Multivariate Sliding-Mode Singular Spectrum Analysis for the Decomposition of Multisensor Time Series

Abstract

In this letter, the multivariate automatic singular spectrum analysis (MA-SSA) and multivariate sliding-mode singular spectrum analysis (MSM-SSA) algorithms are proposed as multivariate extensions to automatic singular spectrum analysis and sliding-mode singular spectrum analysis (SM-SSA), respectively, for the decomposition of multisensor time series or multichannel signals. The MA-SSA is evaluated using hierarchical clustering after the diagonal averaging step of multichannel singular spectrum analysis of a multichannel signal. The MSM-SSA uses a sliding window-based analysis and MA-SSA for obtaining the reconstruction components from the multichannel signal. The MSM-SSA is analyzed and tested using both multichannel synthetic and real-world (electroencephalogram) signals, wherein the quality reconstruction factor is used to quantify the reconstruction of the synthetic signal. The accuracy, sensitivity, and specificity are measured in a focal versus nonfocal seizure classification task, thereby, showing the reliability and robustness of the MSM-SSA.