Variable Length Methods for Detecting Anomaly Patterns in Time Series

There has been much interest in mining anomaly patterns in time series. However, different datasets may have different lengths of anomaly patterns, and usually, the length of anomaly patterns is unknown. This paper uses k-distance of a pattern and median to define anomaly factor, the degree of anomaly, presents- definition- of- anomaly pattern based on it and two algorithms, algorithm 1 and algorithm 2. Algorithm 1 uses quadratic regression to segment time series, and obtains the range of length patterns. Algorithm 2 uses DTW (dynamic time warping) and variable methods to calculate similarity of patterns dynamically, detects anomaly patterns in a given time series automatically. We demonstrate the effectiveness of our detection algorithm for anomaly patterns with both synthetic and ECGs data sets, and the experimental results confirm that our methods can detect anomaly patterns with different lengths.