WCFormer: a wavelet-enhanced CNN-transformer hybrid network for bearing fault diagnosis using multi-sensor signal fusion

Abstract Bearings are core components of rotating machinery, and their failures can cause significant production accidents. Current multi-source data fusion methods and independent network architectures show limited diagnostic performance on small and medium-sized datasets. To address multi-sensor data fusion and feature extraction in bearing fault diagnosis, we propose MCWT-WCFormer, a lightweight convolution-Transformer hybrid network with three key modules: MCWT, CSAN, and WAFN. 1. MCWT transforms multi-source signals into grayscale images through wavelet transform, stacks them into RGB format, integrating time-frequency information. 2. CSAN generates spatial and channel information descriptors and dynamically weights feature maps. 3. WAFN extracts high-frequency features of the Transformer’s Key by introducing wavelet transform convolution, realizing joint learning of local-global features. MCWT-WCFormer optimizes efficiency and performance by leveraging the inductive bias of CNN and the scalability of the Transformer. Cross-evaluated on HUST-gearbox and SHU-TSTB datasets, MCWT-WCFormer achieves 98.12%±0.17% and 98.03%±0.12% accuracy, respectively, with a single sample diagnosis time of about 4.2ms while having the lowest complexity (43.25 GFLOPs) and parameters (5.91 M). It can be integrated into industrial digital twin systems cost-effectively while supporting new metrics like cyclic-correntropy. It is extendable to rotating machinery health management like steam turbines and wind turbine gearboxes.