Composite Fault Identification in Multi-Part Coaxial Structure Equipment Based on Convolutional Neural Network-Transformer Neural Network

Abstract In order to address the limitations of condition monitoring of multipart coaxial structure equipment, which are caused by the difficulties in placing sensors at the vibration source due to special structures or environmental constraints and the inability to accurately identify weak faults, a fault diagnosis method based on a convolutional neural network (CNN) and a transformer neural network is proposed. First, a neural network based on a transformer encoder is constructed. The original signal is then processed by a signal preprocessing method, after which the original signal and the processed signal are fused to form the input. Finally, the rich and complementary fault features are extracted by resplicing the data shape. The sliding window is utilized to stitch the one-dimensional timing signal for the initial time and input the model. Subsequently, the data are stitched for a second time and input into the one-dimensional CNN layer and embedded layer. Subsequently, a multilayer transformer encoder is introduced to capture global information. The experimental results demonstrate that the proposed method exhibits an accuracy rate of 97.2%, accompanied by a discernible clustering effect. This approach effectively addresses the challenge of accurately extracting weak fault features from composite faults, offering a promising avenue for practical applications.