Research on Adaptive Model Pooling Method for Data Stream Anomaly Detection Based on Concept Drift Identification Strategy
Mingwei Li, Zimeng Fan, Lei Song, Lili Guo · 2025
Anomaly detection in data streams is critical for intelligent diagnostics, where accurate pattern recognition supports equipment health monitoring. Existing adaptive model pool frameworks face challenges in distinguishing concept drift from true anomalies due to passive adaptation mechanisms, leading to imbalanced detection performance. This study proposes a dynamic identification strategy incorporating concept drift priors. The approach integrates reservoir computing for drift pattern capture and dynamic model pruning with feature repository matching, enabling active learning of distribution shifts while maintaining parameter stability during anomalies. The synergistic mechanism enhances differentiation between concept drift and anomalies, improving detection robustness. Experimental evaluations on benchmark datasets and spacecraft telemetry data demonstrate the framework’s superior accuracy in both anomaly detection and drift recognition compared to existing methods.