Anomaly Detection for Non-Stationary and Non-Periodic Univariate Time Series

Abstract

This study proposes an anomaly detection method called wavelet autoencoder anomaly detection (WAAD) for non-stationary and non-periodic univariate time series. The method first applies discrete wavelet transform to time series of a sliding time window to obtain wavelet transform coefficients. It then uses an autoencoder to encode and decode (reconstruct) these coefficients. WAAD calculates the reconstruction error for every time window. An anomaly is assumed to occur for specific conditions of the errors. By five NAB datasets, the performance of WAAD is evaluated and compared with other methods to show its superiority.