An Anomaly Detection Algorithm Based on Hidden Pattern
基于隐含模式的异常检测算法

It is a difficult problem how to detect such accident of a system. This paper presents a new algorithm, an anomaly detection algorithm based on hidden pattern. Epsilon machine, a new computational mechanics, can discover hidden pattern from the response time series. Causal State Splitting Reconstruction (CSSR), one algorithm of epsilon machine, can infer a set of causal states, which has an analogy to hidden Markov chain. Based on this set, an anomaly measure can be defined, which is the distance of two characteristic vectors. Computing all parts of the time series, an anomaly evolution curve can be got. In simulation analysis of Duffing equation, step changes appear in the anomaly curve, before Duffing oscillator begin to bifurcate. The algorithm proves to be effective in anomaly detection and warning.