O-LoMST: An Online Anomaly Detection Approach And Its Application In A Hydropower Generation Plant

摘要

With the increasing availability of streaming data, industries nowadays are striving for an automated online anomaly detection algorithm that can analyze data stream and detect anomalous patterns in real time. Such an online algorithm should detect anomalies on the fly, without storing all, or a very long stretch of, the historical data. It should be able to update its control mechanism for anomaly detection upon receiving new data. Moreover, the algorithm must work in an unsupervised way; i.e., in the absence of class labeling information a priori. These fundamental requirements limit the application of traditional anomaly detection approaches in streaming scenarios. In this paper, we introduce an online anomaly detection method, based on an offline method recently developed. The prototypical offline method is one of the new approaches that specifically handle the issue of nonlinear manifold embedding in data spaces and use a minimum spanning tree to approximate and capture the manifold structures, leading to a much enhanced detection ability. The primary objective of this paper is to make the offline method applicable to streaming data and address the aforementioned unique online issues. We elaborate the steps of our proposed approach by applying it to a hydropower generation plant and demonstrating how it can contribute to automation in that context.