Study On Time Series Data Mining By Wavelet Transform

Atime series is a data sequence of observations which are ordered in time, whichexists in various fields, such as finance, science observing and engineering, etc. Howto manage and use these time series data efficiently is an important and challengingsubject.At present, main methods of data mining research incldue statistical methods,machine learning, neural network and database methods. In this thesis wavelettransform applied to the time series mining is to be studied, which includes wavelettransform in the time series attribute reduction, time series similarity matching, outlierdetection in time series. By using the multiresolution property of wavelet transform,an approach to perform incremental clustering of time series at various resolutions ispresented, and then, the issues of multi-level time series similarity search techniqueand multi-level discovery of the frequent patterns based on wavelet transformationsare proposed.The main works and contributions of this thesis are:(1) Improve the traditional clustering algorithm using wavelet transformIn order to mitigate the problem associated with the choice of initial centers oftraditional clustering algorithm (such as k-means), W-kMeans algotithm is proposed.Once we compute the Haar decomposition, we perform the k-means clusteringalgorithm, starting at the second level and gradually progress to finer levels. Theintuition behind this algorithm originates from the observation that the general shapeof a time series sequence can often be approximately captured at a lower resolution.(2) Multi-level time series similaritysearch technique based on wavelet transformThe existing methods of multi-level time series similaritymatching consider onlythe slope without taking into account the length of the patterns. To improve theexisting algorithm, based on a new similarity measurement by slope and length (KLmeasure), a more reasonable time series multi-level similar pattern-matchingalgorithm is developed. By using the multiresolution property of wavelet transform, amore effective and reasonable way to solve the problem of uniform time scaling isproposed, which can search long-time pattern matching. (3) Multi-level discovery of the frequent patterns based on wavelet transformTime series have the concept of long-term and short-term, therefore, miningmulti-level time series frequent patterns has important practical significance. Thisthesis proposes the concept of multi-level time series frequent patterns mining for thefirst time, and in accordance with the multiresolution property of wavelet transform,the multi-level frequent patterns mining algorithm is proposed, which first transformthe original sequence with wavelet transformation, and then with the combination ofsegmentation methods based on the important points and Inter-Related SuccessiveTrees SIRSTmethods,frequent patterns of different scales can be discovered.