Latent State Space Modeling of High-Dimensional Time Series with a Canonical Correlation Objective

High-dimensional time series are commonly encountered in modern control systems, especially in autonomous systems. In this work, a novel parsimonious latent state space (LaSS) model is proposed to characterize the latent dynamics, achieving general latent dynamic modeling with dimension reduction. The LaSS model is optimized by alternating estimations of the dimension reduction projection and the latent state space model. Precisely, the latent state dynamics are estimated by stochastic subspace identification methods. Furthermore, the canonical correlation analysis (CCA) objective is employed to acquire the optimal predictability for the extracted latent variables. The proposed LaSS-CCA algorithm is tested on a real industrial case for its effectiveness.