Self-supervised Contrastive Representation Learning for Machinery Fault Diagnosis

Abstract

Supervised deep learning methods have been widely applied in the field of machinery fault diagnosis in recent years, which learns quite well the mapping relationship between the monitoring data and the corresponding labels. In many practical industrial applications, however, the monitoring data are unlabeled due to the requirement of expert knowledge and a large amount of labor, which limits the use of supervised methods for fault diagnosis. In view of this problem, in this work a self-supervised representation learning method named contrastive predicting coding (CPC) is employed to automatically extract high-quality features from one-dimensional machinery monitoring signals without the requirement of labels. The method is validated on a benchmark dataset for bearing fault diagnosis, and the quality of extracted features of different classes is quantitatively evaluated. The results show that features extracted by the CPC method are more representative than those obtained by autoencoders and statistics.