Automated Anomaly Detection Assisted by Discrimination Model for Time Series

As more and more sensor data have been recorded, automated anomaly detection of high reliability is urgently needed to lessen the increasing monitoring burden and reduce the risk of system faults. To this end, this paper proposes a method named Anomaly Detection Assisted by Discrimination Model (ADADM). In ADADM, the anomaly score for each point in sequences is calculated using an LSTM-based Encoder-Decoder model. In the mean time, a discrimination model based on ALSTM-FCN distinguishes the subsequences containing the Known Type Anomalies (KTAs) from others. Finally, we set two different thresholds for above two types of subsequences respectively and determine the anomalous points according to the anomaly scores and the corresponding threshold. Experiments on four real datasets of periodicity/aperiodicity and univariable/multivariable demonstrate that not only ADADM can improve the overall performance significantly, but also can improve the precision and recall of the anomaly detection simultaneously.