Abstract
Accurate estimation of finger joint stiffness is important in assessing the hand condition of stroke patients and developing effective rehabilitation plans. Recent technological advances have enabled the efficient performance of hand therapy and assessment by estimating joint stiffness using soft actuators. While joint modular soft actuators have enabled cost-effective and personalized stiffness estimation, existing approaches face limitations. A corrective approach based on an analytical model suffers from actuator-finger and inter-actuator interactions, particularly in multi-joint systems. In contrast, a data-driven approach struggles with generalization due to limited availability of labeled data. In this study, we proposed a method for energy conservation-based online tuning of the analytical model using an artificial neural network (ANN) to address these challenges. By analyzing each term in the analytical model, we identified causes of estimation error and introduced correction parameters that satisfy energy balance within the actuator-finger complex. The ANN enhances the analytical model's adaptability to measurement data, thereby improving estimation accuracy. The results show that our method outperforms the conventional corrective approach and exhibits better generalization potential than the purely data-driven approach. In addition, the method also proved effective in estimating stiffness in human subjects, where errors tend to be larger than in prototype experiments. This study is an essential step toward the realization of personalized rehabilitation.