Data-Driven Fault Diagnosis and Prognosis Framework for Bearings Using Advanced Signal Processing and Machine Learning Techniques

Abstract

Reliable fault diagnosis and prognosis (FDP) of complex engineering systems is a pressing need to prevent catastrophic failure by avoiding unanticipated problems that could lead to performance deficiencies and adverse effects on safety. In the era of the Internet of Things (IoT), the dramatic increase of sensors, data rates, and communication capabilities continue to drive the complexity of FDP applications to new levels. As a result, governments and commercial industrial communities are looking for new insights to use the massive volume of streaming in from their systems and sensors. Therefore, this dissertation presents a data-driven FDP framework for rotating bearings in large-scale industries based on advanced signal processing and improved machine learning (ML) techniques. This framework comprises of four important modules: robust condition monitoring scheme-based on time-frequency signal analysis, feature engineering (i.e. classical machine learning)- and feature learning (i.e. deep learning)-based reliable fault diagnosis methodology, and data-driven prognostics framework using new health index (HI) and variants of least-square support vector machines for remaining useful life (RUL) estimation. To determine bearing health in operation for detecting an impending failure at an early stage, this dissertation proposes a robust condition mentoring methodology for bearing failures that employs time-frequency analysis (TFA) and optimum sub-band analysis on the stream of acoustic emission (AE) signals to select informative sub-bands of the signal. There is no general consensus on how many samples and which portion of the time-domain signal should be analyzed. To address this issue, wavelet packet transform-based envelope analysis with degree-of-defectiveness ratio (WPT-EA+ DDR) evaluation matrices is applied to quantify each sub-band signal. The results of (WPT-EA+ DDR) are visualized in two-dimensional (2D) analysis tool as a percentage of degree-of-defectiveness ratio (DDR) values. This 2D tool is highly effective to select a narrow-band signal from the stream of signal that contains the most intrinsic and pertinent information about the defects. To verify the effectiveness of the proposed (WPT-EA+ DDR) sub-band selection method, this method is compared with recent state-of-the-art methods. To realize highly reliable data-driven fault diagnosis of a complicated industrial process for identifying the root cause of failures, it is important to exploit useful and discriminatory features from measured data. As a result, this dissertation proposes a hybrid feature selection (HFS) scheme for identifying the most discriminant fault signatures using an improved class separability criterion—the local compactness and global separability (LCGS)—of distribution in feature dimension to diagnose bearing faults. The HFS model consists of filter-based selection and wrapper-based selection. In the filter phase, a sequential forward floating selection (SFFS) algorithm is employed to yield a series of suboptimal feature subset candidates using LCGS based feature subset evaluation metric. In the wrapper phase, the most discriminant feature subset is then selected from suboptimal feature subsets based on maximum frequency of occurrence and maximum average classification accuracy (ACA) estimation of support vector machine (SVM) classifier using them. The effectiveness of the proposed HFS method is verified with fault diagnosis application for low speed rolling element bearings (REBs) under various conditions. Experimental results indicate that the proposed method outperforms the state-of-the-art algorithm when selecting the most discriminate fault feature subset, yielding an improvement of diagnostics performances in average classification accuracy. Furthermore, fault diagnosis in variable operating conditions is still a challenging problem since fault characteristics significantly vary with these changing conditions. An advent of deep learning (DL) has been introduced for fault diagnosis in most recent years. DL can automatically learn useful features from raw signals even with locally distorted and translated of characteristics of information. Therefore, this dissertation proposes a reliable fault diagnosis scheme under variable speed conditions based on adaptive deep convolutional neural networks (ADCNN) with acoustic spectrum imaging (ASI) of acoustic emission (AE) signals as a precise health state. In ASI, the amplitudes of the spectral components of the windowed time-domain acoustic emission signal are transformed into spectrum imaging. ASI provides a visual representation of acoustic emission spectral features in images. This ensures enhanced spectrum images for ADCNN testing and training, and thus provides a robust classifier technique with high diagnostic accuracy. To verify the proposed (ASI + ADCNN), benchmark bearing dataset with variable conditions are used. Finally, this dissertation presents a data-driven prognostic framework for rolling-element bearings (REBs). This framework infers a bearing’s health index by defining a degree-of-defectiveness (DD) metric in the frequency domain of bearing raw signal, named DD-based health index (DD-HI). Then, this dissertation systematically apply least-square support vector machines (LSSVMs) in the forms of Bayesian inference-aided one-class LSSVM (Bayesian-OCLSSVM) for anomaly detection in order to define the time to start (TTS) point of remaining useful life (RUL) prediction and the recurrent least-square support vector regression (Recurrent-LSSVR) model for predicting future values of DD-HI for calculating the RUL. In addition, this dissertation addresses several pertinent challenges, such as failure threshold determination during anomaly detection and RUL estimation, by developing adaptive thresholds. This research conducts extensive experiments on a benchmark dataset using a run-to-failure experiment. The results demonstrate the efficacy of the proposed framework compared to state-of-the-art methods in terms of the accuracy and convergence of the RUL estimation of bearings.