Text

theses

Much of the world’s supply of data is in the form of time series. In the last decade, there has been an explosion of interest in time series data mining. Time series prediction has been widely used in engineering, economy, industrial manufacturing, finance, manage- ment and many other fields. Many new algorithms have been developed to classify, cluster, segment, index, discover rules, and detect anomalies/novelties in time series. However, traditional time series analysis methods are limited by the requirement of stationarity of the time series and normality and independence of the residuals. Be- cause they attempt to characterize and predict all time series observations, traditional time series analysis methods are unable to identify complex (nonperiodic, nonlinear, irregular, and chaotic) characteristics. As a result, the prediction of multivariate noisy time series (such as physiological signals) is still very challenging due to high noise, non-stationarity, and non-linearity. The objective of this research is to develop new reliable frameworks for analyzing multivariate noisy time series, and to apply the framework to online monitor noisy time series and predict critical events online. In particular, this research made an extensive study on one important form of multivariate time series: electrocorticogram (EEG) data, based on which two new online monitoring and prediction frameworks for multivariate time series were introduced and evaluated. The new online monitoring and prediction frameworks overcome the limitations of traditional time series analysis techniques, and adapt and innovate data mining concepts to analyzing multivariate time series data. The proposed approaches can be general frameworks to create a set of methods that reveal hidden temporal patterns that are characteristic and predictive of time series events. In second part of this dissertation provide an overview of the state-of-the-art pre- diction approaches. In the third part of this dissertation, we perform an extensive data mining study on multivariate EEG data, which indicates that EEG may be predictable for some events. In chapter 4, a reinforcement learning-based online monitoring and prediction framework is introduced and applied to solve the challenging seizure pre- diction problem from multivariate EEG data. In chapter 5, it first overview of the most popular representation methods for time series data, and then introduce two new robust algorithms for offline and online segmentation of a time series, respectively. Chapter 6 proposes a general online monitoring and prediction framework, which com- bines temporal feature extraction, feature selection, online pattern identification, and adaptive learning theory to achieve online prediction of complex time series events. Two prediction-rule construction schemes are proposed. In chapter 7, the proposed framework is applied to solve two challenging problems including seizure prediction and ’anxiety’ prediction in a simulated driving environment. The significant prediction results demonstrated the superior prediction capability of the proposed framework to predict complex target events from online streams of nonstationary and chaotic time series.

ETD_4358

Ph.D.

Includes bibliographical references

by Shouyi Wang

http://hdl.rutgers.edu/1782.1/rucore10001600001.ETD.000067015

doi:10.7282/T3348J59

ETD doctoral

The author owns the copyright to this work.

2704384

windows xp

ETD

application/pdf

application/x-tar

2713600

b59a84e33638dfdbf7fb45c5561435930a4ca819