Utilizing an adaptive window rolling median methodology for time series anomaly detection
Dimitris Dimoudis, Thanasis Vafeiadis, Alexandros Nizamis, Dimosthenis Ioannidis, Dimitrios Tzovaras
Abstract
With the rise of industry 4.0 and the amount of data gathered from sensors, anomaly detection has become an extremely important task. Due to the variety of anomalies, a universal approach is not yet possible, therefore, many methods to identify abnormal behaviours on data have been researched. In this paper, a new anomaly detection algorithm is proposed that spots abnormal data points in time series using a rolling median with an adaptive sliding window. The window changes based on two methods, F1 based and T-test. F1 method tries to make the F1 score have only an upward trend, while T-test recognizes trends in time series and adjusts the window accordingly. For the evaluation, two well - known benchmark datasets were employed. Moreover, the proposed algorithm was also tested on a dataset consisting of real industrial machinery sensor observations coming from a furniture manufacturer. In the two benchmark datasets mentioned, the two variants are compared with an ensemble of 7 models. The results indicate that the proposed method achieves, in the most cases, better F1 score compared to the benchmarks.