A multisource large-scale fault diagnosis method for traction drive systems based on a temporal-spatial joint perception graph convolutional network
TL;DR
A multisource fault diagnosis method based on a temporal-spatial joint perception graph convolutional network (GCN) that adaptively updates the adjacency matrix according to the learned Temporal-Spatial feature relationships, enabling dynamic perception of sensor coupling under different fault states and operating conditions.
Abstract
To address the challenges of large-scale fault diagnosis in traction drive systems, such as complex fault categories, strong background noise, insufficient multi-source information fusion, and difficulty in modeling dynamic sensor correlations, this article proposes a multisource fault diagnosis method based on a temporal-spatial joint perception graph convolutional network (GCN). The main research idea is to construct a unified diagnostic framework that integrates signal enhancement, time-frequency representation, global multisensor dependency modeling, and adaptive spatial graph learning. First, singular value decomposition is introduced as a front-end denoising strategy to suppress background interference while preserving fault-related dominant frequency and impulse components. Then, continuous wavelet transform is employed to convert the denoised multi-sensor vibration signals into time-frequency representations, thereby enhancing the expression of non-stationary fault features. On this basis, an Inception V3-like multi-scale feature extraction module is designed to capture fault-sensitive information from different frequency bands. A Transformer module is further introduced to model global contextual dependencies among multi-source sensor feature tokens, while a GCN is used to learn spatial relationships between sensors. Different from conventional static graph-based diagnostic models, the proposed method adaptively updates the adjacency matrix according to the learned Temporal-Spatial feature relationships, enabling dynamic perception of sensor coupling under different fault states and operating conditions. Experimental validation on the Beijing Jiaotong University–Rail Autonomous Operations (BJTU-RAO) bogie dataset demonstrates that the proposed method achieves high diagnostic accuracy and stability in a 51-class large-scale fault diagnosis task, outperforming traditional static GCN networks and several existing fault diagnosis methods. The results indicate that the proposed framework provides an effective solution for multisource, large-scale, and complex fault diagnosis of traction drive systems.