Lower Limb Motion Intention Recognition Based on sEMG Fusion Features

Abstract

Surface Electromyography (sEMG) has been extensively used for gait analysis and robotic control. Since the low accuracy of intention recognition limits the development of exoskeleton robots, a dynamic adaptive neural network algorithm based on the multi-feature fusion of sEMG signals was proposed to realize the accurate identification of eight lower limb movements. Firstly, the original sEMG signals were collected from eighty volunteers through gait experiments. Secondly, the feature parameters of time domain, frequency domain, and sample entropy were extracted and merged. Then, the improved differential evolution (DE) algorithm was proposed to optimize the weight values of each feature for data fusion. Thirdly, the backpropagation neural network(BPNN) used in this study is the most common model for intention recognition. However, BPNN has the problems of low accuracy and slow convergence speed, this paper proposed a novel dynamic adaptive neural network model (GA-DANN). The novel model uses genetic algorithm (GA) to optimize the weights of the BPNN to accelerate convergence. The dynamic adaptive learning rate was introduced to overcome the difficulty of determining the learning rate. The experimental results show that the proposed algorithm with multi-feature fusion information could accurately distinguish eight lower limb motion intentions, and the averaged recognition accuracy reaches 94.89%, which is 10% higher than the traditional BPNN. The averaged recognition time is only 109.67ms. This research will provide technical support for the application of rehabilitation robots.