LAUSR

The supervisory control and data acquisition (SCADA) system provides information that can be used to free humans from laborious monitoring tasks, such as control performance monitoring (CPM). However, the existing CPM methods rely heavily on the quality of SCADA data. In practice, the missing of measurement and computed signals due to some random and man-induced factors will lead to failures of traditional CPM methods. This article develops a robust CPM model for varying-dimensional time-series data resulting from the missing variables in SCADA systems. Two attractive advantages of the proposed model are noticed. First, SCADA data with various variable dimensions and missing patterns can be handled through a structural feature extraction (SFE) module, which constructs specific graphs for input data and explicitly explores the inherent interaction mechanism among variables. A structural vector is then generated to characterize the interaction pattern of multiple variables. Second, the proposed model is designed with the generalization ability by developing parameters-shared node-effect and edge-effect graph neural networks (GNNs). In this way, the method shows good robustness to the previously unseen missing patterns. Experiments on the simulated and real datasets demonstrate the feasibility of this method.