Development of Condition Monitoring Approaches for Rolling Element Bearing Using Singular Spectrum Analysis

Abstract

Condition-Based Maintenance (CBM) has become pivotal in maintenance activities. At the heart of CBM lies the necessity for effective, efficient, and accurate fault detection and prognosis in rolling element bearings, which are foundational components in many industrial machines and operations. Early fault detection in rolling element bearings benefits maintenance personnel in many ways. Identifying faults early can facilitate maintenance activities by providing a larger time window for planning replacement tasks. The vibration signals obtained from bearings are a reliable medium for CBM activities. Each system/machinery component has its unique vibration signature, closely related to the machinery's operating condition. Faults in machinery develop additional dynamic forces that generate vibration signatures within a specific frequency range. Various features from these vibration signals can be useful for estimating the bearing's state. These features can be determined via various signal processing techniques, broadly under time, frequency, and joint time-frequency domain techniques. Joint time-frequency domain techniques, an advancement in CBM, can identify defects while handling the nonlinearity and non-stationarity in signals. These popular techniques are based on the fundamentals of decomposing signals into numerous useful and informative components. Singular Spectrum Analysis (SSA) is one such less-explored technique in the field of bearing fault diagnosis. The Singular Spectrum Analysis (SSA) is a reliable non-parametric method used to separate signals from arbitrary noise, having a broad spectrum of applications ranging from biomedical signals to economics. The method consists of mainly two stages: Decomposition and Reconstruction. The Singular Value Decomposition (SVD) based decomposition process generates a number of singular value components depending upon the energy content. In the traditional approach, SSA emphasizes preserving high-energy singular components for reconstructing signal components from signal noise mixture. This approach has a limitation in identifying weak signal components. For bearing having incipient fault often generates weak fault signals. Hence, a conventional SSA may not be that effective in detecting the incipient faults in the bearing. Moreover, the accuracy of SSA depends upon the combination of window length L and the number of sub-signals considered for reconstruction r. An approach is proposed to estimate the appropriate decomposition and reconstruction parameters for bearing fault diagnosis. It employs a non-linear Teager Kaiser Energy (TKE) operator to enhance the impulsive feature of the raw vibration signature by converting it into a Teager Kaiser (TK) energy signal. For a TK energy signal with N data points, an effective combination of window length and reconstruction parameter has been identified to extract the oscillatory component corresponding to characteristic fault frequency for diagnosis. The method's accuracy is validated from approximately 1000 real-time vibration signals. The proposed method enhances the robustness of the bearing fault detection as it is in line with the classical fault detection technique aiming at detecting the characteristic fault frequencies. The TKEO-SSA-based method excels at classifying bearing defects, achieving 98.6% accuracy. However, the emergence of deep learning (DL), which eliminates the need for feature extraction by learning directly from raw data, has improved the classification accuracy of bearing states. Deep Neural Networks (DNNs), less costly computationally than CNNs, LSTMs, and Autoencoders, can process vibration data more effectively. Using a dataset from Case Western Reserve University containing vibration data from bearings with inner race, outer race, and ball faults, a DNN has been tested. It consisted of multiple hidden layers with a ReLU activation function and a Softmax function at the output layer. Despite these techniques, the DNN's classification accuracy has been found to be inferior to the TKEO-SSA method. To improve the DNN's performance, raw signals are pre-processed with SSA, and these filtered temporal signals have been used as input, along with their corresponding frequency features. This updated method achieved an average 10-fold cross-validation accuracy ranging from 99.76% to 100% for datasets based on various loading scenarios. Fault prognosis is a very important activity of CBM of rolling element bearings. A properly installed rolling element bearing inevitably experiences performance degradation over an extended operational period in rotating machinery. The timely and precise detection of faults in rolling element bearings integral to maintenance and repairs before an unexpected failure occurs remains a formidable challenge in the field of condition monitoring. The identification of the onset of a fault in vibration data, continuously obtained from the bearing at specific intervals, is facilitated with the help of a health indicator (HI). A dynamic health indicator based on the Kolmogorov-Smirnov Distance (KSD) is proposed for detecting anomalies in run-to failure bearing operations. The KSD, derived from a non-parametric statistical test, provides a single value representing the divergence between two samples. The efficacy of this indicator is bolstered by the use of SSA. This non-parametric time series decomposition technique enhances the signal component in a mixture of signal and noise. The effective health state is identified from the SSA-KSD distribution using a statistical change point detection technique based on the likelihood ratio of a signal's mean and variance. The proposed approach is rigorously evaluated using real-world and simulated signals. The results demonstrate the significant superiority of this method, offering promising prospects for its implementation in the field of condition monitoring for rolling element bearings. While the health indicator provides information about the trend of fault progression, knowing the dimension of the fault is more informative for estimating its severity than just knowing the trend of a statistical parameter. This knowledge can further be used to predict the remaining useful life of the bearing. The popular approach for estimating the size (in terms of width) of the spall-type faults involves tracing the entry and exit events of the rolling element interacting with the fault. The time estimated between the entry and exit of the rolling element from a pit like spall can be converted to a geometric estimation of the fault size using vibration signatures. The present approach demonstrates the use of SSA for this task. The vibration signals generated from the interaction of the rolling element with the localized fault are hybrid signals, consisting of a low-frequency stepped response created as the rolling element enters the fault, superimposed on the high-frequency impact generated when the rolling element re-enters the raceway from the spall. Signal information is enhanced by pre-processing the signal with Total Variation Regularization (TVR) filtration. The informative signal, extracted from the raw temporal signal via SSA, aids in the accurate identification of entry and exit events. The proposed method integrates TVR with SSA for fault size estimation and has been validated using simulated and experimental signals from independent resources. The results show strong agreement with the accuracy level of size estimation.