Decision Fusion Network Based on Lightweight Multi-Head Attention Mechanism for Motor Fault Diagnosis

Motor systems are usually monitored by multiple types of sensors to obtain multi-modal measurement data, which contains multiple complementary features. However, existing fault diagnosis methods are mainly based on single modal data set. In addition, fault diagnosis method based on multi-head attention mechanism has been paid increasing attention. And the mechanism has the advantage of extracting global information and parallel computing structure. But this structure has a large number of redundant parameters, which affect the speed of computational prediction and the storage space of the computer. In this paper, a novel multi-modal multi-head attention fault diagnosis method with decision fusion and parameter pruning is proposed to alleviate the above problems. Firstly, multiple sub-models based on multi-head attention mechanism with parameter pruning are trained by different modal data to obtain several fault predicted results. Then, these fault predicted results are fused to achieve more accurate fault diagnosis result. The experiment shows that the proposed method can achieve better performance compared with the models using single modal data.