Fault diagnosis of the harmonic reducer based on dual-path convolutional network with multi-channel hybrid attention mechanism

Abstract Harmonic reducers are prone to wear and other failures due to continuous deformation of the flexure wheel during operation. To evaluate the sorts of faults in harmonic reducers under various operating situations, a fault diagnosis method (MADCNN-BiGRU) based on dual-path convolution neural network (DCNN) with multi-channel hybrid attention mechanism (MCHAM) and bidirectional gated recurrent unit (BiGRU) was proposed. Firstly, a novel MCHAM focuses on the critical information part of the vibration signal and can effectively suppress the noise and redundant information. Secondly, a new soft threshold function is constructed to regulate the attention weights dynamically. Thirdly, BiGRU obtains feature information for different time series locations to identify the faults correctly. Fourthly, a harmonic reducer test rig is established to collect vibration signals of harmonic reducers with different faults under various operating situations. The comparative experimental results demonstrate that MADCNN-BiGRU has excellent capacity for robustness and generalization, and can effectively diagnose under various operating situations.