A NARX-AMB Hybrid Approach for Reduced-Order Modeling of a MIMO Heating-Pressing Process

This paper introduces a novel hybrid approach that combines the regression neural network and the Associative-Memory-Based (AMB) methods for developing a reduced-order model, with physically interpretable parameters, of a MIMO heating-pressing process within a fishmeal plant. Evaluating multiple AMB structures remains computationally difficult due to the unknown system delay, so this structural exploration is conducted within the framework of the NARX (Nonlinear AutoRegressive with eXogenous inputs) model, which, despite having a significantly higher number of parameters, exhibits ease and speed of training compared to AMB modeling. To address this, we propose a three-step approach: (1) Decomposition of MIMO into SISO Subsystems to focus on individual components. (2) NARX Modeling for SISO Subsystems where we explore different neural network configurations, including the number of hidden neurons and the lag time, using the Bayesian information criteria. (3) Integration of AMB Method using the delay time obtained in step 2. To computationally validate this approach, we utilize real data of a heating-pressing process. This model confirms a clear dependence of the model's parameters on specific regressors.