Research On Feature Representation And Clustering Method For Time Series

Time Series is a common form of data, and it is widely used in various fields. Therefore, it is an important research direction of data mining to dig Information and knowledge of time series. Traditional data mining methods are usually based on static data, however, time series is dynamic data with a large amount of data. That led to traditional data mining methods can not effectively mine the information of time series. It is of great importance to process and mine the time series data better.The potential model mining of time series is to find the intrinsic relation of time series with future knowledge. Due to tapping the potential model of large amount of time series data, first, the dimension reduction should be carried out by time series feature representation. Then the closely linked time series are aggregated by unsupervised clustering algorithm. The clustering results could be considered as the potential model of time series.This paper aimed to solve the problem of potential model separation for time series. And the feasibility of potential model separation by means of feature representation and clustering method is analyzed. The appropriate time series feature representation and clustering method for numerical time series is explored.The existing feature representation of time series methods can reduces the dimension of the time series, but the features of the original time series can not be preserved well and the clustering result will be affected. Thus, Combining Discrete Fourier Transform with subsection theory, the Piecewise Discrete Fourier Transform is proposed. The dimension of time series is reduced effectively with this method, and the features of the original sequence are preserved well at the same time.In the absence of prior knowledge of data, cluster analysis can solve the problem of classification of a large number of data and tap its intrinsic relation. Therefore, clustering method can be used to mine the potential model of data. DENCLUE algorithm is a kind of excellent algorithm which can solve random shape of variable density problem, and it is universalizing. However, DENCLUE method has these problems: large number of parameters that need to set, constrained interrelation of parameters led to the difficulty of parameter setting, computational complexities etc. Due to the existence of these problems, the efficiency and performance of using DENCLUE method to find time series pattern separation are not good. In order to improve this problem, cluster point estimation based on cluster evaluation is introduced to propose a novel method based on avoid determining noise threshold in DENCLUE. This improved algorithm which does not use the noise threshold in DENCLUE and reduce computational complexity. Meanwhile, the noise in data sets can also be verified well in our method, and the obtained clusters are more compact, which can be viewed as an improvement for applicability and performance of DENCLUE.In this paper, Piecewise Discrete Fourier Transform and a method based on avoid determining noise threshold in DENCLUE are applied to mine the potential models of time series of pipe temperature of launch field. Combination of dynamic time warping distance and silhouette coefficient is applied to verify whether the models are satisfied with the time sequence of the same model has a higher degree of similarity and different model's similarity is lower. By this standard, whether the potential model is in line with the actual situation can be evaluated.