One-Class Classification and Cluster Ensembles for Anomaly Detection and Diagnosis in Multivariate Time Series Data

Real-world automated systems such as building automation, power plants, and more have benefited from data-driven learning methodologies for anomaly detection and diagnosis. Typically, these methodologies heavily rely on prior knowledge related to abnormal operations, i.e., data points labeled as anomalies. However, in practice, such labelled data points are often unavailable which poses challenges in effective anomaly detection, particularly in diagnosis. In this paper, we propose One-class Classification Cluster ENsembles (OCCEN) anomaly detection and diagnosis approach for multivariate time series data. OCCEN utilizes one-class classification learning methods for anomaly detection followed by the decomposition of anomalies into multiple clusters. Then each cluster is treated as a binary classification problem and classifiers are trained to learn cluster representations. These trained models in combination with explainable AI models are used to generate a ranked list of diagnoses, i.e., features. Finally, we re-rank those features to account for temporal dependencies through the dynamic time-warping technique. The practical evaluation of OCCEN for air handling units (AHU) demonstrates its effectiveness in identifying faults. The framework consistently outperforms the baseline in fault diagnosis, as higher scores are observed for detection and diagnostic evaluation metrics, including F1 score, intersection over union, HitRate@k, and RootCause@k.