An End-to-end Fault Diagnosis Method: Dual-branch Network Feature Fusion Model Combined with MSACNN and Bi-LSTM

With the advancement of informatization and intelligent transformation in the manufacturing industry, the reliability and stability of manufacturing equipment have become crucial. Traditional equipment fault diagnosis methods only use single-branch convolutional neural networks (CNN) to extract vibration signal features, focusing on extracting local detailed features while lacking attention to the global features of the vibration signals. To address this issue, this paper proposes an end-to-end equipment fault diagnosis method based on a dual-branch feature fusion model, which combines the advantages of multi-scale attention convolutional neural networks (MSACNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) in signal feature extraction. This approach focuses on calculating the global relationships of vibration signals while paying attention to local subtle feature information, and the two branches complement and fuse. Furthermore, to address the feature fusion problem of vibration signals, a multi-scale feature fusion module is designed to treat the global and local representations of vibration signals equally in the classifier, thereby avoiding the problem of gradient vanishing in subsequent fully connected layers. Experimental results show that compared to existing fault diagnosis methods, this method can adapt to variable noise interference and has higher accuracy.