Multi-scale Anomaly Detection with Wavelets
Jack Coughlin, Gian Perrone · 2017
We present a novel system for the detection of anomalies in streaming timeseries data, particularly that found in server and application monitoring, as well as "Internet of Things" (IoT) sensor monitoring and applications. By using the Discrete Wavelet Transform combined with online kernel density estimation, we achieve robustness to many different families of data, and detect multiple anomaly types at different time scales. We compare the performance of this method against other methods using the Numenta Anomaly Benchmark, and achieve a normalized score of 65.58 out of 100, approaching the current best of 70.1.