Study Of Complex Pattern Mining In Time Series Data Stream

The rapid growth of information and communication technology makes all areas of human society be descripted as different data models by digital technology. Data scales have explosive growth. The velocity of generation, collection and dissemination of data with different types and structures can achieve real-time level. It’s an eager problem in the age of big data to solve complex problems, by how to effectively store and analyze these time series data stream with characteristics of dynamic, nonlinear, high dimension, complexity and redundancy, and how to explore the evolution of these rules in time series data stream, and how to acquire the knowledge of complex time series data.This dissertation is around the hot and difficult problems in time series data stream mining, and analyses the related models of time series data stream in different resolutions and granularities. According to the characteristics of time series data stream, we design an online wavelet transform method, a trend symbolic representation method, and multi-granularity time varying fractal dimension calculation method. Based on these methods, this dissertation researches segmentation and frequent models mining of time series data stream in multi-resolution. Moreover, considering the related characteristics of financial time series data stream, we research the complex clustering patterns mining technology by utilizing multi-resolution and multi-granularity method.The primary works of this dissertation include:1. An online discrete wavelet decomposition technology is designed for decomposing time series data stream into multi-resolution form, which can solve the problem of boundary extension, and eliminate pattern twist at ends of data sequence.2. A multi-resolution segmentation method of time series data stream is studied, which can segment time series data stream at different resolution at the same time. This method also can construct and update the hierarchical structures of segmentations at different resolution in pace with the changes of time series data stream.3. A trend symbolic representation technology of time series data stream is researched, which can characterize the trend information effectively, and represent the patterns in time series data stream intuitively.4. The related concepts and analysis technology of multi-resolution frequent trend patterns are summarized, by which we can discovery some patterns in time series data stream which have the same combination of trend symbolic, and a similar length ratio corresponding to related trend symbolic.5. A method of computing multi-granularity time varying fractal dimension is studied, based on boundary merging extension, which can help to mine distribution changing of time series data stream efficiently, to find implicit knowlodge and rules in data more comprehensively, and to study evolution rules in different perspective.6. A method of mining the complex clustering patterns of financial time series data stream is studied, utilizing multi-resolution and multi-granularity technology, which can discovery and analyze the evolution patterns of time series data stream at different resolution and granularity.