Self-Tuning Varri Method in Preparing Fatigue Segment

Abstract

An overlapping segmentation method on time series data is often used for preparing training dataset i.e. the population of instance, for classification data mining. Having large number of redundant instances would burden the training process with heavy computational operation. This would happen if practitioners fail to acknowledge an appropriate amount of overlap when performing the time series segmentation. Fortunately, the risk could be decreased if knowledge preferences can be determined to guide on overlapping criteria in the segmentation algorithm. Thus, this study aims to investigate how the Varri method is able to contribute for better understanding in preparing training dataset consists of irredundant fatigue segment from the loading history (fatigue signal). Generally, the method locates segment boundaries based on local maxima in the difference function which are above the assigned threshold. In the present study, the mean and standard deviation have been used to define the function due to the fact that predicting attributes are the key components in defining instance redundancy. The resulting dataset from the proposed method is trained by three classification algorithms under the supervision of the Genetic algorithms-based feature selection wrapper approach. The average performance index shows an additional advantage of the proposed method as compared to the conventional procedure in preparing training dataset.