Exoskeleton Control System In The Surface Emg Signal Processing And Recognition

Systematic research on the exoskeleton system began in the sixties, however, due to technical and material limitations, and poor performance, the development of the exoskeleton was slow. Recently, with the success of the Berkeley Lower Extremity Exoskeleton (BLEEX) and the Hybrid Assistive Leg (HAL) interest in exoskeleton systems re-surged. Lightweight materials with outstanding strength characteristics have been developed, control methods have become highly sophisticated and intelligent, and autonomous power sources have been developed. Another key piece of technology is the use of surface electromyography (sEMG) for controls. Employing pattern recognition techniques, sEMG is regarded as an effective method for providing control commands for the exoskeleton. With an extensive analysis of existing sEMG feature extraction and pattern classification, the thesis discusses different motion patterns of the arm obtained from sEMG signals in order to improve the accuracy rate and the speed of the sEMG recognition. The major studies and innovation of the work are as follows.1. An autoregressive (AR) parameter model was used to extract and characterize the sEMG signals recorded by the MyoTrac Infiniti Clinical T9850USb. In our experiments, we chose the length of Short-Time Fourier Transform to be 512 and the Hanning window with a length of 256. We selected the AR of order 4 and sEMG signals from the biceps collected during the forearm contraction and extension were employed. The results showed that the AR model can improve the signal processing speed with enhanced stability.2. The effective action signals were extracted from the sEMG which were picked up from the human arm undergoing different motions, such as the hands extending or holding and the forearms rotating in or out. The AR model was used to extract EMG signal feature. Then, the back-propagation (BP) neural network with an input layer consisting of 8 neurons (8 was the number of AR coefficients) and an output layer with 4 neurons (4 for the classes of movement) was used. The neural network was trained using 10 groups of movements by healthy trials. Then we used the trained network to identify different motions of the human hand and we obtained reasonably good results.