Predictive Model Based on Dynamic Multitemporal Feature Graph Networks and Application to Industrial Process Anomaly Monitoring

Aiming at the problems of spatial coupling between process parameters and time dependence in actual industrial processes, which lead to insufficient reliability and poor timeliness of process anomaly monitoring, a predictive model of time-series graph slicing networks based on dynamic multitemporal features is proposed, which can effectively reduce false alarms and omissions in process anomalies and explain the causes of anomalies at the same time. First, a method for extracting and building graphs of multidimensional spatiotemporally coupled time-series data based on the data dimensionality reduction method and causality calculation is designed to mine the potential associations between strongly coupled process parameters. Second, a dynamic multitemporal feature graph embedding strategy and a two-channel graph structure learning network are proposed to effectively solve the problem that individual time-dimension features cannot fully utilize the continuity and sequentiality in the time dimension and to realize the long-time dynamic perception of global variables. Finally, the prediction performance of the proposed method as well as the anomaly monitoring performance is validated using the time-series dataset generated by the MATLAB-based Tennessee Eastman (TE) process platform as well as the SWaT process dataset. The experimental results show that the proposed model has an average fitting degree of more than 0.99 for the prediction of target variables, which can identify and explain the anomalies more quickly and accurately compared with other typical methods, and significantly improves the reliability and timeliness of the anomaly monitoring of industrial processes.