Chemical Process Fault Diagnosis Based on Improved ResNet Fusing CBAM and SPP

This paper proposes a fault diagnosis method based on an improved residual network (ResNet) for complex chemical processes. The method can automatically and efficiently extract fault features from the extensive data generated by the chemical operation process. The improvement is carried out in three aspects. Firstly, 1D convolution is introduced in the construction of the model to reduce the number of parameters and training time, and shortcut connections are used to alleviate the network degradation problem of traditional deep neural networks. Second, a residual-CBAM module is proposed by combining residual networks with Convolutional Block Attention Module (CBAM). This module can effectively reduce the interference of invalid targets and improve the characterization ability of the model. Finally, based on the backbone path of the network, the branching path after spatial pyramid pooling (SPP) is introduced to enable the network to extract features from different angles of the feature map and further aggregation, which improves the robustness of the model. The Tennessee-Eastman (TE) process is used as the experimental object to compare the improved ResNet with several other deep learning models. The experimental results show that the improved ResNet model achieves the best fault diagnosis results. The t-SNE method was used to visualize the fault classification process by the improved ResNet model, and the effectiveness of the improved ResNet model was further analyzed and verified.