A universal event-group based time series data mining framework for multi-variate time series classification

The Time Series Data Mining (TSDM) is an active research field due to the massive demands from industrial and other real-world practices. This thesis develops a novel universal TSDM methodology, named the Event Group Based Classifier (EGBC) framework, for multi-variate time series classification. This work was initially motivated by demands from the iron-making industry, and later extended as a generic TSDM technique. The unique feature of the proposed EGBC framework is its three-layer structure, of which each layer works independently and focuses on solving different problems within the TSDM domain. The EGBC framework has been examined and evaluated with different tasks, and the outcomes of which suggested this method has a satisfactory performance and adaptability for sequential data classification under various scenarios. Specifically, in the most complex iron-making industrial environment, the EGBC framework produced significantly better accuracies for predicting abnormal industrial behaviours compared to that achieved by prior existing methods.