EMT-HybridNet: Dual-branch CNN–transformer with edge-guided texture enhancement for bearing fault diagnosis

Abstract

Diagnosis of bearing faults is pivotal for ensuring the safe, reliable, and uninterrupted operation of rotating machinery. However, the efficacy of conventional deep learning-based diagnostic frameworks is often compromised by the limited representativeness of single-channel signal inputs and the limited discriminative capability of shallow feature-extraction architectures, hindering accurate and robust bearing-fault classification. To overcome these challenges, this study introduces EMT-HybridNet, a novel lightweight framework for intelligent bearing fault diagnosis built on three core innovations. First, a dual-representation signal-to-RGB conversion strategy generates complementary time–frequency and transform-free RGB images from raw vibration signals, jointly capturing frequency-domain and time-domain fault signatures that single-representation approaches cannot exploit. Second, an edge-guided multi-scale texture (EMT) enhancement module strengthens fault-discriminative textures and structural patterns in the transform-free branch by enhancing weak fault signatures and suppressing background redundancy prior to feature fusion. Third, a dual-branch CNN architecture with a compact ResNet-50 backbone, learnable weighted feature fusion, and a single-layer Tiny Transformer captures complementary features and global contextual dependencies with minimal computational overhead. The fused features are passed through a SoftMax classifier to identify bearing health conditions. The robustness and effectiveness of EMT-HybridNet are validated on two benchmark datasets, the experimental dataset and the CWRU dataset, through comparison with several state-of-the-art baseline methods, demonstrating superior robustness and effectiveness.