An Explainable Deep Learning-based Approach for Multivariate Time Series Anomaly Detection in IoT

Detecting anomalies from Internet of Things (IoT) data is important for a smooth flow of events. Timely detection of anomalies and raising alarms can help avoid any potential health and safety hazards in an IoT environment such as detecting falling people by activity recognition using wearable sensors. However, most IoT data lacks labels and sometimes when the labels are available, the context is missing, which makes anomaly detection and inference a challenging task. In this paper, we propose an explainable deep learning-based approach to detect anomalies from multivariate time series data using Long Short-Term Memory (LSTM) and the results are explained using the Local Interpre table Model-agnostic Explanations (LIME). The experiments show promising classification results with good accuracy, precision, recall, and F1 Score values. LIME provided useful explanations for anomalous and non-anomalous instances. The effectiveness of LIME for time series anomaly detection is discussed, and its limitations are also highlighted.