Shuffle based Anomaly Detection in Multi-state System

The anomaly events are defined as the points that rare and diverse from the other points in feature space. Conventional anomaly detection methods usually find low-probability events with a learned a probability distribution model, or evaluate the testing samples with the local density of the testing samples. Multi-state system usually has multiple normal states, and state changes at unpredictable points caused by the daily operation such as feed, outlet, flow control, etc. For the multistate system, collecting enough data that contain all possible states are challenging or impossible to users. Furthermore, conventional anomaly detection methods are sensitive to the context of training datasets or the unpredicted phased changes of the testing datasets, or just consider the local density of the testing samples. Motivated by this problem, we transform the model learning problem to a distinction learning problem that learns the familiarity of each testing samples. In order to reduce the effects of the phased changes, we randomly shuffle the testing dataset and use a sliding window to evaluate the familiarity of the testing samples with one-class Support Vector Machine (SVM) method. Our contributions include: (1) reducing the requirement of the prior knowledge; (2) handling the phased changes of the testing datasets, (3) considering the global familiarity of the testing samples. Our proposed method is evaluated on the synthetic datasets, and the real datasets, and experiments results show that our proposed method is superior than the state-of-the-art methods.