Artificial Intelligence Techniques for Bearing Fault Diagnosis

Abstract

The condition monitoring of bearing by using vibration signal is an indispensable mission in today’s manufacture processes. To develop the solution for monitoring the conditions of the bearing, signal processing plays a crucial role for finding out the meaningful information from the vibration signals. The existing signal processing approach present in the current literature mainly focuses on extracting the meaningful information from the bandwidths which contains the fault frequencies. However, when the system is changed, for the new dataset, these approaches failed to extract meaningful information. Moreover, due to noise, and non-stationarity, not always the fault frequencies are appeared into the expected region. The only solution is to explore and analyze the complete dataset from that new environment to make an adjustment with the signal processing approach. Currently, several Machine Learning (ML) and Deep Learning (DL)- based models have attained brilliant results in fault detection and diagnosis under consistent working conditions. Generally, the successful ML models consist of 4 stages, i.e., (1) data preprocessing, (2) feature extraction, (3) feature selection, and (4) classification. However, due to non-linearity, and non-stationarity, it is very difficult to extract and analyze the fault feature information from variable working conditions with the existing preprocessing steps. Therefore, the extracted feature information becomes easily vulnerable when the dataset along with the working conditions are changed. Moreover, the existing feature selection techniques lack of explain ability, which makes it difficult to interpret/justify the performance of the classifier. Furthermore, so far, the existing classifiers are used like a black-box model, which makes it even harder to debug the decision of the classifier. Thus, for every new dataset or application domain, there is a necessity to build a model from scratch. Therefore, to solve all these problems, a concept of explain ability is proposed in two steps for the first time in the field of bearing fault diagnosis: (a) incorporating explain ability of the feature selection process, and (b) interpretation of the machine learning classifier performance with respect to the selected features. In this dissertation, an explainable ML based fault diagnosis model for bearing is proposed with 5-stages, i.e., (1) a data preprocessing method based on a Faster Discrete Orthogonal Stockwell Transformation (FDOST) Coefficient is proposed to analyze the vibration signals for capturing the invariant patterns from both time-frequency, and corresponding phase-angel information for variable speed, and load conditions, (2) a statistical feature extraction method is introduced to capture the significance from the invariant pattern of the preprocessed data obtained by FDOST, (3) an explainable feature selection process is proposed by introducing a wrapper based feature selector - Boruta, (4) a feature filtration method is proposed after the feature selection process to avoid the multicollinearity problem by introducing Spearman’s rank correlation coefficient, and finally, (5) an additive Shapley explanations followed by k-NN is proposed to diagnose, and to explain individual decision of the k-NN classifier for understanding, and debugging the performance of the model. Thus, the idea of explainability is introduced for the first time in the field of bearing fault diagnosis in two steps: (a) incorporating explainability to the feature selection process, and (b) interpretation of the classifier performance with respect to the selected features. Further, to extend the explainable model to mitigate the training time, and to automate the feature extraction, selection, and filtration process, a Transfer Learning (TL) based DL algorithm is proposed by utilizing the visual patterns obtained from FDOST time-frequency coefficient-based Vibration Imaging (VI). The effectiveness of each proposed model is demonstrated on two different datasets obtained from separate bearing testbeds containing different mechanical faults in rotating machinery along with variable load, and speed conditions. Lastly, an assessment of several state-of-art fault diagnosis algorithms in rotating machinery is included.