Parallel Temporal and Spatial Modeling for Interpretable Fault Detection and Isolation of Industrial Processes

The structure of modern industrial processes has been gradually complicated to adapt to diversified production requirements. Process variables generally have temporal characteristics. Meanwhile, complex spatial interactions between variables also pose challenges for process modeling. In this study, a parallel temporal and spatial feature extraction framework is proposed and applied to fault detection and isolation in industrial processes. On the one hand, unlike the existing methods with mixed spatiotemporal information, we design independent temporal and spatial modeling structures. The temporal characteristics of each process variable are simultaneously extracted in the designed temporal submodule. Furthermore, the spatial connections between variables are captured through sparse adjacency network extraction and information fusion. In this way, the temporal and spatial information can be individually observed to provide interpretable monitoring results. On the other hand, considering the different spatiotemporal anomalies caused by various fault types, we establish a targeted isolation strategy to provide reliable fault analysis. For temporal faults, the reconstruction error indicators are designed to quantify the abnormality of the variable. Moreover, a network reconstruction model is developed to measure the spatial structure deviation and locate the fault source. The performance of the proposed method is verified through a real industrial example.