Real-time change detection in time series based on growing feature quantization

An unsupervised time series change detection method based on an extension of Vector Quantization (VQ) clustering is proposed. The method clusters the subsequences extracted with a sliding window in feature space. Changes can be defined as transition of subsequences from one cluster to another. The method can be used to achieve real time detection of the change points in a time series. Using data on road casualties in Great Britain, historical data on Nile river discharges, MODerate-resolution Imaging Spectroradiometer (MODIS) Normalized Difference Vegetation Index data and x simulated data. It is shown that the detected changes coincide with identifiable political, historical, environmental or simulated events that might have caused these changes. Further, the online method has the potential for revealing the insights into the nature of the changes and the transition behaviours of the system.