Segmented Representation of Time Series Based on Key Morphological Points

Abstract

Time series data has the characteristics of high dimension and high noise, data mining directly from the original series is not only time-consuming but also inefficient. Therefore, it is necessary to preprocess the original time series. Among various compression methods, PLR based on feature points has been widely studied because of its simplicity, high efficiency, and easy interpretation. However, most of the feature-based linear representation methods have the problem of low accuracy of the selected segmentation points. Based on the idea of traditional important extreme point extraction method, this paper proposes a time series segment representation algorithm based on key morphological points. The algorithm extracts the key extremum and key stationary value, which can greatly reduce the noise and retain the key points. Compared with IP, SEEP, KP and IEP algorithms, the proposed algorithm is tested in 9 different fields of UCR. The results show that under 98% compression rate, the algorithm has the least fitting error on seven groups of data, and second only to the best case in one group of data, so the algorithm can better fit the original sequence at high compression rate; In addition, the DTW distance calculation results on 9 types of data show that the algorithm can correctly classify 7 types of data, it has the best classification accuracy. The comparison results of difference value of the similarity of different classes show that it has better robustness and can be easily applied to the subsequent application research of the sequence.