Network Model Based On Mutual Information Time Series Database Knowledge Discovery

This thesis suggests a novel, complete approach to knowledge discovery (KD) in Time Series Database (TSDB). The approach based upon Mutual Information of Shannon's Information Theory, probabilistic and statistical methods, filtering with Voronoi graph, and fuzzy logic. The general KD stages in TSDM include: construction of TSDB, cleaning and filtering of time series data, feature extraction, construction and rule deduction of Mutual-Information Network Model (MINM). and fuzzy logic for rule confliction and reduction. The results of KD in TSDB are rules, which can be used to predict time series behavior in the future.First, the thesis points out the difference between TSDB and general database and definition of TSDB, explaining the objectives of TSDB for KD: trend analysis, similarity searching and series model mining. On discussing and analyzing different construction of Special Database, which includes TSDB, we introduce how to remodel TSDB from general database.Then, the paper detailed the construction of MINM, which is used to discover knowledge. After concept of Shannon's Information Theory and Mutual Information being introduced, using difference between priori entropy and post entropy to detect the associating degree between input attributes and target attributes, which controls the network configuration. The network, which consists of several input layers and a target layer, can be used to deduct associating rules between input attributes and target attributes.Upon the contents discussed above, the thesis specifies several steps of using MINM to discover knowledge in TSDB. 1) Analyzing the Exponential Smoothing technology and its shortcoming when used to filtering, the paper proposes a new time series filtering method, which can overcome the shortcoming on base of its adaptive virtues. 2) Trend Segment and Signal to Noise Ratio (SNR) are important features in time series. We show the methods of getting them from filtered time series. 3) Detailed steps of knowledge discovery in time series. 4) In the post-processing step.1 Fuzzy concepts are used to resolve confliction between rules deducted from MIN.Also, we show that the number of rules can easily be reduced by introducing Fuzzy concept.We designed an experiment with the help of Zhejiang Telecom time series database to validate our method. Its result shows the effectiveness of MINM for knowledge in TSDB. At the end of this thesis, we pointed out the important of KD in TSDB in various areas and direction in the future.