Neuro classifiers for condition and bearing health assessment of an electric motor

Abstract

Electric motors (EMs) and their maintenance applications have recently constituted the central core of most industrial processes. Induction motors (IMs) are vastly used among various types of EMs due to their robustness, simplicity, and reliability. Indeed, reliability plays an essential role in all engineering applications. In order to enhance the integrity or safety and prevent any potential devastation or failure in the power systems, the Condition Monitoring (CM) of IMs has been encountered by lots of enthusiasts recently. Due to this, many researchers have focused on investigating and detecting the different types of failures in IMs. According to the information acquired from several surveys, the approximate rates of bearing failures, winding failures, and rotor-related failures are 40%, 30%, and 9%. It can be concluded that the bearing failures are the most probable faults in the IMs. Therefore, a CM technique is vital for detecting the bearing degradations at their earliest stages in an IM. In this way, accurate and quantitative information on the present condition of the IM can be found by tracking the performance of the motor. There are plenty of CM and Fault detection (FD) techniques to diagnose the various faults and their associated frequencies in IMs. However, monitoring the aging levels indicating the overall condition of an IM is not considered vastly. Indeed, aging is a gradual and progressive process arising from several faults in the EMs. The bearing faults can be enumerated as the main problem accelerating the aging in an IM. Therefore, besides the overall condition of an IM, the bearing condition of the IM is also classified in this study. Furthermore, signals acquired from EMs contain lots of information on the condition of the system. The most favorable signal is the vibration since both the electrical and mechanical problems can be revealed through its proper interpretation. In this study, the vibration signals are employed due to their direct representations of the bearing degradations in an IM. Indeed, eight vibration signals have been collected through an accelerated aging experiment on an IM. The experimental data is obtained from an accelerated aging experiment, which was conducted in a research and development project supported by The University of Tennessee Maintenance and Reliability Centre. The Electrical Discharge Machining (EDM) and the thermal-chemical aging operations constitute the two phases of the accelerated aging experiment. In recent years, Artificial Neural Networks (ANNs) have also been used vastly as simple and feasible computational tools for the CM and FD of IMs. There are many research works regarding the classification applications of ANNs in power systems. The preferable architecture for the classification applications of ANNs is the FFNN structure. The significant advantage of ANNs is the fact that they are able to classify various types of faults and their severities in an IM. In this study, two different types of Feedforward Neural Network (FFNN) models are designed with different training optimization algorithms for classifying eight aging levels of an IM. The first model is a shallow Pattern Recognition Neural Network (PRNN) trained by two different optimization algorithms. The Levenberg-Marquardt and the Scaled Conjugate Gradient algorithms are employed as the two optimization algorithms in the developed PRNNs. The second classifier is a deep, fully connected FFNN solved by the limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm. All the classification applications developed in this study benefit from the error Backpropagation (BP) technique. The performances of these classifiers are primarily assessed with the aim of monitoring the overall condition of the IM. The second aim is to evaluate their performances in monitoring and classifying the different bearing conditions of the IM. This study constitutes two parts containing four different ANN-based applications. In the first part, the time-domain statistical features of the provided experimental vibration signals are directly processed by different neuro classifiers to rate the overall conditions of the IM. Indeed, these two classifiers are operated without any preprocessing operations. This part evaluates the performance of the PRNN model trained by the Levenberg-Marquardt and the Scaled Conjugate Gradient algorithms and a multi-layered FFNN solved by the L-BFGS algorithm. The second part represents the bearing aging classification applications by using the aforementioned shallow and deep classifiers. A comparison-based study is executed on the testing performances of the PRNN trained by the Levenberg-Marquardt algorithm and the L-BFGS-based FFNN. In this part, primarily, the high-frequency bands associated with the bearing faults in the spectra of the vibration signals are decomposed by using a one-level Multiresolution Wavelet Analysis (MRWA) technique. In this regard, only the essential frequency bands containing the bearing fault features are employed through a preprocessing process. Additionally, novel spectral-based features are selected as the inputs of the proposed classifiers to reduce the input dimensions of the network. The performances of the networks, with the optimal hidden layer geometry in each application, are also analyzed by computing four metrics. The four metrics available for evaluating the developed classifiers are the accuracy rate, the Recall, the Precision, and the F1-score. In addition, the confusion matrix, Receiver Operating Characteristic (ROC), error histogram, and training loss plots are also used to assess the reliability and performance of the designed neuro classifiers. By comparing the performances of the proposed classifiers, it is deduced that the L-BFGS-based FFNN applications outperform the PRNN applications, generally. In the first part of the study, The testing accuracy rates of the network with the optimal geometry in the PRNN and the L-BFGS-based FFNN applications are 97.2% and 100%, respectively. Likewise, in the MRWA-based study, the testing accuracy rates of the network with the optimal geometry for the shallow and deep classifiers are 97.9% and 100%, respectively. Therefore, it is concluded that the outstanding performance of the MRWA-based deep classifier indicates the novelty and high reliability of the L-BFGS-based FFNN application in grading the different bearing aging levels of an IM. Indeed, among these four applications, MRWA-based FFNN solved by the L-BFGS algorithm has the most important contribution in terms of novelty, high reliability and high performance.