Dual-Stream Cross-Modal Feature Fusion Based on Multi-Scale Attention for Industrial Fault Diagnosis

The characteristics of non-stationarity and non-linearity present a challenging task for the fault diagnosis of bearings. Traditional feature fusion methods did not show satisfying capacity in information extraction and therefore confront difficulties in differentiating bearing fault states. To fully exploit correlated information and enhance cross-domain feature fusion, a novel dual-stream cross-modal feature fusion approach based on multi-scale attention for fault diagnosis is proposed in this paper. To effectively preserve the sequence signals and their temporal dependencies, the Gramian Angular Field (GAF) encoding technique is leveraged to transform raw time-domain signals into two-dimensional images. Besides, a multi-scale channel attention module (MSA) is introduced in the convolutional neural networks which could address the challenges related to inconsistent scales and feature distributions when fusing different modalities, and it enables dual-stream cross-modal (DSCM) data to exchange information and integrate complementary features. Experimental evaluations conducted on the Case Western Reserve University dataset demonstrate that the proposed method DSCM-MSA achieves higher accuracy and efficiency in cross-modal bearing fault diagnosis tasks compared with other methods, and can therefore provide a reliable foundation for industrial fault diagnosis.