Research On Outlier Detection Of Multivariate Time Series

The problems of time series widely exist in the social production and life, so the research on analysis of time series has great significance. In this thesis, the purpose is to implement the anomaly detection of multivariate time series. Anomaly detection has provided people with a lot of valuable information in the financial field, hydrology, meteorology, seismology, video surveillance, medical field and other fields. So its study has great significance. Researches on anomaly detection of multivariate time series mainly related to the similarity measuring function for multivariate time series and the anomaly detection algorithm.Aiming at the existing problems of available similarity measurement functions for multivariate time series, an improved weighed distance function was proposed to measure the similarity for multivariate time series in this thesis. The weighted distance function is based on principal component analysis and piecewise linear representation. The measure method contained three steps:multivariate time series analysis based on PCA, similarity measurement between single time-series datasets (using the distance function based on area) and similarity measurement between multivariate time-series datasets. The anomaly detection algorithm in this thesis was still a local outlier detection algorithm based on k-nearest-neighbor. But the aforementioned improved similarity measurement function was used to calculate the distance between multivariate time-series datasets. Because the similarity measurement function took into account more relevant factors, the anomaly detection algorithm based on the function could get a better anomaly detection result than the other ones based on the other similarity measurement functions.At last, based on some datasets, some experiments were especially designed to analyze the performance of the distance function based on area and the improved similarity measurement function for multivariate time series. The experimental result indicated that the two functions were effective and reasonable. Using China Securities 100 Index as the dataset, compared the anomaly detection algorithm based on the improved similarity measurement function with the one based on EROS. The experimental result showed that the detection result of the anomaly detection algorithm in this thesis was more accurate.