Research On The Similarity-Based Representation And Pattern Search Of Time Series

For the fully use of history data and trying to find the knowledge hidden in the history time series data from the large databases, this thesis devoted in the research on the similarity-based representation and pattern search of time series in details.Base on the all kinds of conception of similarity, this thesis presents a uniform conception of time series similarity. In the new conception, the similarity is defined as the conception depends on the similarity function and the transfer function. The conception of similarity is subjective and the similarity is a concept defined in some special constraints. The problem of similarity is boiled down to the problem to search the right transfer function and similarity function.The key method to solve the problem of similarity search lies in seeking the robust representation algorithms of time series with low compute complexity. This thesis implemented the orthogonal transform such as DFT, DWT; the feature extraction algorithm such as PAA and Landmark Model. The implement emphasizes the quality of speed, the effect of dimensionality reduction, the accuracy, the consistency and the dynamic attribute of the algorithms. This thesis presents a Top_Down algorithm named representation method of time series based on the local extremum feature extraction and verifies the fast speed of the algorithm. Further more, this thesis verifies the correction of the algorithm and discusses the simple application of similarity-based representation of time series combined with the analysis of two examples of the similarity comparison and classification of time series.In the research of the indexing and query methods of time series database, the algorithm of constructing, searching, deleting, querying of R*-tree indexing was given in this thesis. The thesis presents the mixed staging indexing method, the schematic and algorithms of DFT and the representation method of time series based on the local extremum feature extraction to realize the fast exact similarity search. The thesis also discusses the necessary junctures such as the matching model, algorithms of classification and some possible problems in details.Similarity-based representation and pattern search of time series have vast foreground of application. This thesis completes the probingly application of similarity analysis in machine pro-diagnosis, presents the idea of application independent to the research of fault mechanism and the technology of information collection, including the possibility, difficulty and possible way of system realization.