Neural Network-Based Lower Limb Prostheses Control Using Super Twisting Sliding Mode Control

Abstract

This paper presents a method for controlling the prosthetic leg using surface Electromyography (sEMG) signals, Artificial Neural Network (ANN), and Super Twisting Sliding Mode Control (ST-SMC). The triggering signal is extracted from the user’s muscles and intense signal preprocessing that includes filtering, rectification, linearization, and Mean Average Value (MAV) feature extraction. The ANN predicts joint angles for walking, upstairs, and downstairs using the processed sEMG signals of the muscles and measured and filtered target joint angles. The neural network structure is built using Feed-forward Neural Network (FFNN) architecture and Levenberg-Marquardt (LM) back-propagation training algorithm for accuracy, fast convergence, and reliable optimization of nonlinear relationships. The ST-SMC controller regulates the motion of the prosthetic joints according to specified reference trajectories. MATLAB signal analyzers, neural network fitting packages, and Simulink are used to preprocess signals, train the FFNN for dynamic modeling of the system, and design controllers. The proposed ST-SMC is compared with conventional SMC. Simulation results show that training the neural network with processed data increases regression value and decreases trajectory tracking mean squared error (MSE). The controller’s robustness against internal parameter change and external environmental changes is demonstrated through parameter variation and disturbance analysis.