Usage
  • 140 views
  • 316 downloads

Time Series Mining: A Computational Intelligence Approach

  • Author / Creator
    Li, Jinbo
  • Time series has become prevalent in a broad range of real-world applications such as weather, health care, agricultural production, satellite image analysis, speech recognition, industrial process control, and others. This type of data comes as a collection of observations obtained chronologically, describing different aspects of a specific phenomenon. With the increasing availability of time series data, the discovery and extraction of available information (e.g., similar patterns, meaning rules) from them are essential to human. In this dissertation, three main time series mining tasks involving (i) anomaly detection, (ii) approximation /representation and (iii) predictive modeling will be concerned with developing Computational Intelligence (CI) related techniques on the one hand and with their application on complex real-world problems on the other hand. The primary objectives of this thesis are to develop a series of relatively comprehensive frameworks for these mining tasks. Anomaly detection in the multivariate time series refers to the discovery of any abnormal behavior within the data encountered in a specific time interval. Here we develop and carry out two unsupervised and supervised frameworks of multivariate time series anomaly detection for amplitude and shape anomalies, namely cluster-centric anomaly detection models and Hidden Markov Models based model with the aid of the transformation of multivariate time series to univariate time series respectively. In the first unsupervised model, the modified Fuzzy C-Means clustering was used to capture the structure of multivariate time series. A reconstruction error serves as the fitness function of the PSO algorithm and also has been considered as the level of anomaly detected in each subsequence. In the other model, several transformation techniques involving Fuzzy C-Means (FCM) clustering and fuzzy integral are studied. A Hidden Markov Model (HMM), one of the commonly encountered statistical methods, is engaged to detect anomalies in multivariate time series. Before implementing most tasks of time series data mining, one of the essential problems is to approximate or represent the time series data because of its massive data size and high dimensionality. We establish FCM clustering based approximation methods. We carry out a comprehensive analysis of relationships between reconstruction error and classification performance when dealing with various representation (approximation) mechanisms of time series. Furthermore, we also elaborate on a novel Hidden Markov Model (HMM)-based fuzzy model for time series prediction. Here fuzzy rules (rule-based models) are employed to describe and quantify the relationship between the input and output time series while the HMM is regarded as a vehicle for capturing the temporal behavior or changes of the multivariate time series. A suite of experimental studies along with some comparative analysis is reported on both synthetic and real-world time series data sets.

  • Subjects / Keywords
  • Graduation date
    Fall 2018
  • Type of Item
    Thesis
  • Degree
    Doctor of Philosophy
  • DOI
    https://doi.org/10.7939/R33R0Q890
  • License
    Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.