Multivariate Time Series Classification Technology Research

Multivariate time series is a kind of data that have the ability of space and time description at the same time. It is widely used in the description and expression of the dynamic behavior characteristics of things, and therefore is a kind of very important data. The classification of multivariate time series for the identification of dynamic behavior and process is a valuable research work, and has got a lot of attention and research.In order to classify the multivariate time series effectively, comprehensive research for related technologies is done in this paper.These related techniques include:feature extraction for multivariate time series, feature selection and classification techniques.Based on these techniques, classification for multivariate time series can be implemented completely.(1) In order to construct a classification model, it is necessary to get the expression based on features firstly. So the feature extraction technology for multivariate time series is studied in this paper at first.And then, on the basis of analyzing the disadvantage of existing methods, this paper proposes a feature extraction method based on wavelet packet transform.This method can extract features on different time and frequency scale, and obtain as far as possible full and rich feature description, and avoid the information loss in the process of feature extraction.At the same time, it isn’t needed to set any parameters in the process of extraction, and the difficulty causing by the different length of various series is simple to handle. Due to having obtained the complete time and frequency characteristics of multivariate time series, it becomes easier to distinguish different kinds of series.(2) Feature extraction based on wavelet packet transform can get a lot of features from wavelet packet coefficient, which includes a large number of irrelevant and redundant features. If these features are directly used to construct feature vectors for classification, it not only can make the dimension of feature vector have very high, but also will have a negative impact eventually on the performance of classification model classification.So this paper studies the feature selection technology, by which the features effective for classification will be reserved and the irrelevant and redundant features will be excluded from the initial features set. In the process of research,the feature selection algorithm ReliefF is analyzed firstly. Considering its shortcomings, an improved algorithm ReliefF-SFB-SVM is proposed, which obtain better performance than ReliefF. Finally, we compare ReliefF-SFB-SVM with ReliefF and mRMR on actual performance through experiments on real data sets. The experimental results show that ReliefF-SFB-SVM can get higher classification accuracy by using less features than ReliefF, and has similar classification performance but higher running speed compared to mRMR.(3) Based on feature vector representation from feature extraction and feature selection technology for multivariate time series, the classification model for multivariate time series can be constructed. As a result, this paper studies the classification algorithm. In the research process, considering the SVM classification model has strong generalization ability and can solve the nonlinear separable problem, a new tree-structured Multi-category SVM classification model TSM-SVM is proposed, to make use of the advantages of SVM and obtain multi-classification ability at the same time. Finally, we compare TSM-SVM with PNN and OVA-SVM on actual performance through experiments. Experimental results show that TSM-SVM has faster classification speed than OVA-SVM and higher classification accuracy than PNN, and gains ideal training speed at the same time.