Multisensor Process Performance Assessment Through Use of Autoregressive Modeling and Feature Maps

Nicolas Casoetto, Dragan Djurdjanovic, Rhett Mayor, and Jun Ni, Dept.of Mechanical Engineering, University of Michigan, Ann Arbor, Michigan, USA Jay Lee, Industrial and Manufacturing Engineering Dept., University of Wisconsin-Milwaukee, Milwaukee, Wisconsin, USA

Abstract
This paper presents an algorithm for quantitative assessment of process or machine performance. The algorithm is based on the fusion of multiple sensor inputs and matching of the currently observed system signatures against those observed during its normal behavior. In contrast to traditional monitoring techniques, neither faulty data nor historical data are needed for the performance assessment to be made. This quantitative information about process performance can be used to take appropriate maintenance action in a timely manner. Effectiveness of the methods presented in this paper was experimentally demonstrated in multisensor welding process assessment.
Keywords: Continuous Process Based Assessment, Multisensor Modeling, Feature Maps Assessment, ConditionFusion, Autoregressive

Introduction
In today’scompetitivemarket, production costs,lead time, and optimal machine utilization are crucial values for companies.It is well known that machine or processbreakdownsseverelylimit a company’s effectiveness.Scheduledmaintenanceis intended to eliminate machine or process breakdowns through scheduledmaintenanceoperations, regardless of the actual machine/process state.Actual time intervals in which maintenanceis performed are determinedusing the reliability theory andmachine or processlife cycle information. Nevertheless, practice often resultsin this unnecessary loss of productivity due to maintenance performedwhen the processor machine is still able to perform at an acceptablelevel, or due to unpredicted breakdownsbefore the scheduled maintenanceoperation is performed.:
This paper is an original work and has not been previously published except in the Trunsuctionx of NAMRUSME, Vol. 31. 2003.

Nowadays, increasingly sophisticated sensorsare installed on machines more and more often, while the speed of computers increases in great leaps. Under these circumstances, condition-based maintenance (CBM) emerges as a more appropriate and efficient tool for achievingnear-zerobreakdowntime through a significant reduction and, when possible, elimination of the downtime due to process/machine failure or any other unacceptablebehavior. According to Engel et al. (2OOO), two methods exist for condition-based assessment (CBA). One approach involves an estimation of current condition based on the recognition of indications of failure. When implementing this CBA paradigm, expert and a priori knowledge about the assessed machine or process is necessarybecauseone needsto know the corresponding failure modes to assessthe current machine or processperformance by comparing it against the performance that characterizesthose known failure modes;hence,the application-specific nature of this CBA paradigm. For example, Swanson (2001) used a fuzzy logic neural network to predict failure of a tensioned steel band with seededcrack growth. Ray and Tangirala (1996) built a stochastic model of fatigue crack dynamics in mechanicalstructures to predict remaining service time. Vachtsevanos and Wang (2001) give an overview of different CBA algorithms and suggest a method to compare their performance for a specific application. The second type of CBA characterizes the machine or process performance through the overlap of the most recent system signatures and those observed during the healthy process or machine behavior. In this paper,the latter type of CBA has been implemented to shift the maintenancefocus from the analysis and quantification of process/machinefail-

64

enabling a generic approach to a wide range of applications. The system performance is described using generic signal features.it is crucial to correctly select the signatures representing this normal operation state. However. Methods General Description of the Methods The performance assessment methodology proposed in this paper is basedon assessing overlap the between the signatures describing the current machine or process performance and those observed during the normal system operation. a generic sensorfusion based on statistical reasoning was accomplished through the merger of sensor features. As a consequence. such as vibrations. this work was in essence application-specific with the need to a priori know of the system failure modes to compare the current system behavior against those failure modes. and Lee (2002). 1996) follows the other CBA paradigm where the focus is placed on quantifying the performance degradation and no a priori failure knowledge is needed. In the case of encountering an unforeseen failure mode. White 1998). Then the newly arrived signatures can be compared against those learned during the normal system operation. or DC motor current signals are fed into it. utilization of the information from multiple sensors mounted on the system whose performance is being assessed highly desirable according to Hall (1992) is and Hall and Llinas (2000). Ni. the sensor fusion was heavily dependent on the prior characterization of the failure modes of the process/machine that was being “assessed. Multiple sensor readings are processed into a domain where generic signal features can be extracted to describe the current machine or process performance.The problem in accomplishing this task is that the machine degradationcycle is usually composed Sensors cv Flowchart Figure of Performance I Assessment Algorithm 65 . but the work presented in that paper was still focused on the description and detection of process failure. The merger of multiple sensor readings was accomplishfd through either feature level. the work by Lee (1995. the sensor fusion and performance assessment methods presented in that work could easily give spurious and wrong results. the sensorfusion in this work was accomplished through the direct hashing of sensors in a cerebellarmodel articulation controller (CMAC) neural network (Albus 1975).and multiple sensor readings are merged using an application-independent and algorithmic-rather than application-specific and rule-based-sensor fusion strategy. Higher overlaps between the two setsof signaturessignify a better system performance. However. Furthermore. 22/No. 1 2003 ure toward detection and description of performance degradation expressed in the drift of the newly arrived system signatures away from those observed during the normal. and the performance assessmentcan be based on the overlap between the normal and current system signatures. the process/productperformance was assessed a rather ad hoc manner and clearly a in novel method of assessingthe performance degradation was needed. Regardlessof the adopted approach to CBA. signatures representing the normal system behavior must be selected to learn the description of the normal system operation. or model-based sensorfusion (Steinberg. However. First. On the other hand. The goal of the work presented in this paper was to achieve generic process/machine performance assessmentbasedonly on the information about the normal system operation. In Djurdjanovic.neither faulty data nor expert knowledge is needed to set the performance assessmentprocess.Journul of Munufucturing Systems Vol. The diagram in Figure I describesthe basic steps necessary to assess the system performance by comparing the currently observed system operation with the normal one. Bowman.and was therefore also strongly application-dependent. healthy machine/process operation. decision level. Byington (2000) demonstrated several CBM applications based on multisensor fusion. which could be unfeasible when highly dimensional sensorreadings. Because this entire approach to performance assessmentrelies on appropriately describing the normal system operation. in each case.

It can be causedby. the machine temperature that increases up to a steady state or a mechanical wear-in of certain parts and components. and Tiemey (1983) to the set of extracted signal features to determine the end of the transient period.The learning process followed by the operation period during which tt PSD features are extracted from the newly arrive signals and the overlap between the newly obtaine signal features with those observed during the no ma1 system behavior is used to assessthe currel system performance.which correspon to the signal processingblock in Figure I. Singh. Frequency domain description of the signal. transient period normal operating period 8 degrading period fault. It is proposed to apply the Schruben’s test introduced in Schruben (1982) and Schruben.Jourmd of Munufucturin. 22/No.q Vol.. Theseperiods are: l l arefirst analyzedto determine the signals that shoul be used to represent the normal system behavio The performance signatures observed during no] ma1 operation are stored into a Feature Map (FPI/ during the learning process. is the sensorreading at time 9 & are the parametersof the AR model. tive measureof system performance will be refem to as the performance Confidence Value (CV). after which the system signatures can be used for describing the normal system operation.Therefore. Normal period period operating state 5. denotesresidual noise at time t..(l-Tye-’wT) . for example. as shown in Figure 2. which is based minimizing the forward prediction error (Pandit a wu 1993). Learning 3. The system performance assessmentmethod proposed in this paper can be summarized as follows. t. 1 2003 Systems r Confidence I. machine breakdown.OlO< 66 . where: l l X. These features is fit to the sensor signals to describe the dynamj of the signal. Fault occurs L Figure 2 Evolution of Performance Confidence Value (CV) Over Time of severalperiods. the featu map basedprocessperformanceassessment.are calculated usi the modified covariance method. can now be obtained as follows: 4 J= (l-r. Inclusion of the process signatures observed during the wear-in period into characterizing the normal system behavior could result in false characterization of the normal system operation and subsequently adversely affect the performance assessment. or repair time l Spectral Content Estimation The spectral content of sensor readings was es mated using the autoregressive(AR) modeling tee nique to accuratelydescribeenergy peaks and disce nearby modes in the frequency domain. Degrading state 6. Following Lee. this quantit. First. autoregressivemodel of the form The transient period that precedesthe normal system behavior representsthe warm-up or wear-in period of the system’s operation. E. The model parameters. Autoregressive (AR) models are fit to the time series signals of mt&iple sensors and the roots of the corresponding dR model characteristic equation are used to extract the Power Spectral Density (PSD) peaks of individual sensor readings.e-jwT)(l-~~ie-jOT). <bi. this wear-in period should be identified and avoided when the normal system behavior is being learned. Transient 2. The following subsections describe:the AR mode ing basedspectral contentestimation.

= I Y. If the speedof the PSD estimation algorithm is critical.I > 0. On the other hand.?is the residual noise variance. and rk = pkeiek arethe roots of the autoregressive characteristicpolynomial I+ Ii1 lpkXk=o P=l The power spectral density (PSD) of the signal X. However. if the root gets closer to the unit circle. but application-specific adjustments can be made. Though not applied in this paper. The order of the autoregressivemodel is set asthat of the adequate 67 . for each AR root r. k = 1.e’@. (For more information on the methods of spectral estimation. . the performance degradation can be assessed using the evolution of the signal featuresextracted from the AR model of order p of the newly arrived signals. the spectral content of all signals was estimated only at roots r. The AR-based spectral estimation method is gearedtoward accurately describing the PSD peaks by determining the roots of the AR model of the signal and evaluating the signal PSD only at the frequencies corresponding to those AR roots. The logarithm of the PSD intensity was utilized to better localize the PSD peaks in the FM.. pp. The order p of the adequate AR model was found using the AIC criterion &scribed in Pandit and Wu (1993). given as The closer the AR root. the more prominent the PSI3 peak is at the frequency o1 corresponding to that root. 159-163.where T is the sampling interval. the correspondingPSD peakis more prominent. the normal processcharacterization may not be good becauseof the high feature variability. From then on. is the magnitude of the frequency-domain description of the signal X. is to the unit circle in the complex plane. to extract only the most prominent PSD peaks. andthe other representing the logarithm of the PSD intensity at that frequency). features corresponding to a degraded state may be learned.9. = p. local PSD the peak&r be obtained. p model for the signals describing the norma] machine or process behavior. cr. During the learning period..The initial wear-in or warm-up period can be identified and disregarded by applying the Schruben’stest to the extracted PSD signal features. one may refer to Marple 1987 or Kay 1988 and references therein.. if the learning period is too long. a two-dimensional feature map (FM) is created. Hence. 2. methods exist to improve the FM resolution by moving the cells toward the features observed during the period of learning the normal process or machine behavior (Eldracher. then the most prominent PSD peakscan also be estimated using the well-known fast Fourier transform (FFT). after which the normal machine or process behavior can be learned. the threshold for the module of the AR roots at which the PSD was calculated was set to 0. Thus.eachFM is split into regularly spaced two-dimensional cells (one dimension representingthe frequency of the PSD peak.) In this method. Pomp1 1994).9. If this number is too small. Feature Maps and Performance Assessment For each sensor. . equidistant frequency samples. such that p. the duration of learning period should be carefully chosen and is Thus. r. . Becauseit is assumed that there is no a priori information on the localization of the features. the FM learnsto characterize the normal process or machine behavior. Staller. The first dimension of the feature represents the frequency domain. As explained earlier. In this paper. and the second dimension represents the logarithm of the corresponding power spectral density (PSD). the frequency content of the signal was calculated only at frequencies v such that wT= 0. (1). and to extract local PSD peaks. where the spectral content of the signal is estimated at predetermined. the number of process signaturesneeded for the learning period still needs to be determined. two parametersneedto be set: the order of the autoregressivemodel and the threshold of the module of the AR roots for which the spectral content will be estimated using Eq..

the smaller the variance of a given sensor CV.Journul oj Munujucrurin(: Systems Vol. with fixed weights. and o&. Throughout this process. as follows: l freq/ and PSD’ l &a. y) is the sum of squaresof the activation levels of each sample of that PSD peak. the FM cells are updated to model the distribution of the PSD peaks corresponding to the normal process of machine behavior. MVi’.sensor readings were collected andprocessperformancewas assessed according to the methods described in the previous section. and any shifts in the featuresextracted from that sensorreading become more emphasized.the updatereferred to as activation level (AL) is determined for each FM cell (x. shown in Figure 3 (Continental Acia 781 model no. The weights associated with each individual sensor CV are chosen as the inverse of the variance of that individual sensor CV observed during the learning period. Bowman. White 1998). The PSD peak features describing the newly arrived signals are extracted and are fed to the FM to calculate the match value (MV). l CV of an individual sensor reading corresponds to equal weighting assignedto eachPSD peak. y)AL/ (x. and spot-resistance welding was performed until those electrodes wore out. For each PSD peak i of thejth process signature. respectively. 601 ES-5).. Dynamic focusing through the weights associated with PSD peaks will be addressed future work. y) = c ALP/(x.Another possibility for calculating the performance CV is to utilize dynamic weighting where the weights associated with each PSD peak would be increased for those peaks that migrate away from the normal behavior and thus depict the performancedegradation. The final activation level for the ith PSD peak at the FM cell centeredat (x. The data are collected tbru a data acquisition card (NI606OE). Thus. estimated from the set of signals chosen to represent the normal process system behavior. During the learning period. as follows: MV. the final CV is computed as the weighted averageof all individual sensorCVs. in The performance CVs calculated from individual sensorreadings can now be merged into an overall multisensor CV. and sig- Once the learning period is over.y) as follows: where are the PSD peak frequency and logarithm of its intensity. In this paper. 22/No. it was observed that the FM is able to learn the nominal process behavior after 20 to 30 signals representing the normal process behavior were presentedto it.‘(x. y) The performance confidence value (CV) based on individual sensbr readings is now computed as the averageMV of all the PSD peaks extracted for that particular sensor. and each of the sensor readings was digitally sampled at the sampling rate of 5 kHz. the higher the weight associated with it. A new pair of 06mm brass spot welding conical electrodes was set on the machine. correspondingto the ith PSD peak of the jth process signature. x and y denote the center of the two-dimensional FM cell. I 2003 process dependent. the activation levels are fixed. Experimental Results Experimental Setup The experiment has been conductedon the welding machine. Future research is needed on dynamic focusing on the sensors with migrating PSD peaks as the migration occurs. This sensor fusion corresponds to the decision-level sensorfusion accordingto the Joint Directors of Laboratories (JDL) data fusion model (Steinberg. respectively. are the sampled variances of the ith PSD peak frequency and intensity.This method of calculating the 68 . The multisensor CV fusion performed this way is also static. A total of 3550 welds were made. In the experiments presented in this paper. Three sensorsused to assess the welding process measured the electrode current and voltage as well as the force that occurs between electrodesduring the welding process.

4x 0. The welding process degradation mirrored in the migration of the extracted signal featuresincurs ever smaller and smaller overlaps between the newly arrived process signaturesand the learned signatures. The next 30 recorded welds were usedto learn the activation levels of the featuremap.I. andthese the welds were therefore skipped in the training period. CVs shown in Figures 5 and 6 indicate a clear degradationin the welding processperformance after some 1500-2000 welds. Welding Machine Figure 3 Used for Experiment Discussion natures from every 10th weld were recorded. as shown in Figures 5 and 6.475 0. and Figure 6 gives the result of sensorfusion where the performance CV was calculated as the weighted averageof the individual sensorCVs.485 il. As this overlap decreases. 69 . which renderstraditional statisticalpatternrecognitiontechniques inappropriatefor this application. * The total of 200 welds was made when those 20 welds were recorded.the CVs show a decreasing trend. the features extracted from the current. Note that the distribution of these features is cumbersome to parameterize. aswell as how it looked after it was used to perform 3550 welds.SV Nomalired PSD intensity Histogram of PSD Features Roots Characterizing Figure 4 Extracted from the Current One of the AR(50) Signals Figure 5 shows the CVs computed for each sensor reading individually.. This migration is illustrated in Figures 7u and 7b for feature 16 of the current sensor. Figure 8 shows the electrode used in this experiment before any welds were donewith it.47 11. thus yielding the set of 355 recorded welds. During the 3550 welding operations performed in this experiment. voltage.465 Il. with weights equal to the inverse of the individual sensors’CV variance observedduring the learning period. which in this case can be attributed to the wear processof the welding electrode. Figure 4 shows the histogram of the power spectraldensity (PSD) characteristics of one of the AR(50) roots characterizingthe current sensorsignal. One should note that this degradation process precedes the actual process failure when the Results Autoregressive models of order ‘50 were fit to all sensorreadingsand features wereextractedasdescribed previously.5 ll.491 0.4Y 0. This migration is causedby the degrading characteristics of the welding process. and force sensorreadings migrate. Schruben’stest was individually applied to each extracted feature and yielded that the first 20 recordedwelds* represented wear-in period.

Weld numkr Weld number Confidence Values (0’s) Calculated Using Figure 5 of Welding Process Performance Individual Sensor Readings Figure Overall Confidence of Welding Process 6 Values (CVs) Performance welding performance becomes unacceptabledue to excessive electrode wear.5 (a) 1.0.S. Figure 9 showsthe evolution of one of the PSD features extracted from the current sensorreadings.500 2lUl1) 2sw XKKI 3SlKI 4001.K) 2lKKl 2.111 . 2.100 0 SO0 lO(Kl . It is also apparentthat the CVs associatedwith the earliest few welds are lower.501) XKKI 75110 4.000 . (b) Evolution of PSD for Feature 16 70 . thus always ensuring the high quality of the welds. Inclusion of the wear-in period could hamper the performance assessment.and this is why the Schruben’s test is used to eliminate these signatures from the learning period.117 The wear-in period is obvious as the features grate toward some steady state (normal behav within the first 15-20 recorded welds (correspo to the first 150-200 performed welds). which is a reflection of the electrodewear-in period. Conclusions and Future Work Quantitative methods for multisensor-based 1 cess assessment are presented in this paper.x 0 0 5. 1 7. The degradation process can be used to change the “almost-worn” pair of electrodes before the processperformance becomes unacceptable. It is this ste state apparent after the first 15-20 recorded w( that characterizes the normal welding process 1 formance against which the newly arrived proc signatures need to be compared to assess the 1 cess performance. (b) (a) Evolution of Frequency for Figure 7 Feature 16.

High overlap of the currently observed and normal process signatures results in a high performance confidence value (CV).” TTLIIZS.S.and this problem will be thoroughly investigatedin the future work (knowing “when something is going to happen”). (1975). Furthermore. while drift of the-currently observed system signatures away from those observed during the normal system behavior causesa drop in the performance CV. In addition. References Albus. will be used to facilitate both the diagnostic and prognostic features through this “peer-to-peer” paradigm. pp220227.Journal of’ Manufacturing Vol. 1 2003 06mm Welding and After Figure 8 Electrode Before (right) 3550 Welds (left) quency content of multiple sensory inputs is extracted through the use of autoregressivemodeling.” but also “what is happening”). J. which would enrich the diagnostic information in the process assessment (knowing not only that “something is happening. Measurement and Control. methods for autonomous selection of the order of the AR model and of the FM resolution arebeing investigated to fully enable the “plug-andplay” implementation of the newly introduced methods. Researchwork is currently under way to facilitate generic processperformance assessmentin the presence of nonstationary sensory signals with timevarying spectral content (Djurdjanovic. Future work in the development of the AR/ from I Power Current Spectral Sensor Figure Density Readings 9 Features During Extracted First 500 Welds FM-based approachto processperformance assessment also includes tests of the FM in recognizing faulty states once those states are presented to it. Systems 22/No. Journal oj Dpnnmic Systems. Acknowledgment This researchwas supported in part by the NSF I/UCRC Centerfor Intelligent Maintenance Systems at the University of Michigan and University of Wisconsin-Milwaukee. “A new approach to manipulator control: the cerebellar model articulation controller (CMAC). as well as the performance of similar systems in operation (“brother-systems”). 71 . special attention was dedicated to reducing and eliminating the need for expert human involvement in setting up the parametersof the featureextraction methods and feature map based pattern matching. to facilitate the “plug-and-play” implementation of the proposed process assessmentmethods. Lee 2002). Historical data on the process performance. Furthermore. the methods described in this paper can be applied in a wide rangeof applications. Finally. performing additional experimental tests with the methods presentedin this paper in applications other than welding needs to be done. A major advantage of the newly proposed methods for performance assessmentis their generality achieved through extraction of generic signal features and generic methods of signature matching. Ni. of’ AWE. Also. Thus. This quantitative information about process performance can be used to take appropriate maintenance action in a timely manner. forecasting of the processbehavior and performance is essential for predictive condition basedmaintenance. 2002). a link must be established with a decision-making module that will enable optimal maintenance action in a timely manner (Chen et al. and matching of the currently observed process signatures against those observed during the normal process behavior is accomplished through the use of feature maps.

“Optimal tests for initialization bias in simulation output. pp29712977. maintenance. Ray. He received his PhD of Natal.” Int ‘I Jownal qj’Compzrter Integrated A4fb’. J. deputy director of the NSF Engineering Ccntcr for Rcconfigurable Manufacturing Systems.. Swanson. pp457-469. and Wu. manufacturing process modeling and control. Authors’ Biographies Nicolas Casoetto is a PhD candidate and graduate the Dept.N.” Proc. n6).C. of IEEE Acrospacc Conf. (2000). Lee his PhD in mechanical engineering from George Washington U His research interests include innovative design methods for scl nancc products. Dr. S. and Tiemcy. Mathemuticul Techniques irr Multi-Sensor Data Fusion.” Proc. Dr Ni’s research interests manufacturing science and engineering.” Computers in Industry (~30).” Proc. Quebec City. Wu Manufact of Michigan. (1993).. and Wang. ar systems. statist ity design and improvement. n5). Malabar. CRC Press.J. L. D. H..M. Hall. ~~443-449. J. Jr. Rhett Mayor is a rcscarch search Center at the University BS degrees from the University fellow in the S. FL: Kricger Publishing Co. 72 . and business ir strategies. Kay. (1998).: and Hess. NY Operation Research (~3 1. C. Singh. L. (~8. (1995).W.s. Wu Manufacturing Research Ce University of Michigan. “Fault prognosis using dynamic wavclct neural networks. Hall. (2001). with special focuses on machining.M. Handbook ofSensor F~uion.. F. trial and Manufacturing Engineering at the University of Wisco waukce. Digital Spectral Ana[vsi. and web-enabled intelligent monitoring. (1987). 22/No. NJ: Prentice Hall. (~6). Lee. A. micro/mesa systems and manufactt ccsses. Canada. of 2001 IEEE Int’l Symp. pp79-84. “Prognostics.D. K. He received his MS and PhD degrees fron versity of Wisconsin-Madison. J. Englcwood Cliffs. Internet-based machine degradation assessment maintcnancc. (1996). Modern Spectral Estimutiorz. c-factory and c-manufacturing.. of IEEE Aerospace Conf. “A gcncral prognostics tracking algorithm for prcdictivc maintenance. L.Journal qf’ Manufircturing Vol.L. (2001).M. “Measurement of machine pcrformancc degradation using a neural network model. A. Steinberg. (2000). A. (1983). Schruben.. “Revisions to the JDL data fusion model.: Bongort. S. ~~193-209. Jay Lee is the Wisconsin Distinguished Professor in the Dept. and Tangirala. Marple.. of Mechanical at the University of Michigan. L. S. 1 2003 Systems Engel.E. B. “Stochastic modcling of fatigue crack dynamic for on-line failure prognostics. Gilmartin. He is also director of the NSF Industry/University Co Research Center for Intelligent e-Maintenance Systems. S. and Llinas. Lee.W. (v6).. G. Englcwood Cliffs.L. “Detecting initialization bias in simulation output. Artcch House Inc. S-M. n4). P. ~~370-380. on Control $wterns Technology (~4. the real issues involved with predicting lift remaining. Vachtsevanos.J. of Mechanical Engineering at the University rcscarch a of Mich of En Dragan Djurdjanovic is an adjunct assistant professor operations and a research fellow in the Dept. (1982). of 3rd NATO/IRIS Conf. S. (1988).D.” Proc. (1996). and White.M. Durban. NJ: Prcnticc Hall.” l/Z?/% Trans. “Machine performance monitoring and proactive maintcnancc in computer-integrated manufacturing: review and pcrspcctivc. cds. (1992). on Intclligent Control.” Conzell Universir?: Ithaca. Schruben. and co-direr NSF IndustryUnivcrsity Coopcrativc Rcscarch Center for Inte Maintcnancc Systems. Time Series andsystem Analysis with Application. Bowman. L.” Operations Research (~30. Jun Ni is director ofthc S. Pandit. n3).