A Residual Multihead Self-Attention Network Using Multimodal Shallow Feature Fusion for Motor Fault Diagnosis
Juan Feng, Jinya Su, Xiaoliang Feng
Abstract
Fault diagnosis (FD) is a crucial task in motor-based industrial systems, but traditional signal analysis-based methods face various challenges that make it difficult to achieve high-level performance. This article proposes a novel approach called residual multihead self-attention network using multimodal shallow feature fusion (ResMHA-MMSFF) for motor FD. The method first applies an improved feature fusion technique based on a gating mechanism (GM) to combine the datasets provided by both vibration and current signals. The fused features are then introduced into the residual multihead self-attention network to obtain more detailed information. Finally, a SoftMax function is adopted to classify faults and diagnose motor system faults. Comparative experiments under different loads and speed conditions are conducted to demonstrate the effectiveness of the proposed method, which outperforms existing single-signal or other deep learning-based methods.