Data Representation Based on Interval-Sets for Anomaly Detection in Time Series

Anomaly detection in time series is a popular topic focusing on a variety of applications, which achieves a wealth of results. However, there are many cases of missing anomaly and increased false alarm in most of the existing works. Inspired by the concept of interval-sets, this paper proposes an anomaly detection algorithm called a fuzzy interval-set and tries to detect the potential value of the time series from a new perspective. In the proposed algorithm, a time series will be divided into several subsequences. Each subsequence is regarded as an interval-set depending on its value space and boundary of the subsequence. The similarity measurements between interval sets adopt interval operations and point probability distributions of the interval bounds. In addition, an anomaly score is defined based on similarity results. The experimental results on synthetic and real data sets indicate that the proposed algorithm has better discriminative performance than the piecewise aggregate approximation method and reduces the false alarm rate significantly.