Research And Application Of Cluster Algorithm In Time Series

There are a lot of time series data. Clustering analysis of time-series data can not only obtain the hidden valuable time-related information, and knowledge acquisition can be achieved. then take activities under the guidance of the knowledge. However, the time series data is very large in real life. So, it is necessary to reduce dimensions before applying the cluster analysis to the time series.The main results of this paper are:(1) A weighted value of similarity measure matrix is proposed. Multiple time series obtains the characteristic matrix and the singular values as a matrix and vector respectively with SVD. After hadamard transform the characteristic matrix and its correlation coefficient matrix eigenvalues as a matrix and weights respectively.(2) Multivariate time series clustering analysis based on EMD and SVD is proposed. After filling the default value of the original data and normalizing. Firstly, we obtain the trend of multivariate time series by using of EMD and smooth the sequence. Secondly, multivariate time series are reduced dimension of by using of SVD. Finally, the improved K-means algorithm is applied to the characteristic matrix and corresponding weight data that represents of original multivariate time series.(3) Multivariate time series cluster analysis based on Hadamard transform is proposed. After filling the default value of the original data and normalizing. Firstly, obtaining the trend of multivariate time series by using of Hadamard transform and reducing dimension by using of wavelet transform. Secondly, calculating the weight of obtained characteristic matrix. Finally, the improved K-means algorithm is applied to cluster the pretreatment data.In this paper, multivariate time series clustering method based on EMD and SVD and Hadamard transform are introduced. Both methods have their advantages:in the first approach, time series is decomposed by using of EMD and the trend of time series is extracted. Due to the original EMD method can filter the noise sequence; the resulting trend series can accurately reflect the original sequence trend so that sequences can become more clear and relatively small loss of information. So the dimensionality reduction clustering on the basis of trend sequence can improve the clustering effect. Furthermore, the SVD decomposition of the trend series can be unified sequences of different lengths to the same scale. Usually, most of the time series are long. After extracting sequence characteristic by SVD, the dimension of sequence characteristic is nothing to do with the length of the sequence, which makes the clustering of unequal length sequences possible; In the second method, sequences dimension reduced by Hadamard transform. Because hadamard transform make data relatively high concentration of energy, which make very short data represent the original sequence and maintain the trend of transformation of original sequence. Cluster analysis on the basis of dimension-reducted sequence can increase the time required for clustering and clustering accuracy significantly. The experimental results show that these two kinds of clustering method has realized multiple time series clustering effectively, and the effective clustering of two parallel clustering algorithms are analyzed and compared.