Feature-distance-based Online Anomaly Detection for Dynamic Systems

Traditional statistical anomaly detection methods struggle with the high-dimensional and highly dynamic characteristics. With the advancement in computing and monitoring capabilities, machine learning-based anomaly detection methods for dynamic system equipment have become a major research focus. However, due to the infinite number of dynamic operating states inherent in dynamic equipment, it is difficult to enumerate them all. Therefore, this paper introduces a dynamic feature extraction method based on a deep temporal model with an encoder-decoder structure. This method accumulates long-term historical dynamic features and constructs a rule engine for online anomaly detection based on inference engine-driven rule extraction, aiming to cover the system's dynamic operating space as comprehensively as possible. Taking a typical dynamic thermal system, gas turbine, as an example, the paper performs feature extraction, feature library construction, and anomaly rule extraction on the training dataset. The proposed anomaly detection method based on feature distance are verified. With the accumulation of time and personnel calibration, rule extraction can be further quantified, and there is potential to cluster and classify dynamic operating conditions within the feature space.