Segmental Semi-Markov Model Based Research In Online Series Pattern Detection Method

In recent years, online series pattern detection technique has attracted increasing interest in time series data mining communities, which is necessary for real-time supervision and control of time series, and is of unique significance in many applications, such as to help people find out and fix up the problems, analyze the abnormal condition, make technical analysis of investment in securities business, analyze human motion. Up to now, numerous approaches have been proposed for online series pattern detection, and they can be classified into two categories: 1) methods based on the sliding window and linear scan, such as Euclidean and DTW distance algorithms;2) methods based on probability model, such as segmental semi-Markov model. Althouth the former ones are simple and easy to realize, their computational loads are very expensive. By contrast, the latter ones are complex but shows better performance. However, the existing segmental semi-Markov model can only detect the matching sequences with approximately equal length to that of the query pattern, i.e., without time scaling. Online detection of similar patterns under arbitrary time scaling of a given time sequence is still a challenging problem.In this paper, we propose an improved segmental semi-Markov model which modifies the existing model from the following two aspects: 1) Introducing the offset distribution to replace the observation distributions for every segment;2) Introducing the amplitude (Y coordinate) difference distribution for every segment. At the same time we give the effective algorithms to model the query pattern and estimate the parameters of the model. Experiments on two synthetical datasets and two real datasets demonstrate the good performance of the model.