Detection of time series anomalies based on data mining and neural network technologies

The article touches upon the problem of discovering subsequence anomalies in time series, which is currently in demand in a wide range of subject domains. We propose a new semi-supervised method to detect subsequence anomalies in time series. The method is based on the concepts of discord and snippet, which formalize, respectively, the concepts of anomalous and typical time series subsequences. The proposed method includes a neural network model that calculates the anomaly score of the input subsequence and an algorithm to automatically construct the model’s training set. The model is implemented as a Siamese neural network, where we employ a modification of ResNet as a subnet. To train the model, we proposed a modified contrast loss function. The training set is formed as a representative fragment of the time series from which discords, low-fraction snippets with their nearest neighbors, and outliers within each snippet are removed since they are interpreted as abnormal, atypical activity of the subject, and noise, respectively. Computational experiments over time series from various subject domains showed that the proposed model, compared with analogues, has on average the highest accuracy of anomaly detection with respect to the standard VUS-PR metric. The downside of the high accuracy of the method is the longer time spent on model training and anomaly detection compared to analogues. Nevertheless, in applications of intelligent building heating control, the method provides a speed sufficient to detect subsequence anomalies in real time.