Unsupervised Multi-sensor Anomaly Localization with Explainable AI

Multivariate and Multi-sensor data acquisition for the purpose of device monitoring had a significant impact on recent research in Anomaly Detection. Despite the wide range of anomaly detection approaches, localization of detected anomalies in multivariate and Multi-sensor time-series data remains a challenge. Interpretation and anomaly attribution is critical and could improve the analysis and decision-making for many applications. With anomaly attribution, explanations can be leveraged to understand, on a per-anomaly basis, which sensors cause the root of anomaly and which features are the most important in causing an anomaly. To this end, we propose using saliency-based Explainable-AI approaches to localize the essential sensors responsible for anomalies in an unsupervised manner. While most Explainable AI methods are considered as interpreters of AI models, we show for the first time that Saliency Explainable AI can be utilized in Multi-sensor Anomaly localization applications. Our approach is demonstrated for localizing the detected anomalies in an unsupervised multi-sensor setup, and the experiments show promising results. We evaluate and compare different classes of saliency explainable AI approach on the Server Machine Data (SMD) Dataset and compared the results with the state-of-the-art OmniAnomaly Localization approach. The results of our empirical analysis demonstrate a promising performance.