a simulation based approach to detect wear in industrial
TRANSCRIPT
A simulation based approach to detect wear in industrial robots
Sathish V1, Srini Ramaswamy2 and Sachit Butail3
Abstract— In this paper we simulate, detect, and locatewear in industrial robots operating over long time-periods onrepetitive tasks. We simulate wear using the MATLAB roboticstoolbox by increasing the Coulomb friction coefficient on selectrobot axes. Disturbance and measurement noise are included ineach axis while computing the inverse and forward kinematics.To compare change across all configurations within a repetitivecycle, the output torque data is transformed into a normalizeddistribution sampled once every hour. We identify wear assignificant change in torque values from normal operatingconditions, and detect wear using principal component analysis;the source axis is located by cross-correlating the output datafrom all axes through time. Simulation results with a rangeof friction values tested for three different tasks across allaxes show that wear detection on the basis of comparison withnormal operating condition depends on the manipulator task,whereby onset of wear was detected in five out of six axes forone task and two out of six axes in the rest. Wear location wascorrectly identified in all successful detections except one.
I. INTRODUCTIONFriction has been reported as the primary cause behind
more than 50% of the faults in heavy industrial robots [1].Friction in robot joints can occur due to inadequate lubri-cation [2], high operating temperatures [3], accumulation ofwear debris [4], and heavy load [3]. Unattended friction canlead to wear and tear, which eventually results in criticaldamage [3], [5]. Friction effects in robotic manipulators arealso widely studied for their importance in determining theoverall performance in terms of accuracy and stability [2],[6], [7], [8].
In industrial robots, wear and tear is typically avoided aspart of preventive maintenance schedules. These schedulesare arrived on the basis of estimated life expectancy ofvarious robot components—the actual condition of the robotis ignored [9], a scenario that leads to high maintenancecosts. Condition-based maintenance is an alternative strategythat uses the data collected from various parts of the robots toassess the condition of the robot. Data from multiple robotsis then subject to conservative thresholds on measurementdata such as torque, temperature and voltage, which are thenused to generate different types of events such as informationevents, warning events, and error events. While low thresh-olds may be used to avoid damage of expensive equipment,
1Sathish V is Associate Principle Engineer at ABB India De-velopment Center and also a graduate student at Indraprastha In-stitute of Information Technology Delhi (IIITD), New Delhi, [email protected]
2Srini Ramaswamy is Global R&D Project Manager-Engineering (PowerGeneration) at ABB. [email protected]
3Sachit Butail is an Assistant Professor in the Department ofElectronics and Communication Engineering, Indraprastha Instituteof Information Technology Delhi (IIITD), New Delhi, [email protected]
these are bound to raise the rate of false alarms from error orwarning events leading to consumer complacency and misseddetections [10].
Fault monitoring of industrial systems is performed usinga wide variety of techniques that can be broadly classifiedinto knowledge-based, model-based, and data-driven meth-ods [11], [12], [13]. Among these, knowledge-based methodsrely on user expertise to arrive at customized thresholdsfor output signals that are input to a monitoring system[11]. These are specific to robot types and the operatingconditions and are therefore non-transferable across differentsituations [10]. Model-based methods rely on friction modelsto predict the occurrence of a fault [9], [14]. In [14],for example, a friction model is used to estimate physicalparameters of an industrial robot which are then utilizedfor fault diagnosis. While modeling approaches focus onan accurate representation of friction effects [7], [15], theirdirect application in fault detection and prediction requiresthe additional step of calibrating the robot to determine robot-specific model parameters, and may therefore be difficultto implement on industrial systems that have a diverse lineup [16]. Instead, data-driven approaches rely on statisticalmethods to characterize normal and faulty situations. Whiledata-driven methods are designed for model independence,they require high-volume data inputs to perform reliably [10],[17], [18], and rely on the variability in the training datafor an exhaustive fault detection framework [19]. Followinga recently proposed approach [20], we use an approximatemodel to generate high-volume synthetic data of wear dueto low maintenance across multiple scenarios. The resultingdata is used to develop and validate a wear detection andidentification approach.
We model wear as a slow exponential growth in Coulombfriction spread over a period of five days to simulate in-stances of oversight in maintenance [4]. Realistic values ofdisturbance and measurement noise are introduced [21]. Toallow detection and location of wear irrespective of the jointconfiguration, we transform the periodic torque values into anormalized distribution, and compute Kullback-Leibler (KL)divergence from the initial distribution as the new outputdata. The resulting output for all axes is then subject toprincipal component analysis (PCA) to compare with normaloperating conditions. Wear is identified as an error in thedata fit beyond a fixed confidence interval and located bycross-correlating the output data between axes. The approachis tested by simulating an increase in friction in all of sixrobot axes, one at a time, in three different tasks simulatedusing the open source MATLAB robotics toolbox [22]. Thecontributions of this paper are: (i) we add the capability
2015 IEEE International Conference onAutomation Science and Engineering (CASE)Aug 24-28, 2015. Gothenburg, Sweden
978-1-4673-8183-3/15/$31.00 ©2015 IEEE 1570
of simulating wear due to friction in industrial robots inan open source software [22]; (ii) we implement PCA andcross-correlation based approach to detect the onset of wearin robots and identify the source location axis; (iii) wevalidate and analyse the proposed detection and identificationapproach on simulated data.
This paper is organized as follows: Section 2 gives a back-ground of manipulator dynamics, PCA, and cross-correlationas a measure of detecting the delay between two signals.Section 3 describes the approach for simulating a gradualincrease in friction in the presence of disturbance and noise,and its implementation on the MATLAB robotics toolbox[22]. Section 4 presents the wear detection and sourceidentification approach. Section 5 provides results from datagenerated from an exhaustive set of experiments performedon three different tasks and finally, Section 6 summarizes thefindings and discusses future work.
II. BACKGROUND
A. Manipulator Dynamics
The simulation model used in this study is based on asix degrees-of-freedom (dof) manipulator called PUMA 560(Programmable Universal Machine for Assembly). Given thejoint angles q ∈ R6, velocity q ∈ R6 and acceleration q ∈R6, the required torque τ ∈ R6 is given by the dynamicmodel
τ = M(q)q + C(q, q) + F (q) +G(q), (1)
where M,C,F and G are inertia matrix, coriolis and cen-trifugal force matrix, friction matrix, and the gravity torquematrix respectively.
Equation (1) can be represented in a discrete state-spaceform for a single axis i with the state variable xi =[qi, qi, qi]T, ∈ R3 as
xik+1 = Aixik +Biτ ik + ωik
yik = Cixik + νik,(2)
where τ ik ∈ R1 is the input torque at time step k, yi =qi ∈ R1 is the measured output, ωik ∈ R1 is disturbanceand νik ∈ R3 is the measurement noise. Ai, Bi, and Ci areknown system matrices [23], [24].
B. Principal Component Analysis and Cross-correlation
We use PCA and cross-correlation to detect and locateonset of wear.
Principal Component Analysis: PCA is a multivariateanalysis technique [10] used to classify patterns in data[25]. Briefly, PCA finds principal components (PCs) that canbest approximate the variation in original data [26]. We usePCA to detect a change in the overall torque values overextended periods of time. Given n torque values from allsix robot axes, a torque matrix is constructed as T ∈ Rn×6.PCA eigen-decomposes the matrix T into eigenvectors Pm∈ R6×m, and scores matrix Tm ∈ Rn×m, as [27]
T = TmPᵀm + E, (3)
where m is selected so that the principal componentsaccount for more than seventy five percent of the variationin data [28], and E is the residual error matrix.
The accuracy with which new data can be projected ontothe principal components represents the degree of data fitand can be calculated using Q statistic [29]
Q = T (I − PmPTm)T ᵀ (4)
In our case, we use PCA to detect onset of wear by settinga threshold on Q-statistic. In other words, an onset of wear isidentified if the Q value rises above a data-driven threshold[30]. We use 95% confidence level to set threshold on theQ-statistic [31].
Cross-correlation: Cross-correlation compares twosignals at different time lags by correlating them. GivenN pairs of observations of two time series τ i and τ j , thecross-covariance at lag δt is [32]
c(δt)τ i,τj =
{∑N−δtt=1 (τ it − τ i)(τ
jt+δt − τ j)/N, δt ≥ 0∑N
t=1−δt(τit − τ i)(τ
jt+δt − τ j)/N, δt < 0
(5)and the cross-correlation coefficient for the time series τ i,
τ j , is
R(δt)τ i,τj =c(δt)τ i,τj
sτ isτj
, (6)
where τ i, τ j are the average of series τ i and τ j ; sτ i , sτj
are the standard deviation of τ i and τ j , respectively.The time delay δt at which the two signals give the
maximum cross-correlation R(δt)τ i,τj corresponds to the lagat which the signals are best aligned. In the case of twosimilar signals, δt at maximum R corresponds to the delaybetween them. A positive value of δt indicates that τ j istrailing τ i; conversely, a negative value indicates that τ j isleading τ i. We use the delay in cross-correlation to determinethe direction of fault propagation.
III. SIMULATING WEAR INDUCED DUE TO FRICTION
To simulate wear, we increase Coulomb friction coeffi-cient over an extended period of time (5 days) representingoversight in scheduled maintenance. The increase is set upso that the maximum torque stays within the recommendedoperational limits.
Denoting the Coulomb friction coefficient for positive andnegative joint velocity as µ+ and µ−, the change in frictionwith time k is modeled as
µ+k = r+0 +
c11 + exp(−(c2k)4)
µ−k = r−0 −
c11 + exp(−(c2k)4)
,(7)
where r+0 and r−0 are set so that at k = 0, the initial valuesof Coulomb friction, µ+
0 and µ−0 are equal to the default
values available from the PUMA560 Robotics toolbox [22],and c1 and c2 are constants computed by solving (7) forincrease in friction at the end of the five days. For each k, the
1571
updated value of Coulomb friction coefficient is substitutedin equation (1) to arrive at Friction matrix F [22]
F (q) = V q + µ, (8)
where V denotes the coefficient of viscous friction. Figure1 shows a sample change in friction torque as a result ofchange in Coulomb friction using (7) (Note that disturbanceand noise are added in the same sample as per Section IIIB).
0 10 20 30 40 50 60 70 80 90 100 110 12027.5
28
28.5
29
29.5
30
Time (hours)
Fric
tona
l Tor
que
(Nm
)
Student Version of MATLAB
Fig. 1. Change of frictional torque due to increase in Coulomb friction.Increase in frictional torque in a single axis as the Coulomb friction isincreased up to 10% of it’s initial value over 5 days (120 hours).
A. Implementation on the robotics toolbox
The MATLAB robotics toolbox is a set of open sourceMATLAB functions that allow ease of simulation and analy-sis of robotic manipulators [22]. The toolbox consists of user-defined functions that can help visualize and create a robottrajectory, compute input torque using inverse dynamics (1),and compute output torque using forward dynamics [22].We use the robotic toolbox to simulate increase in Coulombfriction along with disturbance and measurement noise in aPUMA 560 manipulator.
B. Adding disturbance and measurement noise
In order to represent a realistic scenario where disturbanceand measurement can cause high false alarm rates, we adddisturbance and measurement noise to the simulations ofmanipulator tasks. In particular, unknown disturbances mayoccur during the operation of a robotic manipulator due tominor variations in load torques and change in lubricanttemperatures [19]. Measurement noise may occur due tosensor degradation and high or low operating temperatures.
Disturbance ωk in (2) is added in the form of a zero-mean Gaussian process with unit standard deviation scaledto one percent of the average torque [33]. Using (2) asreference, here, we assume C ∈ R3×3 to be an identitymatrix. Similarly as in [33], [34] the measurement noise νk isalso modeled as a zero mean Gaussian process. To obtain thenoise levels in q and q, νk is then scaled with the inverse ofthe sampling time-step δt and δt2 respectively. The modifiedmeasurement equation is
yik = Cxik +
11δt1δt2
νk, (9)
where C ∈ R3×3 is an identity matrix, νk is 0.0001 rad,based on the average accuracy of the encoder in robotic
manipulators [21] and δt = 0.5 s. Once y is obtained, weuse inverse dynamics (1) to obtain the measured torque.
TABLE IALGORITHM TO SIMULATE WEAR DUE TO FRICTION
Input: Start and end position of trajectory, duration of operation, timestep between successive operation, and torque operationallimits
For each time step k
1) Generate joint position,velocity, and acceleration, q, q, q using for-ward kinematic equation (jtraj)
2) Calculate torque for each time step using inverse dynamic equation(rne). Introduce friction increase by modifying Coulomb frictioncoefficient (7) along with disturbance as in (2) modeled as zero meanGaussian process
3) Compute q, q, q for each axis.4) Add measurement noise as per equation (9) to q, q, q5) Calculate measured torque using inverse dynamic equation (rne)
Table I lists the schematic of the algorithm for simulatingwear in robots using the MATLAB robotics toolbox [22].
IV. DETECTING AND LOCATING WEAR
In this section, we describe the wear detection and locationmethod. The torque in an axis is dependent on configurationand may vary drastically depending on the task to be per-formed. Sampling the torque at fixed intervals of time doesnot guarantee measuring it at the same configuration becauseof changes in velocity and acceleration, again due to friction.Further, sampling the torque at a fixed configuration mayalso lead to completely ignoring a problematic configurationof joints. To make the comparison of torque independentof configuration, we implement the proposed wear detectionapproach on the KL distance [19] of the normalized torquedistribution. The torque is sampled every ten seconds for thefirst ten minutes of every hour to obtain sixty values. Thisensures that we obtain the torque values for six possibleconfigurations in a trajectory that repeats itself once everyminute. These sixty values are then normalized with respectto maximum torque and binned into twelve equal-sizedintervals. The KL distance measures the similarity betweentwo distributions [35], [36], and therefore in our case servesthe compare full trajectories within a cycle instead of singleconfigurations.
A. Identification of onset of wearWe use PCA based data-driven approach that does not
depend on predefined thresholds to detect changes in axistorque due to increase in friction. In particular, we use thetorque distribution in normal operation as training data togenerate principal components using (3). Torque from allsix axes are used together to create a single normal-workingprofile. Fig. 2 shows the KL distance of each axis from areference distribution. The number of PCs are selected asdescribed in Section II-B and ranges between 3-4 in ourcase. Anomalies are detected by checking if the data fallswithin the thresholds defined using the Q-statistic (4). Figure3 shows a sample training and testing procedure whereone-hundred and twenty hours of normal working operation
1572
is used to obtain the output data for training. Friction isintroduced at the hundred and twenty first hour and detectedusing the Q statistic threshold.
0 10 20 30 40 50 60 70 80 90 100 110 1200
2
4
6
8
10KL Divergence Vs Samples
Time (hours)
K-L
dist
ance
(nat
) Axis 1Axis 2Axis 3Axis 4Axis 5Axis 6
Student Version of MATLAB
Fig. 2. KL Distance Vs Time. KL Distance is calculated between the initialdistribution and distribution of torque sampled once every hour. Frictionincrease of 10% is introduced in axis 1.
0 20 40 60 80 100 120 140 160 180 200 220 24010
-5
100
105
Training Testing
Q R
esid
uals
(log s
cale
)
Calibration
Test
Model-specific Q Limit
95% Confidence Level
Time (hours)
Student Version of MATLAB
Fig. 3. PCA analysis based on KL distance. PCA model is developedwith a five day training data (0-120 hours on the left) and tested the modelwith a 10 % increase in friction over five days (120-240 hours). The onsetof wear is detected at approximately 3 days after introducing the friction.Q-Residuals show the goodness of fit of the PCA model to test data.
B. Identification of location of wear
We cross-correlate the output data (torque distribution foreach axis across time) to locate the source axis of the wear.In particular, using lag at maximum cross-correlation as anindicator of leading or trailing occurrence of a change intorque, we identify the source axis as one that has all positivelags.
TABLE IILAG (HOURS) AT MAXIMUM CORRELATION BETWEEN AXES
Axis 1 Axis 2 Axis 3 Axis 4 Axis 5 Axis 6
Axis 1 0 -3 -3 -27 0 -4Axis 2 3 0 0 0 -38 -2Axis 3 3 0 0 0 6 0Axis 4 27 0 0 0 -22 5Axis 5 0 38 -6 22 0 32Axis 6 4 2 0 -5 -32 0
Table I shows the lag (in hours) at maximum cross-correlation for all pairs of axes where the friction wasincreased in axis 1. The output from axis 1 has a positivelag with all other axes. For a single change in output datathrough time, this shows that axis 1 was first to register achange, thus indicating the source of wear. Fig. 2 shows theKL distance of each axis from a reference distribution.
V. EXPERIMENTS AND ANALYSIS
The details of simulation experiments is given in Table III.We introduce increase in friction corresponding to 10% and20% change over five days (this resulted in the overall torquenot exceeding more than 50% of the maximum alloweddesign torque for any axis). We simulate the friction changein all six axes, one axis at a time.
-1 0 1-1
-0.5
0
0.5
1Top view
x-position
y-p
ositio
n
-1 0 1-1
-0.5
0
0.5
1Side view
x-position
z-p
ositio
n
Student Version of MATLAB
(a) Task 1
-1 0 1-1
-0.5
0
0.5
1Top view
x-position
y-p
ositio
n
-1 0 1-1
-0.5
0
0.5
1Side view
x-position
z-p
ositio
n
Student Version of MATLAB
(b) Task 2
-1 0 1-1
-0.5
0
0.5
1Top view
x-position
y-p
ositio
n
-1 0 1-1
-0.5
0
0.5
1Side view
x-position
z-p
ositio
n
Student Version of MATLAB
(c) Task 3
Fig. 4. End-effector trajectories for three different manipulator tasks. Topand side view of a single run of the trajectories are shown for each task.The tasks represent simple pick-and-place operations. Green circle denotesstart and red circle denotes the stop position
TABLE IIIEXPERIMENTAL DESIGN
Trajectory % Friction Increase No. of simulation runs
1 10 61 20 62 10 62 20 63 10 63 20 6
We designed three different tasks to exhaustively validatethe detection and location approach described in this paper.The tasks represent pick-and-place operations and involve therobot end-effector reaching its workspace limits and rotationof all joints (Fig. 4). Duration of one cycle for each task isset to one minute repeated over five days (120 hours). The
1573
top and side view of the end-effector for each of the threetrajectories are shown in Figure 4.
Table IV shows the detection and source location perfor-mance for all thirty-six simulations. Onset of wear on thebasis of a data-driven threshold using PCA is both detectedand located in all axes except axis 2 for task 2, two axesfor task 1, one axis for task 3 with 10 % friction increase,and four axes in task 3 with 20% friction increase. While weincreased the friction coefficient by the same amount in allaxes, the maximum absolute torque at the time of detectionor at the end of simulation period (Table V) shows that theincrease does not have the same effect on all axes. Betweenall tasks, task 2 has the lowest loads for axes two and three ofthe manipulator, and has the highest load for the sixth axis.At the same time, task 2 has a low coefficient of variation(Table VI) for most of the axes between all tasks. This showsthat the detection of wear is dependent on the torque loadand variability on an axis. Between axes, 4-6 are closer toend-effector and their movement and rotation depends on thetask at hand and it is likely that the disturbance and noise,also introduced into the training data, do not allow the changedue to increase in friction to manifest.
TABLE IVWEAR DETECTION AND LOCATION WITH THE NUMBER OF HOURS AT
THE TIME OF DETECTION
Friction Increase 10% Friction Increase 20%Axis T1 T2 T3 T1 T2 T3
1 (60) (60) × (50) (50) (40)2 × × × × × ×3 (100) (90) × (85) (78) (105)4 × (80) × × (60) (90)5 × (115) × × (95) ×6 × (90) (95) × (90) (90)
Task 2 trajectory is different from that of tasks 1 and 3in terms of the region of workspace it explores. The lowcoefficient of variations could possibly result from a simplerpath that has less workspace region to cover compared totask 1 and task 3. Both tasks 2 and 3 involve the rotation ofsixth axis, an effect of which is indicated in the similaritybetween the absolute torque values for axis 6 for these twotasks as compared to task 1.
TABLE VMAXIMUM ABSOLUTE TORQUE (Nm) AT THE TIME OF DETECTION OR
AT THE END OF FIVE DAYS IF WEAR WAS NOT DETECTED
Friction Increase: 10% Friction Increase: 20%Axis T1 T2 T3 T1 T2 T3
1 28.520 29.705 29.070 29.724 29.475 28.7992 81.842 64.176 71.769 82.533 64.665 71.7353 26.240 9.248 25.620 26.421 9.484 25.4424 2.292 1.381 2.226 2.292 1.409 2.2495 2.434 3.183 2.822 2.446 3.231 2.8036 0.000 0.912 0.834 0.000 0.928 0.873
The detection and location of the onset of wear is thesame for tasks 1 and 2 for both ten and twenty percentincrease in friction with the detection taking place earlierfor twenty percent than ten percent. For task 3, the wear is
detected in three additional axes when the friction increaseis 20 percent. This is expected since the anomalous changein torque distribution will be quicker with a larger increasein friction over the same time period. The time of detectionitself indicates that disturbance and noise can overshadow anincrease in friction for over three days before any change intorque can be detected.
TABLE VICOEFFICIENT OF VARIATION OF ABSOLUTE TORQUE (Nm) AT THE TIME
OF DETECTION OR AT THE END OF FIVE DAYS
Friction Increase 10% Friction Increase 20%Axis T1 T2 T3 T1 T2 T3
1 0.733 0.049 0.047 0.733 0.048 0.0512 0.210 0.227 0.230 0.225 0.233 0.2333 0.766 0.254 0.350 0.773 0.242 0.3514 0.625 0.192 0.824 0.625 0.189 0.7985 0.758 0.614 0.675 0.757 0.607 0.6776 0.731 0.409 0.445 0.692 0.405 0.422
Studies on fault detection and prediction using simulation,focus on introducing faults artificially by incorporating sud-den load changes in operation [20], [37]. In [20], two typesof faults: actuator fault and sensor fault, are introduced anddetected immediately using statistical differences. The sourceof the actuator fault is identified by comparing regressioncoefficients between normal and faulty condition. In [37],actuator faults in two-link manipulators are detected withnegligible delay in all of the nine test cases. In contrast, thewear simulated and detected in this paper does not createsudden changes in manipulator performance but indicativeof an alarm event. Accordingly, although the onset of wearis detected hours after it is introduced, the actual fault mayappear still later. Finally, using our approach, the source axiswas correctly identified in all cases where the onset of wearwas detected.
To ensure that the algorithm can be implemented inpractice, we ensured that the measurement data matchesthe format of the output from the remote servicing agentand the normal operating condition is characterized on thebasis of sixty days of normal operation, a time period forwhich we expect data to be easily available. The computationof secondary measures such as KL divergence and cross-correlation lags is computationally simple and can be imple-mented on the remote servicing agent. Finally, open issuesinclude identifying optimal thresholds that likely depend onrobot and environment specific disturbance and measurementnoise values for usable fault time prediction.
VI. CONCLUSIONSIn this paper, we present an approach to detect and locate
the onset of wear in industrial robots operating over longtime periods. Wear is simulated on a PUMA 560 robotusing MATLAB Robotics toolbox by increasing Coulombfriction in the presence of process and measurement noise.PCA and cross-correlation is used to create data-driventhresholds for detecting wear, and locating the source axis.Multiple simulation experiments are conducted on output
1574
data represented in the form of torque distribution data andcompared during periodic cycles.
Our results show that the detection and location of weardepends on the manipulator task. Low variability in load andmovement of axis that may result from different trajectoriesaffect the change in torque as a result of adding friction.Further, given disturbance and noise in the system, the timeof detection shows that the onset of wear may take morethan three days to manifest.
Future studies will focus on two directions: first, onvalidation of this approach on real data collected from thefield, and second on developing methods to characterize andidentify the cause behind an increase in friction that mayresult from multiple sources [2], [3], [4].
ACKNOWLEDGMENTThis work was supported by ABB, India Development
Center (IDC). Authors are thankful to Dr. Mikael Norrlof(ABB Sweden) for discussions related to remote servic-ing agent operation, Dr. Magnus Larsson (ABB India) forreading and commenting on the manuscript, Mr. RajeshwarDatta (ABB India) for discussions on related work, and theanonymous reviewers for their constructive comments.
REFERENCES
[1] M. R. Kermani, M. Wong, R. V. Patel, M. Moallem, and M. Ostojic,“Friction compensation in low and high-reversal-velocity manipu-lators,” in Proceedings of the IEEE International Conference onRobotics and Automation (ICRA). IEEE, 2004, pp. 4320–4325.
[2] B. Bona and M. Indri, “Friction compensation in robotics: anoverview,” in Proceeding of the 44th IEEE Conference on Decisionand Control, and European Control Conference, 2005, pp. 4360–4367.
[3] A. C. Bittencourt, “On modeling and diagnosis of friction and wear inindustrial robots,” Ph.D. dissertation, Linkoping University, Linkoping,January 2012.
[4] A. Bittencourt, “Friction change detection in industrial robot arms,”Masters Degree Project, KTH electrical engineering, Stockholm, 2007.
[5] K. Kato, “Wear in relation to frictiona review,” Wear, vol. 241, no. 2,pp. 151–157, 2000.
[6] H. Olsson, K. A strom, C. Canudas de Wit, M. Gafvert, and P. Lischin-sky, “Friction Models and Friction Compensation,” European Journalof Control, vol. 4, no. 3, pp. 176–195, Jan. 1998.
[7] A. C. Bittencourt, E. Wernholt, S. Sander-Tavallaey, and T. Brogardh,“An extended friction model to capture load and temperature effects inrobot joints,” in Proceeding of the IEEE/RSJ International Conferenceon Intelligent Robots and Systems (IROS), 2010, pp. 6161–6167.
[8] P. Hamon, M. Gautier, and P. Garrec, “Dynamic identification ofrobots with a dry friction model depending on load and velocity,”in Proceedings of IEEE/RSJ International Conference on IntelligentRobots and Systems (IROS), 2010, pp. 6187–6193.
[9] A. C. Bittencourt and P. Axelsson, “Modeling and experiment designfor identification of wear in a robot joint under load and temperatureuncertainties based on friction data,” IEEE/ASME Transactions onMechatronics, vol. 19, no. 5, pp. 1694–1706, 2013.
[10] O. Pettersson, “Execution monitoring in robotics: A survey,” Roboticsand Autonomous Systems, vol. 53, no. 2, pp. 73–88, Nov. 2005.
[11] M. L. Visinsky, J. R. Cavallaro, and I. D. Walker, “Robotic faultdetection and fault tolerance: A survey,” Reliability Engineering &System Safety, vol. 46, no. 2, pp. 139–158, 1994.
[12] H. Sneider and P. M. Frank, “Observer-based supervision and fault de-tection in robots using nonlinear and fuzzy logic residual evaluation,”IEEE Transactions on Control Systems Technology, vol. 4, no. 3, pp.274–282, 1996.
[13] L. H. Chiang, R. D. Braatz, and E. L. Russell, Fault detection anddiagnosis in industrial systems. Springer, 2001.
[14] B. Freyermuth, “An approach to model based fault diagnosis ofindustrial robots,” in Proceeding of IEEE International Conferenceon Robotics and Automation. IEEE, 1991, pp. 1350–1356.
[15] L. Marton, “Energetic approach to deal with faults in robot actuators,”in Proceeding of the 20th Mediterranean Conference on & Automation(MED). IEEE, 2012, pp. 85–90.
[16] A. Robotics, “Product range: Improving productivity, quality andworkplace safety,” ABB Robotics, 2010.
[17] G. Vachtsevanos, F. Lewis, M. Roemer, A. Hess, and B. Wu, “Faultdiagnosis and prognosis performance metrics,” Intelligent Fault Diag-nosis and Prognosis for Engineering Systems, pp. 355–399.
[18] M. H. Terra and R. Tinos, “Fault detection and isolation for roboticsystems using a multilayer perceptron and a radial basis functionnetwork,” in Proceedings of the IEEE International Conference onSystems, Man, and Cybernetics. IEEE, 1998, pp. 1880–1885.
[19] A. C. Bittencourt, K. Saarinen, S. Sander-Tavallaey, S. Gunnarsson,and M. Norrlof, “A data-driven approach to diagnostics of repetitiveprocesses in the distribution domain–applications to gearbox diagnos-tics in industrial robots and rotating machines,” Mechatronics, vol. 24,no. 8, pp. 1032–1041, 2014.
[20] R. Muradore and P. Fiorini, “A pls-based statistical approach for faultdetection and isolation of robotic manipulators,” IEEE Transactionson Industrial Electronics, vol. 59, no. 8, pp. 3167–3175, 2012.
[21] B. W. Mooring, Z. S. Roth, and M. R. Driels, Fundamentals ofmanipulator calibration. John Wiley & Sons, NY, 1991.
[22] P. I. Corke et al., Visual Control of Robots: high-performance visualservoing. Research Studies Press Taunton, UK, 1996.
[23] J. Choi, S. Kang et al., “External force estimation using joint torquesensors for a robot manipulator,” in Proceedings of the IEEE Inter-national Conference on Robotics and Automation (ICRA), 2012, pp.4507–4512.
[24] R. Johansson, A. Robertsson, K. Nilsson, and M. Verhaegen, “State-space system identification of robot manipulator dynamics,” Mecha-tronics, vol. 10, no. 3, pp. 403–418, 2000.
[25] M. S. Pan-Ning Tan, Vipin Kumar, Introduction to Data Mining.Pearson, 2013.
[26] D. Pan, Z. Liu, L. Zhang, Y. Liu, and M. J. Zuo, “A study onapplications of principal component analysis and kernel principalcomponent analysis for gearbox fault diagnosis,” in Proceedings of theInternational Conference on Quality, Reliability, Risk, Maintenance,and Safety Engineering (QR2MSE). IEEE, 2013, pp. 1917–1922.
[27] J. MacGregor and T. Kourti, “Statistical process control of multivariateprocesses,” Control Engineering Practice, vol. 3, pp. 403–414, 1995.
[28] R. M. Penha and J. W. Hines, “Using principal component analysismodeling to monitor temperature sensors in a nuclear research re-actor,” in Proceedings of the maintenance and reliability conference(MARCON). Citeseer, 2001, pp. 6–9.
[29] Q. Chen, U. Kruger, M. Meronk, and A. Leung, “Synthesis of T 2 andQ statistics for process monitoring,” Control Engineering Practice,vol. 12, no. 6, pp. 745–755, 2004.
[30] V. Sathish, S. D. Sudarsan, and R. Srini, “Event based robot prog-nostics using principal component analysis,” in Proceedings of IEEEInternational Symposium on Software Reliability Engineering (ISSRE).IEEE, 2014, pp. 14–17.
[31] J. Devore, Probability and Statistics for Engineering and the Sciences.Thomson Learning, 2002.
[32] C. Chatfield, The Analysis of Time Series, An Introduction. Chapmanand Hall, 2012.
[33] S. J. Qin and T. A. Badgwell, “A survey of industrial model predictivecontrol technology,” Control engineering practice, vol. 11, no. 7, pp.733–764, 2003.
[34] M. Athans, “The role and use of the stochastic linear-quadratic-gaussian problem in control system design,” Automatic Control, IEEETransactions on, vol. 16, no. 6, pp. 529–552, 1971.
[35] T. Dasu, S. Krishnan, S. Venkatasubramanian, and K. Yi, “Aninformation-theoretic approach to detecting changes in multi-dimensional data streams,” in In Proc. Symp. on the Interface ofStatistics, Computing Science, and Applications. Citeseer, 2006.
[36] A. Bittencourt, K. Saarinen, S. Sander Tavallaey et al., “A data-drivenmethod for monitoring systems that operate repetitively–applicationsto wear monitoring in an industrial robot joint,” in Fault Detection,Supervision and Safety of Technical Processes, vol. 8, no. 1, 2012, pp.198–203.
[37] W. E. Dixon, I. D. Walker, D. M. Dawson, and J. P. Hartranft,“Fault detection for robot manipulators with parametric uncertainty: aprediction-error-based approach,” IEEE Transactions on Robotics andAutomation, vol. 16, no. 6, pp. 689–699, 2000.
1575