Decoupling Transformer with Convolutional Fusion for Mechanical Composite Fault Diagnosis

Abstract

Compound faults often occur in mechanical systems under complex and variable operating conditions, which can seriously affect the health level of mechanical systems, so it is crucial to study the decoupling diagnosis of compound faults. In addition, industrial big data can greatly boost the accuracy and reliability of fault diagnosis results. Therefore, based on data fusion and attention mechanism, we propose a new composite fault diagnosis method called decoupling Transformer with convolutional fusion (DTCF). First, we construct the input embedding sequence as the model input. Second, the multichannel sensor data are adaptively fused using a convolutional layer. Then, the encoder encodes the signal based on global self-attention, which is representation learning. Finally, the decoupler iteratively generates decoupled labels. Two compound fault datasets are used, including the gearbox dataset and the bearing dataset, experiments on which show that the proposed method has higher accuracy, better generalization ability on the smaller training dataset, and stronger robustness against noise than CNN-based or MLP-based models. In addition, the visual analysis of attention weights makes the model interpretable.