Estimation of lower limb rehabilitation exoskeleton torque by combining dual-joint parameter identification and neural network

Abstract

Lower limb rehabilitation exoskeleton is widely used for rehabilitative training. Estimating the torque of the lower limb exoskeleton can help identify the patient's intent, thereby enhancing engagement in rehabilitative training. Parameter identification (PI) is used to estimate torque. However, the presence of unmodeled dynamics and external disturbances poses challenges for achieving reliable torque estimation. Consequently, achieving accurate torque estimation is a primary research focus in this field. This study combines dual-joint parameter identification and neural network, for estimating joint torque in lower limb rehabilitation exoskeletons. This method enhances the performance of parameter identification optimization algorithms by employing Markov-based Particle Swarm Optimization and Gradient Descent Algorithm (MPG). Additionally, it independently identifies the parameters of the hip and knee joints, thereby enhancing the accuracy of torque estimation for each joint. The estimated physical parameters of the model and joint state variables are then utilized as inputs to the neural network for estimating the torques during the lower limb exoskeleton training process. MATLAB simulation demonstrates that employing MPG for parameter identification enhances fitness by 37.59 % and 15.24 % when compared to Particle Swarm Optimization(PSO) and Gradient descent (GD), respectively. Through experimental verification conducted under controlled disturbances, method for combining dual-joint parameter identification and neural networks (DPI-BP) demonstrates its effectiveness in accurately estimating torque in lower limb rehabilitation exoskeletons. Angle, velocity, acceleration, inertia matrix, Coriolis matrix, gravity matrix and friction matrix of hip and knee joints are taken as inputs for DPI-BP. The application of DPI-BP results in a reduction of torque estimation errors, specifically by 0.12 Nm and 1.40 Nm(P<0.001), corresponding to a decrease of 66.57 % and 14.35 % when compared to the PI and Backpropagation (BP) methods, respectively. The torque estimation error of hip and knee joints are 0.86 Nm and 0.54 Nm.