Unsupervised Continual Source-Free Network for Fault Diagnosis of Machines Under Multiple Diagnostic Domains

Abstract

Data-driven based intelligent fault diagnosis (IFD) approaches have been broadly developed. In actual industry, not all data of mechanical equipment is accessible, especially in an era of increasing attention to data privacy protection. In addition, as the machine consistently operates, new data of various working conditions is continuously collected. The existence of these problems puts forward more stringent requirements for the application scenarios of algorithms, that is, how to train an intelligent model without source samples, and how to preserve the diagnostic knowledge learned from the previous task as the new data continually is collected. To exploit the application range, a novel unsupervised continual source-free network (UCSN) is proposed for IFD of rotating machinery. A one-dimension convolutional neural network is adopted as the feature extractor to extract invariant features. Based on the extracted features, a local structure clustering is used to process the target samples, which can cluster the same fault samples and separate the different fault samples simultaneously. Meanwhile, the sparse domain attention is utilized to preserve the learned knowledge and to improve the generalization of the well-trained model. In this way, the application scenarios of the IFD can be significantly expanded. Extensive experiments on a popular public bearing dataset and an actual industry dataset are carried out to verify the effectiveness of the proposed UCSN. Comparisons with ablation methods on the same experimental circumstance validate the superiority of key techniques and the proposed UCSN. The experimental results indicate that the proposed UCSN provides a promising tool for IFD of rotating machinery.