A Bearing Fault Diagnosis Method using Transfer Learning and Dempster-Shafer Evidence Theory

Abstract

Rolling element bearings are among the most important components in rotary machines. The reliable operation of rotary machines highly depends on the performance of bearing. Therefore, bearing fault diagnosis is a critical task in the industry. Signal-based fault diagnosis for bearings has applied extensively deep learning algorithms because of their ability to automatically extract features from fault signals measured from rotary machines. However, designing a deep learning model for any fault diagnosis problem is not a trivial task since each deep model has a complex structure and a huge number of hyper-parameters and trainable parameters. Each hyper-parameter of a deep learning model has a profound impact on the performance of that model. The selection of appropriate hyper-parameters is often conducted manually based on the Trial & Error method and experiences of the designer. Transfer learning is a technique that adopts already existing machine learning models into new domains. This technique helps to save the designing and training time of machine learning models, especially deep neural networks. In this paper, transfer learning technique is exploited to the problem of bearing fault diagnosis. A pre-trained deep neural network in the domain of image classification is adopted and modified to extract features from vibration signals measured by multiple sensors. The effectiveness of the proposed method is verified by experiments conducted with actual bearing data set supplied by Case Western Reverse University Bearing Data Center.