Lightweight bearing fault diagnosis using cross-attention fusion of vibro-acoustic data and multi-scale depthwise separable convolutional network

Abstract Bearings are critical components in large rotating machinery, but fault diagnosis based on a single signal source often fails to accurately identify faults. Therefore, this paper employs a fusion diagnostic method combining vibration and acoustic signals. To address the challenges of feature redundancy, noise sensitivity, and high computational complexity in multimodal sensor fusion for bearing fault diagnosis, this paper proposes a lightweight vibration-acoustic multi-sensor cross-attention fusion network (VAMCAFN) to achieve the integrating vibration and acoustic data. This method dynamically weights bearing fault-related feature channels through a channel feature relabeling module, suppressing redundant information and reducing the impact of irrelevant data on fusion diagnosis effectiveness. We design a multi-scale separable convolutional module to concurrently extract bearing impact and modulation features across different time scales, highlighting local and periodic fault characteristics while reducing the number of parameters and computational complexity. We propose a gated cross-attention fusion module that quantifies the global dependencies between acoustic and vibration fault modes using an attention weight matrix, enabling dynamic alignment and complementary enhancement of cross-modal features. Experiments demonstrate that under complex noise conditions, the VAMCAFN diagnostic accuracy significantly outperforms unimodal models. Combining lightweight architecture with high noise tolerance, it provides a reliable solution for complex industrial scenarios.