Spatiotemporal Transform Network-Based Anomaly Detection and Localization of Distributed Parameter Systems

Due to complex spatiotemporal couplings, it is difficult to detect and locate spatiotemporal abnormal sources for distributed parameter systems (DPSs) with unknown governing equations. In this research, a spatiotemporal transform network-based anomaly detection and localization framework is proposed for unknown DPSs. Considering the orthogonality, the spatial basis functions (SBFs) are optimized by the nonlinear space–time separation network to achieve the minimal reconstruction error. The Gaussian process regression is used to identify the temporal dynamics, based on which the temporal statistic is constructed. A comprehensive statistic is designed by considering the temporal dynamics and spatial dissimilarity for reliable detection. With the spatial construction, the weighted absolute error of SBFs is constructed for anomaly localization. The anomaly detectability is proven by theoretical analysis. Experiments on a lithium-ion battery demonstrate the effectiveness and superiority of the proposed method in detecting and localizing battery internal short circuits.