Multiattention Spatiotemporal Fusion Graph Neural Network for Chemical Process Fault Diagnosis

Fault diagnosis of complex industrial processes becomes challenging due to temporal and spatial dependencies in process data. This means that the emergence and evolution of faults are affected not only by temporal factors, but also by the spatial relationship between individual components and locations. To address these challenges, multiattention spatiotemporal fusion graph neural network (MA-STGNN) is proposed in this article. First, fault features are represented as fully connected graphs. Considering that some nodes are not closely connected to each other, edge-based self-attention is employed to obtain sparse graphs with dual channel. Second, a spatiotemporal fusion graph convolution neural network (STGCN) block by a fusion of convolutional neural networks (CNNs) with graph convolutional neural network (GCN), is proposed to capture the temporal and spatial dependencies of the graphs in each channel. Moreover, in order to alleviate information loss, graph-based interlayer attention is employed to compute weighted sums of the outputs from each STGCN block, and then the graph embeddings are obtained. Third, node-based self-attention is used to focus on nodes in the graph embedding that are more critical to the target task, so as to generate a graph encoding for each channel. By concatenating the dual-channel graph encodings, the graph representation is obtained. Finally, the graph representation goes through fully connected layers and then uses a softmax classifier to complete the fault diagnosis task. Experimental results on the Tennessee Eastman (TE) process, three-phase flow facility (TFF), and PROcess NeTwork optimization (PRONTO) benchmark case demonstrate the high accuracy and robustness of the model in fault diagnosis.