Sliding Window Online Anomaly Detection for Dynamic Systems with Unsupervised Algorithms

Online Anomaly detection in operational data from dynamic systems is challenging due to the inherent complexity and variability of these datasets. This research introduces an innovative approach that leverages the Sliding Window Online Anomaly Detection method combined with online learning to enhance unsupervised algorithms for online anomaly detection within time-series operational data. Compared to traditional offline anomaly detection models, this method offers a significant advantage by allowing the training set to be updated incrementally in real-time. Several unsupervised algorithms were evaluated under different conditions: insertion of outliers, insertion of random noise and real fault data. Our method shows that traditional time-series data anomaly detection algorithms can be further optimized in conjunction with online learning to perform well in dynamic systems.