DEVELOPMENT OF DEEP LEARNING-BASED BEARING FAULT DIAGNOSIS METHODS

Abstract

This thesis is a systematic research on bearing fault diagnosis using machine learning algorithms, especially on deep neural network or deep learning, an emerging topic of machine learning. In the machine health monitoring area, bearing fault diagnosis is an important part because rolling element bearings are indispensable elements in rotary machines. Bearings are not only the most critical components but also the main contributor to the system failures, 45 - 55 % of equipment failure cases caused by broken of bearings. Any unexpected failure of bearings may cause sudden breakdown of the machine, even of the entire system, leading to huge financial losses. The condition monitoring of a bearing can be considered as a pattern recognition task which has been successfully solved by intelligent diagnosis methods. According to the current literature, a general intelligent diagnosis methodology includes four steps as follows: data acquisition, feature extraction, feature selection, and feature classification. Deep learning algorithms can learn multiple layers of representations from input data by deep architectures with many layers of data processing units. The output from a layer will be the input for its successive layer. Each layer can learn a higher level of data presentations from its preceding layer output. Therefore, DL architectures can automatically extract multiple complex features from the input data without human engineers. The ultimate goal of this research is to develop fault diagnosis systems using existing deep neural networks and also to find novel deep learning algorithms.