并行图学习与动态通道融合的水处理时空异常检测方法
By combining parallel graph learning with dynamic channel fusion, the problems of topology sensing and channel noise reduction in multivariate time series anomaly detection in water treatment systems are solved, achieving efficient detection of minute leaks and cascading faults, and improving the accuracy and robustness of detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CIVIL AVIATION UNIV OF CHINA
- Filing Date
- 2026-06-12
- Publication Date
- 2026-07-17
AI Technical Summary
Existing methods for detecting multivariate time-series anomalies in water treatment systems have shortcomings in handling physical coupling characteristics, masking mechanisms, and anomaly scoring mechanisms. They are unable to simultaneously maintain the accuracy of topology sensing and the effectiveness of channel noise reduction, and they are also unable to capture early signals of cascading faults.
A parallel graph learning and dynamic channel fusion approach is adopted. Features are extracted through a multi-scale temporal convolution module, and the first and second branches are processed in parallel. The first branch constructs a dynamic adjacency matrix to extract spatial topology enhancement features, and the second branch generates a collaborative mask matrix to perform channel collaborative feature fusion. Orthogonal constraints are used to decouple the features, and the dynamic adjacency matrix is used as a topology prior to generate a collaborative mask, thereby improving feature discriminability and redundancy removal capability.
It improves the accuracy and robustness of water treatment systems in detecting minute leaks and cascading faults in complex, high-dimensional noise environments, enhances the ability to capture physical dependencies between sensors, and reduces the impact of environmental noise.
Smart Images
Figure CN122412922A_ABST