A Kernel Design Approach to Improve Kernel Subspace Identification

Subspace identification methods, such as canonical variate analysis (CVA), are noniterative tools suitable for the state-space modeling of multi-input, multi-output processes, e.g., industrial using input-output data. To learn nonlinear system behavior, kernel subspace techniques commonly used. However, issue design must be given more attention because type can influence kind nonlinearities that model capture. In this article, a new is proposed CVA-based identification, which mixture global and local to enhance generalization ability includes mechanism vary each process variable into response. During validation, hyper-parameters were tuned random search. The overall method called feature-relevant mixed CVA (FR-MKCVA). Using an evaporator case study, trained FR-MKCVA models show better fit observed data than those single-kernel CVA, linear neural net under both interpolation extrapolation scenarios. This work provides basis future exploration deep diverse designs identification.