Approximation of Nonlinear Dynamics of the Twin Rotor MIMO System Using Neural Network

Abstract

This paper presents a hybrid residual-learning framework for approximation of Twin Rotor MIMO System (TRMS) dynamics using a neural-network-based nonlinear correction model combined with a nominal linear state-space representation. The proposed approach predicts the next system state by augmenting the analytical linear model with a learned residual correction term. Unlike conventional methods, the residual model incorporates derivative-based dynamic features computed from consecutive measurements, allowing improved representation of transient system behavior and inertial effects. Since pitch and yaw dynamics are strongly coupled, a joint neural approximator is employed for simultaneous modeling of both channels. Experimental trajectories collected from the TRMS platform were used for dataset construction, training, and evaluation. The proposed framework was evaluated using onestep-ahead prediction accuracy metrics and compared with the nominal linear model. Experimental results demonstrated that the residual neural model significantly improves approximation accuracy and more effectively reproduces nonlinear transient dynamics. Additional experiments involving residual clipping showed that aggressive clipping degrades approximation quality because nonlinear correction terms are suppressed. The obtained results confirm the effectiveness of derivative-based residual learning for coupled nonlinear dynamical systems and demonstrate the applicability of the proposed approach for realtime system approximation tasks.