MGTN-DSI: A multi-sensor graph transfer network considering dual structural information for fault diagnosis under varying working conditions

Abstract

In recent years, mechanical systems have increasingly integrated multiple sensors to monitor equipment status more effectively. However, extracting domain-invariant features from multi-sensor data and adapting the diagnostic model to significant variations in operating conditions remain challenging tasks. To address these issues, a novel multi-sensor graph transfer network considering dual structural information (MGTN-DSI) is designed for fault diagnosis under varying working conditions. Firstly, we develop an advanced multi-sensor feature extraction mechanism.Specifically, on the one hand, a multi-sensor collaborative fusion layer is proposed to uncover the intrinsic connections among different sensor data. On the other hand, a relationship graph between data points of the fused high-level features is constructed using a graph convolutional network with constrained filters. Secondly, a joint alignment method using a virtual discriminator is proposed to simultaneously align both subdomain and global distributions. The extensive experiments conducted on a public fault diagnosis dataset and a practical fault diagnosis test platform indicate that the proposed MGTN-DSI has higher accuracy and better generalization ability than other state-of-the-art comparison methods.