Spatial–temporal masked graph autoencoder for rotating machinery fault diagnosis with limited data

Abstract

Due to the difficulty and high cost associated with obtaining samples of faults in rotating machinery, current intelligent fault diagnosis methods struggle to extract sufficiently diverse fault features, leading to poor diagnostic performance. Additionally, existing deep learning–based approaches often focus solely on extracting temporal or spatial information from sensor signals, ignoring spatial–temporal correlations. Addressing these issues, a novel spatial–temporal masked graph autoencoder (STMGAE) framework is proposed for fault diagnosis in rotating machinery with limited data. In this proposed methodology, a masked strategy is applied to graph-structured data constructed from multi-sensor signals, enabling the model to enhance feature learning capability by training on incomplete graph-structured data masked accordingly. Furthermore, a spatial–temporal graph attention encoder module is introduced to capture temporal and spatial dependencies. To effectively reconstruct the masked portions of the graph-structured data, a correlation similarity decoder is designed, which achieves the desired outcome by capturing correlation similarity between nodes at different granularities, thus improving the model’s performance. Experimental validation of STMGAE’s effectiveness in fault diagnosis with limited data is conducted on two publicly available datasets, demonstrating superior performance compared to existing methods.