Uncompromised Accuracy: Fast and Reliable Multivariate Anomaly Detection for Satellite Signals

Abstract

In the realm of multivariate anomaly detection (AD), deep neural networks (DNNs) have garnered attention. However, relying solely on a single DNN model may not achieve the optimal balance between accuracy and time efficiency. Nonlinear variants of Kalman filter models (extended kalman filter (EKF), unscented kalman filter (UKF)) are known for their efficient time complexity but often compromise accuracy. On the other hand, deep learning-based models like Transformers and recurrent NNsexcel in accuracy but introduce complexity challenges. This article introduces the selective points AD method, which strategically merges accurate and time-efficient algorithms by leveraging a selection of multiple models. The optimal model fusion that maximizes the accuracy-to-time ratio (ATR) is determined by assessing the estimated covariance from both sets of algorithms. The results demonstrate a superior ATR by at least 30% and 33% compared to the best existing method for soil moisture active passive and Mars science laboratory rover datasets, respectively.