الفهرس 
Myoelectric prosthesis control
Training and control systemsOnly 14 pages are availabe for public view
 |  | from | 77 |  |  |
| | | from | 77 | | |
Abstract
For decades, bio-signals analysis and processing have been used for diagnosis of neural and biomechanical problems. Bio-signals have been dealt with just as indicators of human’s vital processes for diagnosis and evaluation, e.g. ECG for heart beats monitoring. The advances in biomedical instruments have enabled the acquisition of most bio-signals e.g. EEG, ECG and EMG. Consequently, the study of interpretation of such signals led to efficiently interface human neural system with smart devices, which maybe prostheses that move per amputee’s intention interpreted from a human’s bio-signal.
Myoelectric hand prostheses are active prostheses that get use of surface EMG of the forearm muscles remaining after amputation, using surface electrodes. Being plug and play; commercial prostheses are mostly myoelectric, mostly they use pattern recognition control scheme, which lacks simultaneous motion, and perform robotic unnatural inter-pattern motion. Accordingly, the use of regression models for the estimation of hand kinematics proportionally to sEMG signals has proved more simultaneous and natural motion.
The objective of this study was to introduce proportional speed control on robotic hand motion, where each finger has its own estimator model to achieve non-robotic performance of the hand. Each finger has four regression models to cover the motion of the finger over four patterns. A pattern recognition classifier was trained to classify four hand gestures; accordingly, the regression models of the fingers is to be altered according to the classifier decision. A commercial sEMG sensing armband was used in the acquisition of training data that can be used later in the development of the prosthetic control system. The reproduction of data for linear (least-square fitted model) and non-linear (ANN) regression models were investigated, where the ANN proved better reproducibility of finger speeds compared to those captured by Leapmotion IR cameras during training sessions.
The models also were trained on reduced RMS features, where the selected features are only the channels that are allocated over the active muscles during performing the patterns which resulted in a reproducibility of 89.27±1.92%. These results demonstrate the robustness of the multi-regression models system over wide range of motion.