An Efficient Anomaly Detection Model Based on Tensor Decomposition and VARIMA for High-Dimensional Multivariate Time Series

Abstract

This article presents an edge-cloud collaboration anomaly detection model for high-dimensional multivariate time series. It is capable of dealing with both point anomaly and pattern anomaly. The transformation of data to tensor is carried out by sliding window with full consideration of the time dimension. The high dimensionality of data is tackled with tensor dimensionality reduction. An efficient iterative tensor decomposition method with low rank approximation is developed to rapidly obtain an optimal core tensor. It retains key information of the original tensor and achieves dimensionality reduction at the same time. A key matrix factorization technique is employed to circumvent large amount of iterative calculation for singular vectors of matrices. For anomaly detection, a tensor-based statistical prediction model is devised to generate a predicted tensor. For the purpose of comparison, a reverse technique is used to transform the predicted tensor to the form of original data. The final anomaly detection is performed with least significant Difference and majority voting. Extensive experiments are conducted with two notable real-world datasets in a specific edge-cloud environment. Our proposal is compared with six other popular methods in terms of performance metrics precision, recall, F1-score, AUC and delay. Experimental results show that our method is superior to the six other methods in both edge-cloud and pure cloud settings.