Degradation Assessment And Residual Life Prediction Of Rolling Bearings Based On Multiple Features

Rolling bearings, as critical components of mechanical systems, play an important role in determining the performance of the entire system. The maintenance plan may be inefficient or catastrophic failures of the entire system may be incurred if the health status/damage level of rolling bearings cannot be detected or predicted in a timely manner. Therefore, degradation assessment and prediction of rolling bearings are of help to engineers to monitor the damage level of a system and prevent unexpected failures. In most recent years, tremendous attempts have been made to assess and predict the degradation of rolling bearings from the data-driven perspective. The signals and features collected from sensors can, more or less, reflect the degradation of rolling bearings. Nevertheless, many existing methods merely considered a few representative features and cannot integrate multiple features from various domains. This thesis devotes to investigate a set of new data-driven degradation assessment and remaining useful life prediction approaches based on multiple features extracted from vibration signals. The four basic elements of data-driven approaches, i.e., feature extraction, feature selection, degradation assessment and residual life prediction, have been studied to achieve the aforementioned goal. The major contributions of this thesis are summarized as follows:(1) As the collected vibration signals contain a plenty of noises, the wavelet threshold denoising approach is firstly used to pre-process the raw signals so as to remove the high-frequency noises. Secondly, the feature extraction is executed to extract useful features, such as the time-domain features, the frequency-domain features, the wavelet node energy and so forth, from the original vibration signals collected from sensors mounted on horizontal and vertical directions. Additionally, the entropy characteristics of vibration signals are also extracted, including the time-domain information entropy, the frequency-domain information entropy, the Hilbert entropy, the sample entropy, and the wavelet packet energy entropy. Thirdly, an unsupervised feature selection strategy with three selecting criterions is utilized to select the most sensitive features among 96 features extracted from our study to facilitate residual life prediction of rolling bearings.(2) The original vibration signals of rolling bearings during operational stage are non-stationary and non-linear in nature. The Local Mode Decomposition(LMD) that excels at processing AM-FM complex signals is used to decompose the original vibration signals into a finite number of Product Functions(PFs), each of which has a special physical meaning. PFs can somewhat reflect the trend of weak fault signals. The time- and frequency-domain features can be extracted from the PFs of the LMD. Then, the Principal Component Analysis(PCA) is performed to reduce the number of dimensions of the extracted features. The indicators of degradation assessment are developed by incorporating the PCA and the correlation analysis. The effectiveness of the proposed indicators is demonstrated via a set of experimental tests.(3) The indicators of degradation assessment are critical to predict residual life of rolling bearings. In this work, the Relative Root Mean Square(RRMS) that overcomes the heterogeneity of rolling bearings is firstly used to assess the degradation trend of rolling bearings. The lifecycle of rolling bearings is divided into three phases by the pre-specified thresholds for the initial flaw and the final failure. Secondly, by taking advantage of the support vector machine(SVM) in dealing with small samples and the relevance vector machine(RVM) in terms of yielding probabilistic outputs, a new multi-feature residual life prediction approach based on SVM and RVM is developed. The effectiveness of the proposed method is verified by the data sets collected from an accelerated life test of rolling bearings. The errors of predictions are also compared between SVM and RVM.