Window Size Dependency on Anomaly Detection Based on Discord for Time Series
Makoto Imamura, Takaaki Nakamura · IEEJ Transactions on Electronics Information and Systems · 2015
A large facility such as plants and buildings have many and various sensors so that “training data less” and “parameter tuning free” are required for anomaly detection. Discord method that finds the nearest neighbor subsequence with high speed is a state of the art satisfying the above needs. However, the selection of window size for discord is a remaining issue. We notice that discord behavior in transient state is different from that in steady state and show that the optimal window size depends on the anomaly shape because of a discord-shift in transient state (In a transient state, the nearest neighbor subsequence of an anomaly subsequence shifts from the normal subsequence corresponding to the anomaly subsequence originally.). We also define an evaluation index to compare the detection accuracy for anomalies with different window sizes and specify the condition when a discord-shift occurs in a transient state. Lastly, we consider a window size tuning method on the basis of the above knowledge.