Anomaly detection in time-series data environment
Doyeon Kim, Taejin Lee · 2020
Typical label data detect anomaly due to the relationship between inputs and labels, but time-series data are more demanding in detecting anomalies because they detect anomaly based on time-varying values. To solve this problem, this paper proposed Stacked-Autoencoder based data detection technique with ICS dataset among time series data. The Loss value was calculated as CDF and determined to be a suspicious event if it was greater than the arbitrarily specified threshold value. The experiment was carried out by designating 0.5, 0.7, 0.9 and 0.98, and 0.98 showed the best result with an accuracy of about 96%.