The Research On Linear Representation For Time Series

With the development of social economy, computer and information technology and the storage technology, a large amount of important data are gradually stored. It is one of the popular problems in data mining how to extract the valuable knowledge what they want in database.Time series which is arranged according to the time sequence reflect the order of the attribute value in time or space. The classification、clustering、anomaly detection、the modeling and similarity inquires of time series can realize the extraction of the desired knowledge. Nowadays, the time series data mining is an important research direction in data mining.Because of the time series data with the properties of massive volume, noise and short-term volatility frequently, it is difficult to obtain a satisfactory result to directly do the above mentioned works on raw data set. It is necessary to reduce the dimension of time series. Various methods have been put forward such as linear fitting method and linear division method, aiming at grasping the local feature and keeping the main features of the sequence. Doing so, the efficiency of data mining is improved and the computation is significantly reduced.Based on the work related to time series linear represent, this dissertation carries out the following research.A new linear fitting method for time series is proposed. The turning points are picked up in terms of the slope changes of the adjacent point connections, and the set of key points of the time series is obtained by merging these points with extreme points. The method can not only eliminate the noise points, but also more precisely locate the key points of the series. Test results show the method can much better reflect the original time series. Comparing this method with the existed ones, the smaller fitting errors are achieved.The method for time series piecewise linear representation based on the function is introduced. Considering the actual situation of the different influence on the different segments in terms of time property of time series and the dynamic growth data of time series, a new method FPAA (Function Piecewise Aggregate Approximation) of piecewise linear representation was proposed based on the method of RPAA(Reversed Piecewise Aggregate Approximation) and PAA(Piecewise Aggregate Approximation). The proposed method overcomes the disadvantages of RPAA and PAA by defining the influence factor of function. FPAA has the linear complexity, satisfies lower bounding lemma and supports online segmentation of time series. Compared with the methods of PAA and RPAA, the FPAA method can effectively query time series online.