DRMAF-Transformer: An End-to-End Adaptive Fusion Framework for Fault Diagnosis of Harmonic Drivers

Conventional fault diagnosis methods of Harmonic Drives(HDs) could not work satisfactorily due to their poor performance in complex working environments such as variable loads and speeds. In order to achieve high robustness and accuracy in HDs fault diagnosis under multiple abnormal situations, this paper proposed a fault diagnosis framework based on dimension reconstruction and multi-channel adaptive fusion Transformer (DRMAF-Transformer), which was an a network architecture that combined Fast Fourier Transform (FFT) and improved Vision Transformer (VIT), consisting of three stages: dimension reconstruction based on FFT, two-stage feature extraction, and adaptive multi-channel fusion. The DRMAF-Transformer was experimentally validated on four datasets of HDs from real manufacturing lines. The experimental results showed that, compared to other SOTA fault diagnosis methods, the DRMAF-Transformer could achieve higher recognition accuracy under both fixed and composite working condition data conditions.