CONDITION ASSESSMENT AND RESIDUAL LIFE PREDICTION OF ROLLING ELEMENT BEARINGS

Abstract

Rolling element bearings are widely used in almost all forms of rotating machinery, which in turn constitute a huge fraction of the mechanical equipments used in industries. The failure in bearings is one of the leading causes of catastrophic breakdown in rotating machinery. Such kinds of failures have damaging effects like lengthy industrial downtime, heavy maintenance costs, economic losses and human casualties etc. This undesirable state of affairs demands the development and deployment of effective maintenance systems for failure prevention in bearings. There exist two key issues while dealing with the condition monitoring of bearings. The first issue is that the incipient defects need to be detected as early as possible so that a timely warning could be attained and the sudden machinery breakdown might be avoided. However, only the detection of incipient faults is not sufficient to guarantee the uninterrupted operation of machinery. In real situations, the degradation in bearings propagates in a gradual manner before the event of total failure. Thus, rather than just focusing on the incipient degradation detection, the analysis of performance degradation development in bearings should also be taken into account to avoid the early replacement of bearings and deter a possible increase in unnecessary maintenance activities. Further, even after the occurrence of defects in them, the bearings have significant useful life left and need not to be immediately replaced. A premature replacement of bearings leads to untimely stoppage of the machinery giving invitation to the harmful consequences such as production losses and rise in maintenance costs. Thus, the determination of residual or remaining useful life (RUL) of the bearings becomes important to ensure the maximum utilization of the machinery without being subjected to catastrophic failures. Recently, the condition-based maintenance technique termed as prognostics has received significant recognition from the scientific community for investigating the health of bearings. The aim of prognostics is to provide an intelligent quantitative assessment of the bearing condition and evaluate the RUL. Prognostics is more promising in preventing the fatal breakdowns and reducing the extra maintenance expenses. The present study therefore emphasizes on the construction of efficacious prognostics models for assessing the severity of performance degradation and predicting the RUL of bearings. The prognostics models considered in the study are data-driven in nature and make use of the vibration conditionmonitoring data for their implementation. An experimental-setup is developed through which run-to-failure tests are conducted on 6205 type rolling element bearings under a constant ii operating load of 1 kN and a constant speed of 1500 RPM. The vibration data is acquired through piezoelectric accelerometers mounted on the bearings’ housings and a computer programmed with data-acquisition software. The run-to-failure data of five bearings with different lifetimes and failing with different types of defects i.e. inner race defect, outer race defect and ball defect are utilized in the current study. To address the concerns of incipient fault detection and trending the degradation in bearings, two condition-assessment (CA) methods based on machine-learning methods called as self-organizing map (SOM) and k-medoids are developed for classifying the degradation states in bearings. The SOM-CA method utilizes the time-domain and frequency domain methods for extraction of fault features from the vibration signals, which certainly remains as one of the key challenges in bearing condition monitoring. The extracted features are later supplied to the SOM algorithm to generate health indicator (HI) for predicting the severity of faults in bearings. The advocated technique is applied on the collected experimental signals and the results demonstrate that the HI deduced from feature space composed of both time-domain and frequency-domain features provided superior outcomes in terms of incipient fault detection and existence of outliers as compared to the HI computed using feature space of time-domain features only. Further, it is revealed that the suggested SOM-HI outperforms the conventional root mean square (RMS) and kurtosis features. In the second CA approach based on k-medoids, the time-frequency domain technique known as empirical mode decomposition (EMD) is utilized to fulfill the dispute of feature extraction from the bearing signals. The k-medoids model takes the time-frequency domain features called as singular value and energy entropy as inputs in order to produce HI for bearing health prediction. The experimental results reveal that the k-medoids approach has better performance than the existing approaches like RMS and kurtosis features, the SOM and the Fuzzy c-means (FCM). Afterwards, two prognostic models exploiting the supervised machine learning algorithms namely nonlinear autoregressive with exogenous inputs-neural networks (NARXNN) and support vector regression (SVR) are proposed for accomplishing the objective of bearing RUL prediction. In the first RUL prediction approach based on NARX-NN, a new monotonic HI is developed by utilizing the Mahalanobis distance (MD) and the cumulativesum (CUMSUM) techniques. The MD fuses the time-domain features extracted from the bearing signals filtered by continuous wavelet transform (CWT) and the CUMSUM is applied to extract a monotonically increasing trend from the MD. The available failure histories of the bearings are used to train the NARX-NN model and then the trained model is used to determine iii the RUL of new bearings. The MD-CUMSUM and the age of the bearing are set as the inputs and the life-percentage of the bearing is set to output while training the NARX-NN model. Similarly, the SVR-based RUL prediction model first constructs a monotonic HI called as Confidence value (CV) which is derived using the SOM model already developed for CA of bearings. The CVs and the ages of the bearings from the training group are utilized to train the SVR with the output set to life-percentages of the bearings. The SVR model after training is then ready to predict the RUL of test bearings. The new monotonic degradation indicators i.e. SOM-CV and MD-CUMSUM developed for training purposes help to reduce the noise and minimize the overfitting issues while learning the input data. The two approaches employ the bearing life-percentage for predicting the RUL and are therefore referred to as the lifepercentage prediction models in the thesis. The experimental results show that the suggested approaches predict the bearing RUL with sufficient degree of accuracy and the predicted RUL shifts nearer to the actual RUL during the final stages of bearing lifetime. Further, the proposed indicators i.e. CVs derived from SOM and the MD-CUMSUM performed better than the corresponding CVs and CUMSUMs of RMS and kurtosis features. In addition, the NARX-NN and the SVR models are found to provide more accurate predictions than the generally used feed-forward neural network (FFNN) models. Finally, a RUL prediction model based on multi-step ahead time-series prediction technique is proposed. In contrast to the life-percentage models, the time-series prediction model forecasts the HI time-series to a certain predefined failure threshold in order to compute the RUL. The proposed time-series prediction model first utilizes the Ensemble empirical mode decomposition (EEMD), Gaussian mixture models (GMM) and Jensen-Rényi divergence (JRD) to construct suitable monotonic HI usable for the purpose of RUL determination. Then, SVR in combination with multi-step ahead prediction technique is employed to project the developed bearing HI to future time-periods until an already established failure threshold is reached, thereby facilitating the approximation of bearing RUL. Meanwhile, the particle swarm optimization (PSO) technique is utilized to tune the parameters of the SVR, which promises greater accuracy and increased computational efficiency of the SVR model. The prediction results are again demonstrated to be convincing for degradation states in the late life of the rolling element bearings. The newly developed indicator i.e. the CVs derived from EEMD, GMM and JRD are realized to be superior than the GMM-negative-log likelihood probability frequently employed in the earlier published literature.