Sensor fusion transformer for interpretable anomaly detection via multi-sensor time series data

Abstract

Advances in the Internet of Things (IoTs) and smart sensing technology have facilitated the development of multi-sensor networks for detecting anomalous behaviors from the system monitored in various areas. general, sensor data in the multi-sensor network are highly correlated and have nonlinear dynamics with time dependencies. It is crucial to detect abnormal patterns from a monitoring system at a much earlier stage and provide the interpretability for the anomalies to proactively respond to them. We propose a sensor fusion transformer (SFT) prediction model with a parallel architecture that permits data representations using time and sensor-dimensional features. SFT mainly aims to detect time trends and identify sensor inter-relationships in multi-sensor systems. To identify influential sensors related to the anomalies, we propose an algorithm by decomposing activated feature maps generated from the SFT architecture. The effectiveness of the proposed method is validated using real-world datasets collected from a variety of multi-sensor systems. Finally, through an ablation study and sensitivity analysis, we validate that the proposed method has potential to timely manage anomalies in the early stage by identifying the sensors that may measure root causes of the anomalies and pinpointing the times of abnormal changes in multi-sensor time series data.