Abstract
A new method of time series segmentation is developed using differential evolution. Traditional methods of time series segmentation focus on single variable segmentation and as such often determine sections of the time series with constant slope (i.e. linear). The problem of segmenting multivariate time series is significantly more involved since several time series have to be jointly segmented. Thus the concept of boundary becomes ill-defined since each time series may not be exactly synchronized and change identically in time. The problem is rectified by minimizing the mean of the variance of the slopes determined in each segment. Performance of the method is measured in terms of the classification rate and the accuracy of determination of boundaries. Experimental evidence shows the effectiveness of the method when applied to synthetic and real-world data compared with multivariate time series clustering approaches.