Trend extraction based on variable time window length median filter

Time series trend extraction is of great interest in data mining research recently. For example, it can be applied in system monitoring to detect faults and abnormalities in process industry and medical care. In this paper, a trend extraction algorithm based on variable time window length median filter is presented. The trend filtered can not only reflect the basic geometry of noised signal, but also preserve minor variations in order to facilitate following feature extraction techniques to detect abnormal phenomena. The Variable Time Window Length Median Filter Segmentation (VTWLMFS) algorithm proposed in the paper is firstly formulated in terms of basic concepts and implementation steps, then competed with latest piecewise linearization segmentation algorithm. The main advantage of VTWLMFS is its on-line monitoring ability, promptly capturing signal time-varying features. Simulations also show VTWLMFS better recovers the useful signal under typical criteria. An application is field-pipelines small leakage detection, through using VTWLMFS algorithm, online trend of key variable become a useful tool to locate the leakage point along pipelines.