| Recently,Non-homogeneous Dynamic Bayesian network was proposed to model the ubiquitous non-stationary time series.It broke through the limit of homogeneous property in Dynamic Bayesian Network,and thus became a research hotpot in the field of Probabilistic Graphic Model.Change-points detection is an indispensable task in the process of modeling non-stationary time series by Non-homogeneous Dynamic Bayesian network.Relevant ex-isting work ignores the prior knowledge hidden in the time series.In order to fully mine the relative information of change-points,we propose an algorithm called APK-RJ-MCMC.The APK-RJ-MCMC algorithm computes the Euclidean distance between the back window and the front window of each time point,and view it as relative in-formation of change-points to estimate the probability of each certain time point to be a change-point.Then we combine above estimation into change-points sampling.We adjust the ratio of the proposed probabilities of the birth,death and shift moves during the sampling to adjust their acceptance probability.So the time point which has higher prior value is more probable to be sampled.We evaluate the proposed APK-RJ-MCMC with both artificial data set and real gene expression data set,and define the occurrence frequency of the real change-point as the evaluation standard.The experiment result shows that the proposed APK-RJ-MCMC performs better than RJ-MCMC algorithm in change-point detection.In the sample result of artificial data set,APK-RJ-MCMC prevails over RJ-MCMC with 16 percent higher true change-point occurrence frequen-cy and 7 percent higher AUC score.In the sample result of real gene data set,these two indicators are improved 21 percent and 24 percent respectively.Besides of exploring the change-points detection technology of Non-homogeneous Dynamic Bayesian networks,this paper also applies Non-homogeneous Dynamic Bayesian network to predict abnormal condition in exhaust gas treatment system.The method p-resented in this paper can be divided into four stages,raw data preprocessing stage,regression model learning stage,output density of next moment predicting stage and abnormal condition warning stage.In the preprocessing stage,this paper tested the s-tationary property of the exhaust gas treatment system history data,and conducted irrelevant variable deletion and down-sampling to the data.In the learning stage,this paper built up the evolved regression model between the output and the varied influ-ence factors based on Non-homogeneous Dynamic Bayesian network,then clustered the samples of Non-Homogeneous Dynamic Bayesian networks and obtained multi-ple regression model to present different mechanism.In the predicting stage,this paper used Naive Bayesian classifier to classify recent conditions into corresponding regres-sion model which had the maximum posterior probability,and used the regression model to predict the exhaust gas output density of the next moment.In the warning stage,this paper set classification rules to judge whether the above prediction is ab-normal.Experiment result shows that in the prediction of the density of exhaust gas the method presented in this paper reduces MAE(Mean Absolute Error)by 7.25 and 106.78 respectively,and improves accuracy of abnormal condition prediction by 0.1 and 0.22 respectively compare to ARIMA and GLM. |
