An Adaptive Multi-Scale Fusion and Temporal Attention-Driven Method for Gearbox Fault Diagnosis

With the large-scale deployment of rotating machinery in wind power, rail transit, and high-end manufacturing, gearboxes operating under high-speed and heavy-load conditions are prone to faults such as tooth surface pitting, root cracks, and bearing fatigue. These faults directly affect equipment safety and maintenance costs. However, many existing diagnosis approaches struggle to simultaneously capture multi-scale temporal patterns, explicitly model long-range fault evolution, and maintain robustness in strongly noisy environments. To address these challenges, this study develops an Adaptive Multi-Scale Convolutional Transformer Network (AMSC-TransNet) for vibration-based fault identification. The architecture combines an Adaptive Multi-Scale Convolutional module (AMSC), Transformer-based explicit temporal dependency modeling, and a Convolutional Block Attention Module (CBAM). The AMSC module adaptively adjusts fusion weights for different signal scales, enhancing key temporal features while suppressing operational interference. Fused features are then processed by a progressive learning architecture built with Transformer and CBAM. This enables collaborative optimization from local detail extraction and global dependency modeling to feature recalibration. Experiments on the CWRU bearing dataset and a self-constructed wind turbine gearbox dataset (DDS) show that AMSC-TransNet achieves perfect detection performance on CWRU and reaches 99.844% accuracy on DDS. Significant advantages are also maintained under strong noise (-6 dB). In summary, AMSC-TransNet demonstrates good reliability and engineering potential for weak fault scenarios and noisy environments.