关键词: 时间序列, 异常检测, 核转折点, 降维表示

Abstract:

Clipped representation is one of the important data-driven representation approaches in the fields of time series dimensionality reduction and representation, it transforms the raw time series into a set of sequences composed of 0 and 1. However, the traditional clipped representation ignores the influence of data points on the local trend of time series, and the compression ratio is determined by the data itself, which means the user has no choice to make. In order to address the above shortcomings, an improved clipped representation based on kernel turning points, called KTPC, is proposed in this paper. Based on KTPC, a novel time series anomaly detection approach is proposed, called KTPC-AD. The proposed method first detects the kernel turning points in each time series and represents the raw sequences by KTPC, and then, the anomaly scores of each sequence are calculated by KTPC-AD. Finally, the anomaly sequences are detected. The experimental results demonstrate that KTPC achieves higher processing efficiency, and KTPC-AD not only reduces the time complexity of anomaly detection, but also effectively improves the accuracy of detecting anomalies in time series.

Key words: time series, anomaly detection, kernel turning point, dimensionality reduction representation