A spatiotemporal perception framework for multivariate time series anomaly detection
By combining variable community analysis and spatiotemporal stationarity module, the spatial correlation and temporal dependence of multivariate time series are explicitly modeled, which solves the problem of ignoring the relationship between variables in existing methods, achieves more accurate and stable anomaly detection, and improves the robustness and interpretability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEST PETROLEUM UNIV
- Filing Date
- 2026-05-22
- Publication Date
- 2026-07-03
AI Technical Summary
Existing multivariate time series anomaly detection methods ignore the spatial correlation between variables, which leads to interference from weakly correlated variables in the reconstruction process, unstable reconstruction error signals, and insufficient generalization ability and robustness of the models in high-dimensional complex data scenarios.
The algorithm employs a variable community analysis-driven grouping and reconstruction module and a spatiotemporal stationarity-driven error enhancement module to explicitly model the spatial correlation and temporal dependence characteristics between variables. It groups variables and reconstructs them independently through a community detection algorithm, and combines spatial importance weighting and temporal stationarity smoothing to process the error signal.
It significantly improves the accuracy and robustness of anomaly detection, reduces noise interference, and enhances the interpretability and generalization ability of the model, making it suitable for real-world scenarios where anomaly samples are scarce or expensive to obtain.
Smart Images

Figure CN122333045A_ABST