An In-depth Study of Ankle-Foot Orthosis Dynamics Modeling: Leveraging Non-Parametric Approach Via Artificial Neural Networks

Walking is one of the most important daily activities for human beings. Patients that have abnormal walking gait are caused by foot drops, strokes, and other disabilities. Ankle-foot orthosis (AFO) are widely used to provide practical assistance for patients with injuries or defects in the lower limbs. There are many types of AFOs, including rigid, flexible rigid, and articulated AFOs, depending on the strength of the joint. Modelling ankle-foot orthosis is important because it has always been a challenging task to model ground reaction forces, particularly when the wearable rehabilitation robot is represented using high degrees of freedom and has multiple contact points with the ground. The research is aimed at modelling ankle-foot orthosis (AFO) using a multi-layer perceptron (MLP) neural network. Initially, data collection took place using an experimental rig. Subsequently, the model structure was chosen, followed by parameter estimation through the selected algorithm. Lastly, the models underwent a thorough validation process, which included evaluating their performance using mean-squared error (MSE) and correlation tests. The results showed that the MLP-NN outperformed the conventional method, LS in identifying the AFO system, with lower mean squared prediction error that is 0.000011034 and unbiased results across all models. In contrast to the conventional approach, the MLP-NN offers a good approximation of the AFO dynamic model. Although conventional methods like LS are valuable, the MLP approach exhibits superior performance. These findings provide valuable insights into AFO system modeling, implying that nonparametric methods like MLP neural networks hold significant potential for advancing AFO development and control.