A novel framework based on adaptive Multi-Task learning for bearing fault diagnosis

Abstract

Bearing fault diagnosis is very important for the security and efficiency of electric machines. In recent years, the newly emerging deep learning methods have risen bearing fault diagnosis as a research hotspot again. To achieve better performance and interpretability and absorb novel methods, this paper proposes a novel framework based on adaptive Multi-Task Learning for bearing fault diagnosis, Gramian Angular Fields - Markov Transition Fields - Convolutional Neural Network (GM-CNN), including data augmentation, image encoding methods and adaptive Multi-Task Learning (MTL). Firstly, data augmentation (DA) methods are applied in data preprocessing for tackling the problems of lack of data and weak generalization ability. They include cropping, flipping, and noise injection. Secondly, Gramian Angular Fields (GAF), Markov Transition Fields (MTF), and their combination GAF-MTF are used to encode time series into images which are more proper for convolutional neural network (CNN) to extract features and classify. Then, the processed data are fed into the proposed MTL framework, and tasks for classifying the fault type and severity are trained jointly as they share some general knowledge and this saves more time. Besides, attention is applied to make the MTL adaptive, which is helpful for more balanced training. Some experiments are carried out. And the experiment results show that the proposed framework is relatively simple but more effective, classifying the fault type and severity with high accuracy (basically higher than 99%). It is shown that every step of the framework is important and essential. This paper provides a reference for future studies on bearing fault diagnosis from the perspective of feature extraction and CNN. It could also be applied to other time series classification situations which could be promising directions.