Time Series Segmentation Based On HMM

Time series is a sequence of data which changes with time order.In contrast to traditional static data,time series is a kind of complicated data object which describes the process that things change.Time series segmentation is one of the important tasks of the time series analysis research.It’s a process that divides a long sequence into non-overlapping and ordered subsequences.Also,it’s the basic step of the time series analysis which has wide application space and the important research value.Hidden Markov Model(HMM)is a kind of important time series model.Most of the HMM based time series segmentation approaches fail to consider the field background in practice,so they can’t meet the user’s expectations.In some applications,a minimum period is required before a state change is said to have occurred.Motivated by such applications,a constrained Hidden Markov Model,which controls the transition between underlying states,is proposed in this study.Moreover,a constrained Baum-Welch algorithm and a constrained Viterbi algorithm are also given.The simulation experiment shows that the constrained HMM based method has better segmentation effect than standard HMM based method.The constrained HMM based time series segmentation is applied to the Hamilton’GNP dataset and the Shaoguan’s precipitation data from 1951 to 2013 respectively.The segmentation results of the algorithm proposed in this study more accord with the practical significance than the results of the HMM based approach.This paper also presents a constrained HMM based approach to identify the obscurely grouped pairs.The constrained HMM based approach is added into a general K-Means based consensus clustering framework to evaluate Consensus matrix(CM)and estimate the cluster number.The approach analyses an occurrence sequence gotten by counting the occurrence of each possible element value in an original CM.The sequence is regarded as an observable sequence generated by a Markov chain of three states,which are separately clustered,ambiguously clustered,and jointly clustered respectively.Experimental results on four UCI datasets suggest this approach is more effective than some recent approaches on estimating cluster number.