Research on local-global collaborative anomaly detection algorithm based on entropy weight fusion and its application
Anomaly detection in complex time-series data is a fundamental task in fields such as financial monitoring, industrial systems, and intelligent forecasting. Existing methods often suffer from limited adaptability when simultaneously handling local transient anomalies and global structural anomalies, especially under varying contamination levels and heterogeneous data distributions. To address this issue, this paper proposes a local-global collaborative anomaly detection framework named EWC-LOFAD, which combines local kernel regression residual analysis with adaptive density clustering. A local linear regression model with generalized cross-validation (GCV) bandwidth selection is used to capture short-term fluctuations, while an adaptive DBSCAN mechanism detects global structural outliers. An entropy-weighted fusion strategy is further introduced to dynamically balance local and global anomaly information. Experimental results on synthetic datasets with varying anomaly rates demonstrate that EWC-LOFAD competitively outperforms several benchmark methods including Isolation Forest, Local Outlier Factor, and DBSCAN, achieving up to 40.3% improvement in F1-score under low contamination settings and achieving a high precision of 0.9837 at a 10% contamination level. Applied to exchange rate data, EWC-LOFAD not only identifies anomalies detected by LOF, Isolation Forest, One-Class SVM and DBSCAN, but also reveals hidden anomalies associated with major financial events. Further validation via synthetic anomaly injection on real datasets demonstrates that EWC-LOFAD achieves superior precision, recall and F1-score, verifying its reliability, robustness and practical applicability in complex real-world time series analysis.