Anomaly detection in digital industry sensor data using parallel computing

The article presents the results of case studies on the anomaly discovery in sensor data from various applications of the digital industry. The time series data obtained from the sensors installed on machine parts and metallurgical equipment, and from the temperature sensors in the smart building heating control system are considered. The anomalies discovered in such data indicate an abnormal situation or failures in the technological equipment. In this study, the anomaly is formalized as a range discord, namely a subsequence, the distance from which to its nearest neighbor is not less than the threshold prespecified by an analyst. The nearest neighbor of the given subsequence is a subsequence that does not overlap with this one and has a minimum distance to it. The discord discovery is performed through the parallel algorithm for GPU developed by the author. To visualize the anomalies found, a discord heatmap method and an algorithm for selection the most interesting discords regardless of their lengths are proposed.