Time Series Research Of Anomaly Detection

At present most of methods of time series anomaly detections are for specific type of dynamic data,but the data type are often misjudged t because of the existence of outliers, so it is more reasonable to do anomaly detections before judging the type of the data, many methods of time series anomaly detection are complex and not easy to achieve, such as the method which is described in this article based on Bayesian methods, the theory is very complicated, it need sample and need to repeat many times. the method in literature [40] is not dependent on the model and is relatively simple and easy to implement, but the method can not detected a piece of outliers, and it is easy to make normal points as outliers.in order to solve the problems in the literature[40], one indicator variable and unusual punishment are introduced in this paper and two exception types are defined. the effectiveness of the method given in this article is proved by using the data from Shanghai Stock Market between2004to2009. The main work of this paper:(1) the background and significance of the research about anomaly detection are introduced and then analyze the advantages and disadvantages of some methods.(2) the outlier model which is for time series and a time series anomaly detection methods based on Bayesian are researched.(3) Base on the research about time series anomaly detection above, a time series anomaly detection model which is improved from the literature is given to solve the problem that the model in the literature can not find a piece of outliers and some normal data are considered to be abnormal. two abnormal forms are defined in this paper, one is called the high anomaly, another is called the low abnormal.moreover an indicator variables and the penalty function will be introduced and defined in the paper. finally the validity of the model will be proved through numerical experiments.