Estimation of Muscle States from EMG Signals Using a Nonlinear KKL Observer

Abstract

This thesis presents a data-driven observer for estimating internal muscle states—activation, fiber length, and tendon force—from surface EMG signals under isometric conditions. A simplified Hill-type model (Menegaldo 2017) is used to simulate muscle dynamics. A Kazantzis–Kravaris–Luenberger (KKL) observer, implemented via neural networks, embeds the nonlinear system into a linear latent space for robust state estimation. The observer is validated on synthetic and real EMG signals, including pathological cases. Results show strong generalization and robustness, offering a promising approach for real-time muscle monitoring and control. Future work will integrate KKL with modulating function methods for online parameter identification.