{"title": "A Neural Network Autoassociator for Induction Motor Failure Prediction", "book": "Advances in Neural Information Processing Systems", "page_first": 924, "page_last": 930, "abstract": null, "full_text": "A Neural Network Autoassociator for \nInduction Motor Failure Prediction \n\nThomas Petsche, Angelo Marcantonio, Christian Darken, \n\nStephen J. Hanson, Gary M. Kuhn and Iwan Santoso \n\n[PETSCHE, ANGELO, DARKEN, JOSE, GMK, NIS]@SCR.SIEMENS.COM \n\nSiemens Corporate Research, Inc. \n\n755 College Road East \nPrinceton, NJ 08853 \n\nAbstract \n\nWe present results on the use of neural network based autoassociators \nwhich act as novelty or anomaly detectors to detect imminent motor \nfailures. The autoassociator is trained to reconstruct spectra obtained \nfrom the healthy motor. In laboratory tests, we have demonstrated that the \ntrained autoassociator has a small reconstruction error on measurements \nrecorded from healthy motors but a larger error on those recorded from a \nmotor with a fault. We have designed and built a motor monitoring system \nusing an autoassociator for anomaly detection and are in the process of \ntesting the system at three industrial and commercial sites. \n\n1 \n\nIntroduction \n\nAn unexpected breakdown of an electric induction motor can cause financial loss signif(cid:173)\nicantly in excess of the cost of the motor. For example, the breakdown of a motor in a \nproduction line during a production run can cause the loss of work in progress as well as \nloss of production time. \n\nWhen a motor does fail, it is not uncommon to replace it with an oversized motor based on \nthe assumption that if a motor is not running at its design limit then it will survive longer. \nWhile this is frequently effective, this leads to significantly lower operating efficiencies and \nhigher initial and operating costs. \n\nThe primary motivation behind this project is the observation that if a motor breakdown and \nbe predicted before the actual breakdown occurs, then the motor can be replaced in a more \norderly way, with minimal interruption of the process in which it is involved. The goal is \nto produce a system that is conceptually similar to a fuel gauge on an automobile. When \nthe system detects conditions that indicate that the motor is approaching its end-of-life, the \noperators are notified that a replacement is necessary in the near future. \n\n\fA Neural Network Autoassociator for Induction Motor Failure Prediction \n\n925 \n\n2 Background \n\nAt present, motors in critical operations that are subject to mechanical failures - for example, \nfire pump motors on US Navy vessels - are typically monitored by a human expert who \nperiodically listens to the vibrations of the motor and, based on experience, determines \nwhether the motor sounds healthy or sounds like a problem is developing. Since mechanical \nprobiems in motors typically lead to increased or changed vibrations, this technique can \nwerk well. Unfortunately, it depends on a competent and expensive expert. \n\nIn an attempt to automate motor monitoring, several vendors have \"automated motor moni(cid:173)\ntoring\" equipment available. For mechanical failure monitoring, such systems typically rely \non several accelerometers to measure the vibration of the motor at various points and along \nvarious axes. The systems then display information, primarily about the vibration spectrum, \nto an operator who determines whether the motor is functioning properly. These systems \nare expensive since they rely on several accelerometers, each of which is itself expensive, \nas well as data collection hardware and a computer. Further, the systems require an expert \noperator and frequently require that the motor be tested only when it is driving a known load. \n\nNeither the human motor expert nor the existing motor monitoring systems provide an \naffordable solution for continuous on-line mechanical failure monitoring. However, the \nsuccess of the human expert and existing vibration monitors does demonstrate that in fact, \nthere is sufficient information in the vibration of an electric induction motor to detect \nimminent mechanical failures. \n\nSiemens Energy and Automation has proposed a new product, the Siemens Advanced Motor \nMaster System II (SAMMS II), that will continuously monitor and protect an electric induc(cid:173)\ntion motor while it is operating on-line. Like the presently available SAMMS, the SAMMS \nII is designed to provide protection against thermal and electrical overload an, in addition, \nit will provide detection of insulation deterioration and mechanical fault monitoring. \n\nIn contrast to existing systems and techniques, the SAMMS II is designed to (1) require \nno human expert to determine if a motor is developing problems; (2) be inexpensive; and \n(3) provide continuous, on-line monitoring of the motor in normal operation. \n\nThe requirements for the SAMMS II, in partiCUlar the cost constraint, require that several \nissues be resolved. First, in order to produce a low cost system, it is necessary to eliminate \nthe need for expensive accelerometers. Second, wiring should be limited to the motor control \ncenter, i.e., it should not be necessary to run new signal wires from the motor control center \nto the motor. Third, the SAMMS II is to provide continuous on-line monitoring, so the \nsystem must adapt to or factor out the effect of changing loads on the motor. Finally since \nthe SAMMS II would not necessarily be bundled with a motor and so might be used to \ncontrol and monitor an arbitrary motor from an arbitrary manufacturer, the design can not \nassume that a full description of the motor construction is available. \n\n3 Approach \n\nThe first task was to determine how to eliminate the accelerometers. Based on work done \nelsewhere (Schoen, Habetler & Bartheld, 1994), SE&A determined that it might be possible \nto use measurements of the current on a single phase of the power supply to estimate the \nvibration of the motor. This depends on the assumption that any vibration of the motor will \ncause the rotor to move radially relative to the stator which will cause changes in the airgap \nwhich, in tum, will induce changes in the current. \n\nExperiments were done at the Georgia Institute of Technology to determine the feasibility \nof this idea using the same sort of data collection system described later. Early experiments \nindicated that, for a single motor driving a variety of loads, it is possible to distinguish \n\n\f926 \n\nT. PETSCHE, A. MARCANTONIO, C. DARKEN, S. J. HANSON, G. M. KUHN, I. SANTOSO \n\nTable 1: Loads for motors #1 and #2. \n\nLoad type \nconstant \nsinusoidal oscillation at rotating frequency \nsinusoidal oscillation at twice the rotating frequency \nswitching load (50% duty cycle) at rotating frequency \nsinusoidal oscillation 28 Hz \nsinusoidal oscillation at 30 Hz \nswitching load (50% duty cycle) at 30 Hz \n\nLoad Magnitude \nhalf and full rated \nhalf and full rated \nfull rated \nfull rated \nhalf and full rated \nfull rated \nfull rated \n\nTable 2: Neural network classifier experiment. \n\nFeatures (N) \nPerformance on motor #1 \nPerformance on motor #2 \n\n48 \n100% \n\n63 \n100% \n30% \n\n64 \n92% \n25% \n\n110 \n100% \n55% \n\n320 \n100% \n37% \n\nbetween a current spectrum obtained from the motor while it is healthy and another obtained \nwhen the motor contains a fault. Moreover, it is also possible to automatically generate a \nclassifiers that correctly determine the presence or absence of a fault in the motor. \n\nThe first, obvious approach to this monitoring task would seem to be to build a classifier \nthat would be used to distinguish between a healthy motor and one that has developed a \nfault that is likely to lead to a breakdown. Unfortunately, this approach does not work. \n\nAs described above, we have successfully built classifiers of various sorts using manual and \nautomatic techniques to distinguish between current spectra obtained from a motor when it \nis healthy and those obtained when it contains a fault. \n\nHowever, since the SAMMS II will be connected to a motor before it fails and will be asked \nto identify a failure without ever seeing a labeled example of a failure from that motor, a \nclassifier can only be used if it can be trained on data collected from one or more motors \nand then used to monitor the motor of interest. Unfortunately, experiments indicate that \nthis will not work. \n\nOne of these experiments is illustrated in table 2. Several feedforward neural network clas(cid:173)\nsifiers were trained using examples from a single motor under four conditions: (1) healthy, \n(2) unbalanced, (3) containing a broken rotor bar and (4) containing a hole in the outer \nbearing race. The ten different loads listed in table 1 were applied to the motor for each of \nthese conditions. \nThe networks contained N inputs (where N is given in table 2); 9 hidden units and 4 \noutputs. There were 40 training examples where each example is the average of 50 distinct \nmagnitude scaled FFrs obtained from motor #1 from a single load/fault combination. The \ntest data for which the results are reported in the table consisted of 40 averaged FFfs from \nmotor #1 and 20 averaged FFfs (balanced and unbalanced only) from motor #2. The test \nset for motor #1 is completely distinct from the training set. \nIn the case where n = 110, the FFf components were selected to include the frequencies \nidentified by the theory of motor physics as interesting for the three fault conditions and \nexclude all other components. This led to an improvement over the other cases where a \nsingle contiguous set of components was chosen, but the performance still degrades to about \nrandom chance instead of 100%. \n\nThis experiment clearly illustrates that is is possible to distinguish between healthy and \nfaulty spectra obtained from the same motor. However, it also clearly illustrates that a \n\n\fA Neural Network Autoassociator for Induction Motor Failure Prediction \n\n927 \n\nMeasurements \n\nNovelty \ndetection \n\nNovelty \n\nDecision \n\nDiagnosis \n\nAdaptation \nAlgOrithm \n\nFigure 1: The basic form of an anomaly detection system. \n\nclassifier trained on one motor does not perform well on another motor since the error rates \nincrease immensely. Based on results such as these, we have concluded that it is not feasible \nto build a single classifier that would be trained once and then placed in the field to monitor \na motor. Instead we are pursuing an alternative based on anomaly detection which adapts \na monitor to the particular motor for which it is responsible. \n\n4 Anomaly detection \n\nThe basic notion of anomaly detection for monitoring is illustrated in figure 1. Statistical \nanomaly detection centers around a model of the data that was seen while the motor was \noperating normally. This model is produced by collecting spectra from the motor while \nit is operating normally. Once trained, the system compares each new spectrum to the \nmodel to determine how similar to or different from the training set it is. This similarity \nis described by an \"anomaly metric\" which, in the simplest case, can be thresholded to \ndetermine whether the motor is still normal or has developed a fault. Once the \"anomaly \nmetric\" has been generated, various statistical techniques can be used to determine if there \nhas been a change in the distribution of values. \n\n5 A Neural Network-based Anomaly Detector \n\nThe core of the most successful monitoring system we have built to date is a neural network \ndesigned to function as an autoassociator (Rumelhart, Hinton & Williams, 1986, called it \nan \"encoder\"). We use a simple three layer feedforward network with N inputs, N outputs \nand K < N hidden units. The input layer is fully connected to the hidden layer which is \nfully connected to the output layer. Each unit in the hidden and output layers computes \nXi = (J ( 2::;0 Wi,jXj) , where Xi is the output of neuron i which receives inputs from Mi other \nneurons and Wi,j is the weight on the connection from neuron} to neuron i. The network is \ntrained using the backpropagation algorithm to reconstruct the input vector on the output \nunits. Specifically, if Xi is one of n input vectors and Xi is the corresponding output vector, \nthe network is trained to minimize the sum of squared errors E = 2::~1 Ilxi - xdl2. Once \ntraining is complete, the anomaly metric is mi = IIXi -\n\nxi11 2. \n\n6 Anomaly Detection Test \n\nWe have tested the effectiveness of the neural network autoassociator as an anomaly detector \non several motors. For all these tests, the autoasociator had 20 hidden units. The hidden \nlayer size was chosen after some experimentation and data analysis on motor #1 , but no \nattempt was made to tune the' hidden layer size for motor #2 or motor #3. \n\nMotor #1 was tested using the ten different loads listed in table 1 and four different \n\n\f928 \n\nT. PETSCHE, A. MARCANTONIO, C. DARKEN, S. 1. HANSON, O. M. KUHN, I: SANTOSO \n\nq ,---------------------------~ \n\n....... \n\n,.;-.. , \n,_.-\n\n