Data-Driven Hybrid Framework for Dynamic System Control

Abstract

Nonlinear electrical and electromechanical systems pose significant challenges for observer-based control design. Conventional observer approaches require accurate mathematical models, which often fail under physical irregularities such as sensor noise, measurement delays, and parameter variations. These changes decrease estimation accuracy and degrade control action. This paper proposes a Hybrid ARX– Observer framework to overcome these limitations. It combines the stability of a traditional model-based observer with the flexibility of a data-driven ARX estimator. To guarantee input-to-state stability (ISS), observer gains are computed using a matrix-multiplier technique based on linear matrix inequalities (LMI). The ARX component employs Recursive Least Squares (RLS) adaptation to attenuate measurement noise and capture residual nonlinearities. Both estimates are then fused together using a fusion gain α. This framework is tested on a robotic arm and the results show that the hybrid framework improved the robustness and reduced estimation error up to 4% compared to the conventional observer based estimation, while remaining computationally light compared to fully data-driven alternatives.