Classification of streaming time series based on neural network technologies and behavioral patterns

Data mining of streaming time series is a topical task that occurs in a wide range of subject areas. This article presents the SALTO method (Snippet and Autoencoder based Labeling of Time series coming Online), which combines a neural network model for classifying streaming time series and an analytical method for automatic labeling of training set based on behavioral patterns (snippets). The lightweight architecture of the neural network model used makes it possible to achieve low latency when classifying streaming data. The method involves two stages: preprocessing and classification. At the preprocessing stage, the Preprocessor searches for the optimal value of the length of the subsequence of the input series and performs its labeling. The resulting labels are used to create a training set for the Extractor. The Extractor performs the extraction of the snippet of the input subsequence and assigns it a class based on the similarity of the extracted snippet with snippets from the training set. The experimental results showed that the proposed method performs the classification of a single subsequence in less than 10 ms, which allows SALTO, unlike other competitors, to be used in industrial Internet of Things applications with high performance requirements in accordance with the URLLC (Ultra-Reliable Low Latency Communications) standard.