STR-AF: An Adversarial Monitoring Framework with Spatial-temporal Representation and Adaptive Fusion
Chenying Zhu, Wenqian Zhang, Kai Zhong · 2024
The structure of modern industrial processes has become increasingly complex. Multiple sensors are placed in the system to obtain their operating status. From a spatial perspective, different process variables experience complex interactions through control loops. And because of process dynamics, the variables usually have a temporal characteristic. Often, these two dependencies have different manifestations, indicating different process characteristics. However, the existing methods can not separate spatial and temporal information well, which leads to inaccurate process monitoring performance, and the existing methods still have problems such as model collapse, poor generalization ability and poor accuracy. Therefore, a spatial-temporal representation and adaptive fusion adversarial monitoring framework (STR-AF) is proposed to deal with spatial-temporal characteristics, so as to better identify industrial process faults. Specifically, three different encoders are built to process temporal, spatial, and global information respectively, and integrate the three features organically. Finally, the two decoders are trained against each other to optimize the model. The effectiveness of STR-AF method is verified by experiments.