Anomaly Detection Model Based on Anomaly Representation Reinforcement and Path Iterative Modeling

In complex systems, mutual influence among sensors can lead to gradual accumulation and spread of anomalies, eventually triggering systemic failures. During this process, abnormal features evolve slowly over time, blurring the distinction between normal and abnormal patterns, and anomalies in high-dimensional spaces are difficult to detect, increasing detection difficulty. We propose an anomaly detection model based on spatiotemporal graphs ARR-PIM to address this issue, which obtains anomalous representations from multiple perspectives and models anomalous propagation paths to extract spatiotemporal features. The model consists of two core modules: the anomaly representation enhancement module and the multilevel feature extraction module. The former can enlarge the feature difference and provide prior information for potential anomaly recognition by modeling multigranularity time series as positive and negative sample pairs; the latter learns time-varying spatial features through iterative modeling of the anomaly propagation process. Experiments on six public datasets show that the proposed anomaly detection model ARR-PIM improves the average F1 score by 1.84% compared to 14 benchmark methods, significantly improving the anomaly detection performance of multivariate time series.