Fuzzy dynamic Markov model for Time Series Anomaly Detection

Dynamic Markov model are commonly utilized in time series anomaly detections. The traditional dynamic Markov models define data state using equal width interval segmentations. However, the state definition method is too rigorous. In real applications, the data may vary in a certain range and it cannot always belong to a specific state. Therefore, we proposed a new fuzzy interval segmentation scheme, which dispatch each data into three neighbored fuzzy states to tolerate data variations. Furthermore, we also introduced model updating schemes to adapt to the new data series. The comparison results demonstrated that our method obviously reduced the false detection rate of time series data compared with the traditional dynamic Markov models.