Multidimensional Time Series Analysis for Anomaly Pattern Detection and Interpretation

The interpretability of multivariate time series anomaly detection is crucial for understanding the reasons behind anomalies, enhancing the usability and credibility of models, and ensuring successful real-world applications. Although there has been considerable research on the interpretability of time series, most are based on black-box models and have certain limitations in revealing the underlying patterns of anomalies. To address this issue, this paper focuses on identifying anomaly patterns in multivariate time series using a white-box machine learning approach, which overcomes the interpretability limitations of traditional methods in high-dimensional dynamic data processing. This approach innovatively enhances the application of Dynamic Time Warping (DTW) in higher dimensions to effectively capture anomaly patterns of any length, while identifying both short-term and long-term anomalies. Experiments on the Yahoo dataset demonstrate its high accuracy and reliability in anomaly pattern detection, particularly in explaining the reasons and impacts of identified anomaly patterns. This innovation marks a breakthrough in anomaly pattern interpretation, showing significant practicality and interpretability in real-world applications, and provides vital support for decision-making and predictive modeling in time series analysis.