Dimension Reduction using Self Organizing Map and Its Application to Time Series Anomaly Detection

Abstract

In this paper, we propose an anomaly detection method based on Self Organizing Maps(SOM). In our method, SOM is used to extract the features of the time series as 2-dimentional data using time series convolution SOM, and pareto learning SOM is used to detect the anomaly as outliner. To evaluate the performance of this method, TSB-UAD which is the benchmark suits for time series anomaly detection is used, and our method outperforms other methods included in TSB-UAD and combination of Isolation forest and classical feature extraction method.