Deep learning method for anomaly detection in streaming multivariate time series

The article touches upon the problem of detecting anomalous subsequences of multivariate streaming time series, where the elements arrive in real time, which currently arises in a wide range of subject domains: industrial Internet of Things, personal healthcare, etc. In the article, we introduce a novel method to solve such a problem, called mDiSSiD (Discord, Snippet, and Siamese Neural Network-based Detector of multivariate anomalies). The mDiSSiD method employs the time series discord concept (a subsequence with the most dissimilar nearest neighbor), which is generalized to the multivariate case. Multivariate discord refers to the $N$-dimensional subsequence of a $d$-dimensional time series (where $1\leqslant N\leqslant d$), which is the most dissimilar to all other subsequences of $N$-dimensional time series obtained by composing all the possible combinations of $d$ series of $N$. Anomaly detection is implemented through a deep learning model based on the Siamese neural network architecture. Experimental evaluation of mDiSSiD over real time series from various subject domains showed that the proposed method is on average ahead of state-of-the-art analogs based on other deep learning approaches (convolutional and recurrent neural networks, autoencoders, and generative-adversarial networks) in terms of anomaly detection accuracy.