Anomaly Detection in Time Series via Correlation-Guided Feature Reduction and Residual Distribution Characterization

Time series anomaly detection plays a critical role in ensuring the reliability and safety of modern industrial systems. However, the presence of uninformative yet dominant redundant features, complex temporal dependencies, and structural anomaly patterns poses significant challenges to traditional detection methods. To overcome these limitations, this work introduces a Transformer-based framework that integrates correlation-guided feature reduction to suppress the influence of redundancy, and a residual distribution characterization module to enhance the sensitivity and robustness of anomaly scoring. This joint design not only improves the model's prediction efficiency but also enhances its ability to capture anomaly behaviors in complex temporal patterns. First, a global correlation analysis is conducted to reduce redundant input features and improve model efficiency. Then, a stride-controlled sliding window mechanism divides the data into structured prediction intervals, enabling short-horizon forecasting and multi-window residual analysis. Finally, a residual distribution characterization module is designed to compute fused scores by integrating variance, autocorrelation, and skewness extracted from prediction residuals. On a real system time series dataset, the proposed approach achieves an improvement of 0.036 in F1-score and 0.104 in precision over the multi-window Mean Squared Error method with comparable recall performance. These results validate the effectiveness of both the correlation-guided feature reduction strategy and the residual distribution characterization mechanism in capturing anomalies under complex temporal dynamics.