostationarity A cyclostationary process is a signalwhose statistical characteristics change periodically with time[42] Cyclostationarity is an inherent physical property ofmost signals which is robust against noise and interference[43] Spectral correlation function (SCF) is usually used toexamine and analyze the cyclostationarity of the signal [44]and can be defined as

119878120572119909 (119891) = limΔ119891997888rarrinfin

limΔ119905997888rarrinfin

1Δ119905

sdot intΔ1199052

minusΔ1199052Δ1198911198831Δ119891 (119905 119891 + 1205722) sdot 119883

lowast1Δ119891 (119905 119891 minus 1205722) 119889119905

(2)

where

1198831Δ119891 (119905 V) = int119905+12Δ119891

119905minus12Δ119891119909 (119906) 119890minus1198942120587V119906119889119906 (3)

is the complex envelope value of 119909(119905) corresponding to thefrequency V and the bandwidth Δ119891 and 120572 represents cyclicfrequency and Δ119905 is the measurement interval

As a continuation the spectral correlation function ofdigitally modulated signals was provided in [45] HiddenMarkov Model (HMM) was employed to deal with thefeatures extracted from the cycle frequency domain profilein [46] Simulations show that the presented model wasable to classify the signals at a range of low SNRs Anotherclassification scheme with spectral correlation feature andSVM classifier is developed in [47] The results demonstratethat the algorithm is robust under the condition of low SNRand maintain high accuracy In [48] four features basedon spectral correlation were selected for SVM classifier Theproposed method performs more efficiently in low SNR andlimited training data

Deep learning methods employed spectral correlationfunction features have been applied to signal classificationIn [24] the method using 21 features of baseband signaland a full-connected DNN model is proposed to recognizefive modulation mechanisms The simulations verify that theproposed algorithm outperforms the ANN classifier undervarious channel conditions Deep Belief Network (DBN) isintroduced to pattern recognition tasks in [49] With spectralcorrelation function (SCF) signatures the DBN classifierachieves great success in the field of modulation classificationin various environments In [50] cyclic spectrum statisticsare also adopted in the signal preprocessing and a stackedsparse autoencoder and softmax regression with DL model

are employed to recognize communication signals The com-parison of the proposed methods and several traditionalmachine learning schemes was analyzed by simulationsAnother modulation classification method for very high fre-quency (VHF) signal is presented in [20]The received signalsare transformed into cyclic spectrum then the denoisedspectrum images are fed into CNN model to self-learn theinner features and train classifier to recognize modulationformats finally

33 Transform Features

331 Fourier Transform Fourier transform is a significanttechnique in signal processing and analysis For most signalsfrequency domain analysis is more convenient and intuitivethan time domain and the Fourier transform providesconvenience in decomposing a function of time (a signal)into the frequencies In [51] in order to analyze the signalmodulation method discrete Fourier transform (DFT) andthe calculated amplitude and phase values were employed toclassify MFSKMPSK signals In [52] a classifier adopted fastFourier transform (FFT) has been proposed to classify thereceived MFSK signal

Joint time-frequency transforms play an importantrole in characterizing time-varying frequency informationwhich could compensate for the insufficiency of the one-dimensional solution and process the time and frequencycharacteristics simultaneously The short time-frequencytransform (STFT) is known as a classical time-frequencyrepresentation and has been widely adopted in signal pro-cessing In [53] a novel learning scheme with STFT fea-tures based on DL is proposed which could automaticallyextract features and dodecision-making in complexmissionsThe performance of the proposed method is verified inthe spectrum application of classifying signal modulationtype In [54] the authors propose a DCNN method thattransforms modulation mode to a new time-frequency mapand successfully recognize the hybrid communication signalsunder low SNR condition

332 Wavelet Transform Compared with the Fourier trans-form the wavelet transform (WT) can provide the frequencyof the signals and the time associated to those frequencieswhich offers refined local signals analysis and low computa-tions The paper [55] proposed a WT modulation classifierwithout requiring any a priori knowledge The combinationof WT and SVM is a popular scheme in modulation identifi-cation The wavelet kernel function and SVM for modulationrecognition is employed in [56] The proposed method iscapable of classifying BPSKQPSK andAMsignalswith goodaccuracy The paper extended the SVM classifier with keyfeatures of WT to recognize 8 kinds of digital modulatedsignals

In [57] the authors use higher-order statistical momentsfeatures based on continuous WT and multilayer feed-forward neural network for modulation recognition Takinginto account for various distracters including different SNRfading channels and frequency shift the authors in [17]


propose an improved CNN architecture based on [10] inwhich the pooling layer is employed to retain principalcomponents and reduces the dimension of features Therecognition rate and performance are further enhanced bywavelet denoising technology and appropriate order of thewavelet transformation

333 S Transform S transform (ST) has unique advantagesin possessing both the phase and frequency estimations of theFT and the super-resolution of theWT [58] Early applicationof the ST was in electrical power networks with power qualityanalysis In [59] ST based features extraction for digitalmodulation classification was presented A comparison withthe WT was given by employing different machine learningclassifiers

34 Other Features

341 Constellation Shape In digital modulation the con-stellation is usually used to map the signal into scatteredpointsThe constellation diagram provides an intuitive way torepresent signal basic structure and the relationship betweendifferent modulation modes In [60] the constellation shapewas employed as a reliable signature to realize digital modu-lation recognition

DL has been effectively used in constellation-basedapproaches for automatic modulation classification In [6162] the authors use the DL network combined with IQcomponents of constellation points as features In [63] asimple graphic constellation projection (GCP) scheme forAMC is presented Unlike FB approaches the AMC taskis turned to image recognition technology Reference [64]proposed a DL-based intelligent constellation analyzer thatcould realize modulation type recognition as well as OSNRestimation The authors utilized CNN as DLmodels and treatconstellation map as original processing features Due to theautomatic feature extraction and learning CNN is capableof processing original constellation distribution without theknowledge of any other parameters In [21 65] the signalswere transformed to the images with topological structureand the CNN algorithm was used to deal with the classi-fication In [66] the authors explore a new framework forgenerating additional images in CNN training through usingauxiliary classifier generative adversarial networks (ACGAN)for data augmentation to solve the modulation classificationproblem Reference [67] extended a new semisupervisedlearning model combined with generative adversarial net-works (GANs) The proposed scheme was verified to be amore data-efficient classifier

342 Zero Crossing A zero-crossing sampler is used forrecording the number of zero-crossing voltages of inputsignals which could provide precise information about phaseconversion in wide frequency range Therefore the zero-crossing of the signal can also be applied to modulationclassification The paper [68] develops a modulation recog-nition algorithm with zero-crossing techniques Both thephase difference and the zero-crossing interval histograms

are employed as features The recognizer could identify CWMPSK and MFSK signals with a reasonable classificationaccuracy

35 Feature Learning Feature learning is based on the rawdata or sampling data instead of crafting of expert featuresA recently proposed AMC scheme based on DL modeltakes advantage of CNN classifier [10] The time domain IQsamples are directly fed to CNNmodel and suitable matchedfilters can be learned automatically in different SNRs In[69] the classification accuracy of CNN architectures withdifferent sizes and depths was analyzed The authors alsoprovided a hybrid learning scheme which combines CNNmodel and long short term memory (LSTM) network toachieve better classification performance Reference [70]constructs a CNN model with 5 layers for the recognitionof very high frequency signals Simulation and actual signalsare verified that the frequency offset and noise have a greatimpact on accuracy Reference [19] proposed a novelmodel toobtain the estimation of the frequency offset as well as phasenoise for improving accuracy rate To further improve theperformance of CNN-based recognition schemes in [10] theauthors present a signal distortion correction module (CM)and results show that this CM+CNN scheme achieves betteraccuracy than the existing schemes In [71 72] a heteroge-neous deep model fusion (HDMF) approach is proposedand the two different combinations between CNN and LSTMnetwork without prior information are discussed The time-domain data are delivered to fusionmodel without additionaloperations The results show that the HDMF approach isan excellent alternative for a much better performance inheterogeneous networks

4 Wireless Technology Recognition

Various radio pieces of equipment make an increasing short-age of spectrum resources As a result the interference fortransmissions is unavoidable in a coexistence environmentwhich leads to a decline in spectrum efficiency What is moreserious is that the diversity of communication technologiesand heterogeneous networks cause a more complex electro-magnetic environment Therefore the recognition of signaltechnology is of great significance in spectrum sharing andinterference management

Similar to MR the feature-based approaches can also bedeveloped for WTR Due to the development of machinelearning and deep learning WTR favors explicit featuressuch as time frequency and time-frequency features Theliterature related to DL in MR and WTR is summarized inTable 4

41 Time Domain Features Time features are the mostexplicit representation of a signal such as amplitude phaseand even IQ raw data and can be obtained easily In [73]the author uses time domain features such as IQ vectorsand amplitudephase vectors to train CNN classifiers Theresults demonstrate that the proposed scheme is well suitedfor recognizing ZigBee WiFi and Bluetooth signals In


Table 4 Summary of DL in wireless signal recognition

Modulation recognition Wireless technology recognition

Algorithms

(i) DNN [24 30 31 61](ii) DBN [49 63](iii) CNN [17 19ndash21 54 64 65 70 95](iv) LSTM [69](v) CRBM [53](vi) Autoencoder network [50 62](vii) Generative adversarial networks [66 67](viii) HDMF [71 72]

(i) CNN [73ndash76 79 81 82 96](ii) LSTM [29](iii) NFSC [78]

[74] the amplitude and phase difference representation wereemployed for CNN training procedure The results indicatethat the recognition of radar signals has been realized suc-cessfully with the proposed scheme even under the conditionof LTE and WLAN signals coexisting at the same timeReference [75] extended the deep CNN model for radiosignal classification by heavily tuning deep residual networksbased on previous works A discussion was carried on therobustness of the proposed model under different channelparameters and scales of training sets Reference [76] pro-posed a radio fingerprinting method which adopted CNNmodel and IQ dataset for network training Their method iscapable of learning inherent signatures from different wire-less transmitters which are useful for identifying hardwaredevices The results demonstrated that the proposed methodoutperformed ML methods in the recognition of 5 samehardware devices

42 Frequency Domain Features In contrast to time featuresfrequency features contain more useful characteristics suchas bandwidth center frequency and power spectral densitywhich are essential for wireless technology recognition Whatis more effective frequency domain data can greatly alleviatethe problems of data transmission and storage in actualdeployment process

As one of machine learning fuzzy classifier has beenconcerned in wireless technology recognition In [77] anew fuzzy logic (FL) method was presented to recognizeWLAN BT and FSK signals The power spectral density(PSD) information was used to get the bandwidth andcenter frequency for labelling the signals corresponding tostandards Results demonstrated that the proposed strategyis efficient for explicit signal features extraction Likewiseneurofuzzy signal classifier (NFSC) to recognize nanoNETWLAN Atmel and BT signals by utilizing measured PSDinformation is presented in [78] Results show that the perfor-mance is improved by using wideband and narrowband dataacquisition modes in real-time coexistence environments

As a way forward deep learning has also been developedin recent literatures In [73] the time-domain data wasalso mapped into frequency-domain representation by FFTResults show that the CNN model trained on FFT datahas significant improvement in accuracy compared to timedomain features Similarly a reduced CNN model withfrequency-domain data is proposed in [79] The approachis capable of identifying several common IEEE 802x signals

in coexistence Simulations indicate that the performanceof reduced CNN model is superior to most state-of-the-art algorithms In [29] the authors explore the combinationof averaged magnitude FFT processing and LSTM modelwith distributed sensor architecture The result shows thatthe proposed scheme could classify DVB GSM LTE RadarTetra and WFM signals

43 Time-Frequency Domain Features Time-frequency dis-tribution has unique advantages in the comprehensive anal-ysis of signals In [80] frequency-time representations suchas spectral bandwidth temporal width and center frequencyare extracted for neural networks In [81] the authors appliedCNNs to identify IEEE 802x protocol family operating theISM band The time-frequency power values with matrixform were fed into the CNN classifier for data training Theresult indicates that the CNN model outperforms traditionalmachine learning techniques A semisupervised model basedon CNN is developed in [82] a series of time-slices spec-trum data with pseudolabels scheme are trained in CNNExperiments show that this model performs well in devicesrecognition even at the condition of fewer labeled data intraining process

5 Opening Problems in Signal Identification

Although in last decades large amounts of literatures haveproposed various schemes for signal recognition it cannotbe denied that the technology makes little sense in realelectromagnetic environment Many papers are based onideal assumptions such as AWGN channel or enough priorknowledge and the identified signals are limited to a setof several known types DL is the tendency in future butlarge-scale training data are required to get rid of overfittingand attaching precise labels to large-scale data is difficultto accomplish especially in a short time Semisupervisedand unsupervised mechanisms can be explored to alleviatetedious label operations for DL model in the future Moreinnovative technologies and solutions are envisaged in futureresearch

51 Burst Signal Recognition Burst signal has been widelyused in military field and there will be broad prospects fordynamic spectrum access in the future The burst signalhas the characteristics of short duration and uncertainty ofstarting and ending time which poses great challenges to


signal recognition [83ndash85] Generally the processing flow ofsignal recognition includes data acquisition preprocessingclassification and decision The high quality labels have acrucial role for a satisfactory recognition accuracy but arerelatively time-consuming in burst communication systemsTherefore there are still research prospects on the identifyingburst signals

52 Unknown Signal Recognition In the present literaturesthe type of signal to be recognized is assumed to be withina known set and few literatures are capable of coping withunknown signals Even if the spectral characteristics andbackground noise are obtained in practical situations it isunrealistic to know the set of signal types in advance andunexpected interference may occur at any instant [86 87]Many proposed approaches lack universality for the signalsoutside of the set So a comprehensive set of signal typesor a more reasonable model design may be needed for theunknown signal recognition

53 Coexistence Signal Recognition Signal monitoring andinterference management under the condition of multisignalcoexistence have attracted much attention recently [88 89]With the advent of 5G and the IoT commercializationcrowded spectrum resources are likely to lead to an overlap ofmultiple signalsMoreover the increasing types of signals anddiverse wireless networks will undoubtedly bring challengeson the signal recognition in the coexistence environment[90] Predictably the recognition of homogeneous and het-erogeneous signal in an overlapped bandwidth may be anovel development trend of signal recognition and highlyeffective and accurate DL algorithms will create a newcentre

6 Conclusion

In this paper an overview of modulation recognitionand wireless technology recognition is presented Signalrecognition has made considerable progress on both the-ory and practice in traditional fields of device identifica-tion and interference detection It is noticeable that thearea of signal recognition has extended from modulationrecognition to wireless technology recognition With thedevelopment of machine learning and deep learning sim-ple feature extraction in the preprocessing and even rawdata are becoming a trend to realize signal recognitionHowever many works are based on ideal assumption orsome known conditions and most results are obtained bysimulations There is still a long way for signal recog-nition in real electromagnetic environment and practicalapplication

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Authorsrsquo Contributions

Fangwei Dong (dongfangweisrtcorgcn) is the correspond-ing author of this paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (NSFC) under Grants 61801306

References

[1] X Hong J Wang C-X Wang and J Shi ldquoCognitive radio in5G a perspective on energy-spectral efficiency trade-offrdquo IEEECommunications Magazine vol 52 no 7 pp 46ndash53 2014

[2] A A Khan M H Rehmani and A Rachedi ldquoCognitive-radio-based internet of things applications architectures spectrumrelated functionalities and future research directionsrdquo IEEEWireless Communications Magazine vol 24 no 3 pp 17ndash252017

[3] A Ali and W Hamouda ldquoAdvances on spectrum sensingfor cognitive radio networks theory and applicationsrdquo IEEECommunications Surveys amp Tutorials vol 19 no 2 pp 1277ndash1304 2017

[4] T Liu Y Guan and Y Lin ldquoResearch on modulation recog-nition with ensemble learningrdquo EURASIP Journal on WirelessCommunications and Networking vol 2017 no 1 article no 1792017

[5] O A Dobre A Abdi Y Bar-Ness and W Su ldquoSurveyof automatic modulation classification techniques classicalapproaches and new trendsrdquo IET Communications vol 1 no2 pp 137ndash156 2007

[6] L-X Wang and Y-J Ren ldquoRecognition of digital modulationsignals based on high order cumulants and support vectormachinesrdquo in Proceedings of the 2009 Second ISECS Interna-tional Colloquium on Computing Communication Control andManagement CCCM 2009 pp 271ndash274 China August 2009

[7] A Voulodimos N Doulamis A Doulamis and E Protopa-padakis ldquoDeep learning for computer vision a brief reviewrdquoComputational Intelligence and Neuroscience vol 2018 ArticleID 7068349 13 pages 2018

[8] S P Singh A Kumar H Darbari L Singh A Rastogi and SJain ldquoMachine translation using deep learning An overviewrdquoin Proceedings of the 1st International Conference on ComputerCommunications and Electronics COMPTELIX 2017 pp 162ndash167 India July 2017

[9] T Young D Hazarika S Poria and E Cambria ldquoRecent trendsin deep learning based natural language processingrdquo IEEEComputational Intelligence Magazine vol 13 no 3 pp 55ndash752018

[10] T J OrsquoShea J Corgan and T C Clancy ldquoConvolutionalradio modulation recognition networksrdquo in Proceedings ofthe International Conference on Engineering Applications ofNeural Networks vol 629 pp 213ndash226 Springer InternationalPublishing 2016

[11] A Shen Y Liu Y Zhang et al ldquoThe method of interferencerecognition in mobile communication network based on deeplearningrdquo in Signal and Information Processing Networking andComputers vol 494 of Lecture Notes in Electrical Engineeringpp 296ndash306 Springer Singapore 2019

[12] J Schmidhuber ldquoDeep learning in neural networks anoverviewrdquoNeural Networks vol 61 pp 85ndash117 2015


[13] Y Lecun Y Bengio and G Hinton ldquoDeep learningrdquo Naturevol 521 no 7553 pp 436ndash444 2015

[14] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo The American Associa-tion for the Advancement of Science Science vol 313 no 5786pp 504ndash507 2006

[15] M I Jordan and T M Mitchell ldquoMachine learning trendsperspectives and prospectsrdquo Science vol 349 no 6245 pp 255ndash260 2015

[16] C Jiang H Zhang Y Ren Z Han K-C Chen and L HanzoldquoMachine learning paradigms for next-generation wireless net-worksrdquo IEEEWireless Communications Magazine vol 24 no 2pp 98ndash105 2017

[17] Y Wang J Guo H Li L Li Z Wang and H Wang ldquoCNN-based modulation classification in the complicated communi-cation channelrdquo in Proceedings of the 13th IEEE InternationalConference on Electronic Measurement and Instruments ICEMI2017 pp 512ndash516 China October 2017

[18] J Lemley S Bazrafkan and P Corcoran ldquoDeep learning forconsumer devices and services pushing the limits for machinelearning artificial intelligence and computer visionrdquo IEEEConsumer Electronics Magazine vol 6 no 2 pp 48ndash56 2017

[19] K Yashashwi A Sethi andP Chaporkar ldquoA learnable distortioncorrection module for modulation recognitionrdquo IEEE WirelessCommunications Letters pp 1-1 2018

[20] R Li L Li S Yang and S Li ldquoRobust automated VHFmodulation recognition based on deep convolutional neuralnetworksrdquo IEEECommunications Letters vol 22 no 5 pp 946ndash949 2018

[21] S Peng H Jiang H Wang et al ldquoModulation classificationbased on signal constellation diagrams anddeep learningrdquo IEEETransactions onNeural Networks and Learning Systems pp 1ndash102018

[22] O A Dobre and F Hameed ldquoLikelihood-based algorithmsfor linear digital modulation classification in fading channelsrdquoin Proceedings of the 2006 Canadian Conference on Electricaland Computer Engineering CCECErsquo06 pp 1347ndash1350 Canada2006

[23] V G Chavali and C R C M Da Silva ldquoMaximum-likelihoodclassification of digital amplitude-phase modulated signalsin flat fading non-gaussian channelsrdquo IEEE Transactions onCommunications vol 59 no 8 pp 2051ndash2056 2011

[24] B Kim J Kim H Chae D Yoon and J W Choi ldquoDeep neuralnetwork-based automaticmodulation classification techniquerdquoin Proceedings of the 2016 International Conference on Informa-tion and Communication Technology Convergence ICTC 2016pp 579ndash582 Republic of Korea October 2016

[25] E E Azzouz and A K Nandi ldquoProcedure for automatic recog-nition of analogue and digital modulationsrdquo IEE ProceedingsCommunications vol 143 no 5 pp 259ndash266 1996

[26] A K Nandi and E E Azzouz ldquoAlgorithms for automatic mod-ulation recognition of communication signalsrdquo IEEE Transac-tions on Communications vol 46 no 4 pp 431ndash436 1998

[27] F C B FMuller C Cardoso Jr andA Klautau ldquoA front end fordiscriminative learning in automaticmodulation classificationrdquoIEEE Communications Letters vol 15 no 4 pp 443ndash445 2011

[28] T OrsquoShea and J Hoydis ldquoAn introduction to deep learning forthe physical layerrdquo IEEE Transactions on Cognitive Communica-tions and Networking vol 3 no 4 pp 563ndash575 2017

[29] S Rajendran W Meert D Giustiniano V Lenders and SPollin ldquoDeep learning models for wireless signal classification

with distributed low-cost spectrum sensorsrdquo IEEE Transactionson Cognitive Communications and Networking vol 4 no 3 pp433ndash445 2018

[30] F N Khan C H Teow S G Kiu et al ldquoAutomatic modulationformatbit-rate classification and signal-to-noise ratio estima-tion using asynchronous delay-tap samplingrdquo Computers andElectrical Engineering vol 47 pp 126ndash133 2015

[31] FN Khan C Lu andA P T Lau ldquoJointmodulation formatbit-rate classification and signal-to-noise ratio estimation in mul-tipath fading channels using deep machine learningrdquo IEEEElectronics Letters vol 52 no 14 pp 1272ndash1274 2016

[32] D C Chang and P K Shih ldquoCumulants-based modulationclassification technique in multipath fading channelsrdquo IETCommunications vol 9 no 6 pp 828ndash835 2015

[33] H Wu M Saquib and Z Yun ldquoNovel automatic modulationclassification using cumulant features for communications viamultipath channelsrdquo IEEE Transactions on Wireless Communi-cations vol 7 no 8 pp 3098ndash3105 2008

[34] V D Orlic and M L Dukic ldquoAutomatic modulation classifica-tion algorithm using higher-order cumulants under real-worldchannel conditionsrdquo IEEE Communications Letters vol 13 no12 pp 917ndash919 2009

[35] M R Mirarab and M A Sobhani ldquoRobust modulation classi-fication for PSKQAMASK using higher-order cumulantsrdquo inProceedings of the 6th International Conference on InformationCommunications and Signal Processing ICICS pp 1ndash4 2007

[36] O A Dobre Y Bar-Ness and W Su ldquoHigher-order cycliccumulants for high ordermodulation classificationrdquo in Proceed-ings of theMILCOM2003 - 2003 IEEEMilitary CommunicationsConference vol 1 pp 112ndash117 USA 2003

[37] M L D Wong and A K Nandi ldquoAutomatic digital modulationrecognition using artificial neural network and genetic algo-rithmrdquo Signal Processing vol 84 no 2 pp 351ndash365 2004

[38] A Aubry A Bazzoni V Carotenuto A De Maio and P FaillaldquoCumulants-based radar specific emitter identificationrdquo in Pro-ceedings of the 2011 IEEE International Workshop on InformationForensics and Security WIFS 2011 pp 1ndash6 November 2011

[39] M LWong S K Ting andAKNandi ldquoNaive bayes classifica-tion of adaptive broadband wireless modulation schemes withhigher order cumulantsrdquo in Proceedings of the 2008 2nd Inter-national Conference on Signal Processing and CommunicationSystems (ICSPCS 2008) pp 1ndash5 Gold Coast Australia 2008

[40] A Abdelmutalab K Assaleh and M El-Tarhuni ldquoAutomaticmodulation classification using polynomial classifiersrdquo in Pro-ceedings of the 2014 25th IEEE Annual International Symposiumon Personal Indoor and Mobile Radio Communication IEEEPIMRC 2014 pp 806ndash810 USA September 2014

[41] A Abdelmutalab K Assaleh and M El-Tarhuni ldquoAutomaticmodulation classification based on high order cumulants andhierarchical polynomial classifiersrdquo Physical Communicationvol 21 pp 10ndash18 2016

[42] W A Gardner A Napolitano and L Paura ldquoCyclostationarityhalf a century of researchrdquo Signal Processing vol 86 no 4 pp639ndash697 2006

[43] B Ramkumar ldquoAutomatic modulation classification for cogni-tive radios using cyclic feature detectionrdquo IEEE Circuits andSystems Magazine vol 9 no 2 pp 27ndash45 2009

[44] P D Sutton K E Nolan and L E Doyle ldquoCyclostationary sig-natures in practical cognitive radio applicationsrdquo IEEE Journalon Selected Areas in Communications vol 26 no 1 pp 13ndash242008


[45] A Fehske J Gaeddert and JH Reed ldquoA new approach to signalclassification using spectral correlation and neural networksrdquoin Proceedings of the 1st IEEE International Symposium on NewFrontiers in Dynamic Spectrum Access Networks (DySPAN rsquo05)pp 144ndash150 Baltimore Md USA November 2005

[46] K Kim I A Akbar K K Bae J-S Um C M Spooner andJ H Reed ldquoCyclostationary approaches to signal detection andclassification in cognitive radiordquo in Proceedings of the 2007 2ndIEEE International Symposium on New Frontiers in DynamicSpectrum Access Networks pp 212ndash215 Ireland April 2007

[47] X Teng P Tian and H Yu ldquoModulation classification basedon spectral correlation and SVMrdquo in Proceedings of the 20084th International Conference on Wireless Communications Net-working and Mobile Computing (WiCOM) pp 1ndash4 DalianChina October 2008

[48] H Hu Y Wang and J Song ldquoSignal classification based onspectral correlation analysis and SVM in cognitive radiordquo inProceedings of the 22nd International Conference on AdvancedInformationNetworking andApplications (AINA 2008) pp 883ndash887 Gino-wan Okinawa Japan 2008

[49] G J Mendis J Wei and A Madanayake ldquoDeep learning-based automatedmodulation classification for cognitive radiordquoin Proceedings of the 2016 IEEE International Conference onCommunication Systems ICCS 2016 pp 1ndash6 China December2016

[50] J Li L Qi andY Lin ldquoResearch onmodulation identification ofdigital signals based on deep learningrdquo inProceedings of the 2016IEEE International Conference on Electronic Information andCommunication Technology ICEICT 2016 pp 402ndash405 China2016

[51] P R U Lallo ldquoSignal classification by discrete Fourier trans-formrdquo in Proceedings of the Conference onMilitary Communica-tions (MILCOMrsquo99) vol 1 pp 197ndash201 Atlantic City NJ USA1999

[52] Z Yu Y Q Shi and W Su ldquoM-ary frequency shift keyingsignal classification based-on discrete fourier transformrdquo inProceedings of the IEEE Military Communications Conference2003 MILCOM 2003 pp 1167ndash1172 Boston MA USA 2003

[53] L Zhou Z Sun andWWang ldquoLearning to short-time Fouriertransform in spectrum sensingrdquo Physical Communication vol25 pp 420ndash425 2017

[54] Z Liu L Li H Xu and H Li ldquoA method for recognition andclassification for hybrid signals based on deep convolutionalneural networkrdquo in Proceedings of the 2018 International Confer-ence on Electronics Technology ICET 2018 pp 325ndash330 ChinaMay 2018

[55] K C Ho W Prokopiw and Y T Chan ldquoModulation identifica-tion of digital signals by the wavelet transformrdquo IEE Proceedings- Radar Sonar and Navigation vol 147 no 4 pp 169ndash176 2000

[56] X Z Feng J Yang F L Luo J Y Chen and X P Zhong ldquoAuto-matic modulation recognition by support vector machinesusing wavelet kernelrdquo Journal of Physics Conference Series vol48 pp 1264ndash1267 2006

[57] K Hassan I Dayoub W Hamouda and M Berbineau ldquoAuto-matic modulation recognition using wavelet transform andneural networks in wireless systemsrdquo EURASIP Journal onAdvances in Signal Processing vol 2010 Article ID 532898 p42 2010

[58] R A Brown M L Lauzon and R Frayne ldquoA general descrip-tion of linear time-frequency transforms and formulation of

a fast invertible transform that samples the continuous s-transform spectrum nonredundantlyrdquo IEEE Transactions onSignal Processing vol 58 no 1 pp 281ndash290 2010

[59] U Satija M Mohanty and B Ramkumar ldquoAutomatic modula-tion classificationusing S-transform based featuresrdquo in Proceed-ings of the 2nd International Conference on Signal Processing andIntegrated Networks SPIN 2015 pp 708ndash712 India 2015

[60] B G Mobasseri ldquoDigital modulation classification using con-stellation shaperdquo Signal Processing vol 80 no 2 pp 251ndash2772000

[61] A Ali and F Yangyu ldquoUnsupervised feature learning and auto-matic modulation classification using deep learning modelrdquoPhysical Communication vol 25 pp 75ndash84 2017

[62] A Ali F Yangyu and S Liu ldquoAutomatic modulation classifica-tion of digital modulation signals with stacked autoencodersrdquoDigital Signal Processing vol 71 pp 108ndash116 2017

[63] F Wang Y Wang and X Chen ldquoGraphic constellations andDBNbased automaticmodulation classificationrdquo in Proceedingsof the 2017 IEEE 85th Vehicular Technology Conference (VTCSpring) pp 1ndash5 Sydney NSW June 2017

[64] D Wang M Zhang J Li et al ldquoIntelligent constellationdiagram analyzer using convolutional neural network-baseddeep learningrdquo Optics Express vol 25 no 15 pp 17150ndash171662017

[65] S Peng H Jiang H Wang H Alwageed and Y-D YaoldquoModulation classification using convolutional neural networkbased deep learning modelrdquo in Proceedings of the 26th Wirelessand Optical Communication Conference WOCC 2017 pp 1ndash5USA April 2017

[66] B Tang Y Tu Z Zhang and Y Lin ldquoDigital signal modulationclassification with data augmentation using generative adver-sarial nets in cognitive radio networksrdquo IEEE Access vol 6 pp15713ndash15722 2018

[67] Y Tu Y Lin J Wang and J-U Kim Semi-Supervised Learningwith Generative Adversarial Networks on Digital Signal Modula-tion Classification 2018

[68] S Hsue and S S Soliman ldquoAutomaticmodulation classificationusing zero crossingrdquo IEE Proceedings F - Radar and SignalProcessing vol 137 no 6 pp 459ndash464 1990

[69] N E West and T OrsquoShea ldquoDeep architectures for modulationrecognitionrdquo in Proceedings of the 2017 IEEE InternationalSymposium on Dynamic Spectrum Access Networks DySPAN2017 pp 1ndash6 USA March 2017

[70] H Wu Q Wang L Zhou et al ldquoVHF radio signal modulationclassification based on convolution neural networksrdquo MatecWeb of Conferences vol 246 Article ID 03032 2018

[71] D Zhang W Ding B Zhang et al ldquoAutomatic modulationclassification based on deep learning for unmanned aerialvehiclesrdquo Sensors vol 18 no 3 p 924 2018

[72] D Zhang W Ding B Zhang et al Heterogeneous Deep ModelFusion for Automatic Modulation Classification 2018

[73] M Kulin T Kazaz I Moerman and E De Poorter ldquoEnd-to-End learning from spectrum data a deep learning approach forwireless signal identification in spectrum monitoring applica-tionsrdquo IEEE Access vol 6 pp 18484ndash18501 2018

[74] A Selim F Paisana J A Arokkiam Y Zhang L Doyle and LA DaSilva ldquoSpectrum monitoring for radar bands using deepconvolutional neural networksrdquo in Proceedings of the 2017 IEEEGlobal Communications Conference GLOBECOM2017 pp 1ndash6Singapore December 2017


[75] T J OrsquoShea T Roy and T C Clancy ldquoOver-the-air deeplearning based radio signal classificationrdquo IEEE Journal ofSelected Topics in Signal Processing vol 12 no 1 pp 168ndash1792018

[76] S Riyaz K Sankhe S Ioannidis and K Chowdhury ldquoDeeplearning convolutional neural networks for radio identifica-tionrdquo IEEE Communications Magazine vol 56 no 9 pp 146ndash152 2018

[77] K Ahmad U Meier and H Kwasnicka ldquoFuzzy logic basedsignal classification with cognitive radios for standard wirelesstechnologiesrdquo in Proceedings of the 2010 5th InternationalConference on Cognitive Radio Oriented Wireless Networks andCommunications CROWNCom 2010 pp 1ndash5 France June 2010

[78] K Ahmad G Shresta U Meier and H Kwasnicka ldquoNeuro-Fuzzy Signal Classifier (NFSC) for standard wireless technolo-giesrdquo in Proceedings of the 2010 7th International Symposium onWireless Communication Systems ISWCSrsquo10 pp 616ndash620 UK2010

[79] M Schmidt D Block and U Meier ldquoWireless interferenceidentification with convolutional neural networksrdquo in Proceed-ings of the 15th IEEE International Conference on IndustrialInformatics INDIN 2017 pp 180ndash185 Germany July 2017

[80] A N Mody S R Blatt D G Mills et al ldquoRecent advances incognitive communicationsrdquo IEEE Communications Magazinevol 45 no 10 pp 54ndash61 2007

[81] N Bitar S Muhammad and H H Refai ldquoWireless technologyidentification using deep convolutional neural networksrdquo inProceedings of the 28th Annual IEEE International Symposiumon Personal Indoor andMobile Radio Communications PIMRC2017 pp 1ndash6 Canada October 2017

[82] K Longi T Pulkkinen and A Klami ldquoSemi-supervised con-volutional neural networks for identifying wi-fi interferencesourcesrdquo in Proceedings of the Asian Conference on MachineLearning pp 391ndash406 2017

[83] G Baldini R Giuliani and G Steri ldquoPhysical layer authen-tication and identification of wireless devices using the syn-chrosqueezing transformrdquoApplied Sciences vol 8 no 11 p 21672018

[84] J Yang L Yin L Sang X Zhang S You andH LiuA PracticalImplementation of Td-Lte And Gsm Signals Identification viaCompressed Sensing 2017

[85] D A Guimaraes and C H Lim ldquoSliding-window-based detec-tion for spectrum sensing in radar bandsrdquo IEEE Communica-tions Letters vol 22 no 7 pp 1418ndash1421 2018

[86] P C Sofotasios L Mohjazi S Muhaidat M Al-Qutayri andG K Karagiannidis ldquoEnergy detection of unknown signalsover cascaded fading channelsrdquo IEEE Antennas and WirelessPropagation Letters vol 15 pp 135ndash138 2016

[87] J E Friedel T H Holzer and S Sarkani ldquoDevelopmentoptimization and validation of unintended radiated emissionsprocessing system for threat identificationrdquo IEEE Transactionson Systems Man and Cybernetics Systems pp 1ndash12 2018

[88] F Ding A Song D Zhang E Tong Z Pan and X YouldquoInterference-Aware wireless networks for home monitoringand performance evaluationrdquo IEEETransactions on AutomationScience and Engineering vol 15 no 3 pp 1286ndash1297 2018

[89] L Wei E De Poorter J Hoebeke et al ldquoAssessing the coex-istence of heterogeneous wireless technologies with an SDR-based signal emulator a case study of wi-fi and bluetoothrdquo IEEETransactions on Wireless Communications vol 16 no 3 pp1755ndash1766 2017

[90] S Grimaldi A Mahmood and M Gidlund ldquoReal-Time in-terference identification via supervised learning embeddingcoexistence awareness in iot devicesrdquo IEEE Access vol 7 pp835ndash850 2019

[91] M W Aslam Z Zhu and A K Nandi ldquoAutomatic modulationclassification using combination of genetic programming andKNNrdquo IEEE Transactions on Wireless Communications vol 11no 8 pp 2742ndash2750 2012

[92] C Park J Choi S Nah W Jang and D Y Kim ldquoAutomaticmodulation recognition of digital signals using wavelet featuresand SVMrdquo in Proceedings of the 2008 10th International Confer-ence on Advanced Communication Technology vol 1 pp 387ndash390 Gangwon-Do South Korea February 2008

[93] J Lopatka and P Macrej ldquoAutomatic modulation classificationusing statistical moments and a fuzzy classifierrdquo in Proceedingsof theWCC 2000 - ICSP 2000 2000 5th International Conferenceon Signal Processing Proceedings 16thWorld Computer Congress2000 vol 3 pp 1500ndash1506 IEEE Beijing China August 2000

[94] A Abdelmutalab K Assaleh and M El-Tarhuni ldquoAutomaticmodulation classification using hierarchical polynomial classi-fier and stepwise regressionrdquo in Proceedings of the 2016 IEEEWireless Communications and Networking Conference (WCNC)pp 1ndash5 Doha Qatar April 2016

[95] Z Liu L Li H Xu and H Li ldquoA method for recognition andclassification for hybrid signals based on deep convolutionalneural networkrdquo in Proceedings of the 2018 International Con-ference on Electronics Technology (ICET) pp 325ndash330 IEEEChengdu China 2018

[96] S Grunau D Block and U Meier ldquoMulti-label wireless inter-ference classification with convolutional neural networksrdquo inProceedings of the 2018 IEEE 16th International Conference onIndustrial Informatics (INDIN) pp 187ndash192 2018

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom


signal transmission mechanisms it is difficult to recognizecomplex signals by only estimating single or few parametersTherefore the MR and WTR technologies are focused in thispaper

Traditional algorithms of MR could mainly be separatedinto two groups likelihood-based (LB) and feature-based(FB) approaches [5] LB approaches are based on hypoth-esis testing theory the performance based on decision-theoretic is optimal but suffers high computation complexityTherefore feature-based approaches as suboptimal classifierswere developed for application in practice In particularthe feature-based approaches usually extracted features forthe preprocessing and then employed classifiers to real-ized modulation classification Conventional FB approachesheavily rely on the expertrsquos knowledge which may per-form well on specialized solutions but poor in generalityand suffer high complexity and time-consuming To tacklethese problems machine learning (ML) classifiers have beenadopted and shown great advantages for example supportvector machine (SVM) in [6] Although ML methods havethe advantage of classification efficiency and performancethe feature engineering to some extent still depends onexpert experience resulting in degradation of accuracy rateTherefore the self-learning ability is very important whenconfronted with unknown environment A new dawn seemsto have arrived since the DL performs very well in computervision (CV) [7] machine translation (MT) [8] and naturallanguage processing (NLP) [9] DL architecture has alsobeen introduced to MR for instance the convolutionalneural networks (CNN) model is employed in modulationclassification without expert feature extraction [10] whichdemonstrates excellent performance both on efficiency andaccuracy

On the other hand current communication systems tendto be complex and diverse various new wireless technologiesare constantly updated and the coexistence of homogeneousand heterogeneous signals is entering a new norm As aresult the detection and recognition of complex signals willbe confronted with new dilemma and the methods forsignal recognition are needed to keep up with the paceFortunately some literature on WTR has emerged in whichML and DL model with explicit features have also been doneto realize the recognition of specific wireless technologiesMeanwhile some emerging applications are been exploredsuch as base station detection interfere signals recognitionand localization in mobile communication network [11]

In this paper the need of deep learning in signal recog-nition is reviewed in Section 2 Section 3 introduces themodulation recognition with various approaches In Sec-tion 4 wireless technology recognition is presented Openingproblems are discussed in Section 5 Section 6 concludes thepaper

2 Need of Deep Learning in WirelessSignal Recognition

21 Wireless Signal Recognition With the continuous expan-sion of military and civilian needs communication systems

Wireless signal recognition

Modulation recognitionWireless technology recognition

Frequency and bandwidthestimation

Symbol rate evaluation

middot middot middot

Figure 1 Wireless signal recognition

Input Layer Hidden Layer Output Layer

Figure 2 Neural network framework

and electromagnetic environments have greatly changed overthe past decades the capability of detection and recognitionof communication signals has also made significant progressand gradually become maturity Generally communicationsignal recognition takes advantage of some signal parametersto classify or identify the types of signals These techniquesmay include frequency and bandwidth estimation symbolrate evaluation modulation type classification and wire-less technology identification which could be collectivelyreferred to as wireless signal recognition as shown in Figure 1In this paper we are concerned with the current two maintechnologies namely MR and WTR MR commits to realizethe modulation type recognition so as to evaluate wirelesstransmission schemes and device types while WTR takeswireless technology identification as object for improvinginterference management and electromagnetic environmen-tal assessment

22 Definition of DL Problem The concept of DL originatesfrom the research on artificial neural network [12] and thegoal is to understand data by mimicking the mechanism ofthe human brain [13] The basic neural network frameworkconsists of three parts input hidden and output layer andis shown in Figure 2 Hidden layers maybe one layer ormultilayer and each layer consists of several nodes The





Algorithms





Cons



hellip

b

Outputhellip

1

Sum

Ｒ1

Ｒ2

ＲＧ

Ｑ1

Ｑ2

ＱＧ

f(∙)












Transform features

Other features

Feature learning


Cyclostationarity

Fourier transform

Wavelet transform

S transform

Constellation shape

Zero crossing




















119872119901119902 = E [119910 (119899)119901minus119902 (119910lowast (119899))119902] (1)










119878120572119909 (119891) = limΔ119891997888rarrinfin


1Δ119905

sdot intΔ1199052

minusΔ1199052Δ1198911198831Δ119891 (119905 119891 + 1205722) sdot 119883

lowast1Δ119891 (119905 119891 minus 1205722) 119889119905

(2)

where

1198831Δ119891 (119905 V) = int119905+12Δ119891

119905minus12Δ119891119909 (119906) 119890minus1198942120587V119906119889119906 (3)













34 Other Features













Algorithms
















6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia






Algorithms





Cons



hellip

b

Outputhellip

1

Sum

Ｒ1

Ｒ2

ＲＧ

Ｑ1

Ｑ2

ＱＧ

f(∙)












Transform features

Other features

Feature learning


Cyclostationarity

Fourier transform

Wavelet transform

S transform

Constellation shape

Zero crossing




















119872119901119902 = E [119910 (119899)119901minus119902 (119910lowast (119899))119902] (1)










119878120572119909 (119891) = limΔ119891997888rarrinfin


1Δ119905

sdot intΔ1199052

minusΔ1199052Δ1198911198831Δ119891 (119905 119891 + 1205722) sdot 119883

lowast1Δ119891 (119905 119891 minus 1205722) 119889119905

(2)

where

1198831Δ119891 (119905 V) = int119905+12Δ119891

119905minus12Δ119891119909 (119906) 119890minus1198942120587V119906119889119906 (3)













34 Other Features













Algorithms
















6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia








Transform features

Other features

Feature learning


Cyclostationarity

Fourier transform

Wavelet transform

S transform

Constellation shape

Zero crossing




















119872119901119902 = E [119910 (119899)119901minus119902 (119910lowast (119899))119902] (1)










119878120572119909 (119891) = limΔ119891997888rarrinfin


1Δ119905

sdot intΔ1199052

minusΔ1199052Δ1198911198831Δ119891 (119905 119891 + 1205722) sdot 119883

lowast1Δ119891 (119905 119891 minus 1205722) 119889119905

(2)

where

1198831Δ119891 (119905 V) = int119905+12Δ119891

119905minus12Δ119891119909 (119906) 119890minus1198942120587V119906119889119906 (3)













34 Other Features













Algorithms
















6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia









119872119901119902 = E [119910 (119899)119901minus119902 (119910lowast (119899))119902] (1)










119878120572119909 (119891) = limΔ119891997888rarrinfin


1Δ119905

sdot intΔ1199052

minusΔ1199052Δ1198911198831Δ119891 (119905 119891 + 1205722) sdot 119883

lowast1Δ119891 (119905 119891 minus 1205722) 119889119905

(2)

where

1198831Δ119891 (119905 V) = int119905+12Δ119891

119905minus12Δ119891119909 (119906) 119890minus1198942120587V119906119889119906 (3)













34 Other Features













Algorithms
















6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





119878120572119909 (119891) = limΔ119891997888rarrinfin


1Δ119905

sdot intΔ1199052

minusΔ1199052Δ1198911198831Δ119891 (119905 119891 + 1205722) sdot 119883

lowast1Δ119891 (119905 119891 minus 1205722) 119889119905

(2)

where

1198831Δ119891 (119905 V) = int119905+12Δ119891

119905minus12Δ119891119909 (119906) 119890minus1198942120587V119906119889119906 (3)













34 Other Features













Algorithms
















6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





34 Other Features













Algorithms
















6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





Algorithms
















6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia






6 Conclusion






Acknowledgments


References








































































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





























































































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


a survey on deep learning techniques in wireless signal

Documents