SRMamba-DGCN: Fault alarm recognition method based on spatio-temporal feature fusion

The growing intricacy of industrial processes has made fault alarm recognition a significant challenge in maintaining operational safety, efficiency, and reliability. The dynamic characteristics of industrial process data often exhibit unique and intricate features across both temporal and spatial dimensions. However, traditional fault alarm recognition methods that focus exclusively on original time-domain signals are limited in their capacity to capture these diverse patterns, which may compromise diagnostic accuracy. This study proposes a Self-attention Residual Mamba and Singular Value Decomposition GCN (SRMamba-DGCN) fault alarm recognition method based on spatio-temporal feature fusion, which integrates a residual Mamba unit to extract temporal features and dynamics. Concurrently, a graph convolutional neural network (GCN) is employed to model spatial relationships, capturing structural dependencies. To fuse features from these distinct domains, a multi-head attention mechanism is incorporated, ensuring an optimized combination of spatio-temporal features, enhancing fault identification. The proposed SRMamba-DGCN fault alarm recognition method is validated through extensive simulation experiments on a blast furnace process, a representative case in complex industrial systems. The results demonstrate its superior performance in terms of fault alarm recognition accuracy and robustness compared to existing methods.