Multi-head spatio-temporal attention based parallel GRU architecture: a novel multi-sensor fusion method for mechanical fault diagnosis

Abstract

The sensor signals with multiple measuring points and data types not only bring sufficient information, but also put forward more stringent requirements for multi-sensor fusion efficiency and fault feature extraction capability. The redundancy and conflicts in the information of multi-sensor signals often hinder the accurate extraction of crucial fault features. To address this problem, our study proposes an intelligent mechanical fault diagnosis method, which is based on a multi-head spatio-temporal attention mechanism and parallel gated recurrent units (GRUs) architecture. This method utilizes multiple attention heads to model the correlation information in spatial and temporal dimensions, and employs a parallel GRU network for targeted feature extraction. Finally, it combines local features from different attention heads to achieve flexible scheduling of various spatio-temporal attention modes. This novel application and fusion approach of multi-head attention enables accurate identification of the spatio-temporal value differences in the collected multi-sensor signals from multiple perspectives. Experimental results on multiple mechanical fault datasets show that the proposed method performs well in multi-sensor signals based mechanical fault diagnosis tasks and can maintain effectiveness under small samples and imbalanced data conditions.