Use of neural networks to differentiate wrist movements using muscle signals

Currently, technological advances in bionics are considerable, providing the user with the possibility of regaining the ability to hold the elements and an adequate rehabilitation. Proper placement of electronic equipment on the prosthesis allows reading of muscle signals from the forearm, however measurements are mainly focused on the movement of the fingers when wrist movement is paramount to ensure a greater number of possible movements for the hand, for this reason, the use of processing algorithms as a neural network reduces dependence on this electronic equipment. In this research work, an algorithm of a mechanical control has been designed considering the six movements of the wrist, bending, extension, radial deviation, cubital deviation, pronation and supination using sensors that record data every 0.5 seconds by storing 50 signals per movement for neural network training. For best results, the training process was performed in the Matlab Program using its Deep Learning Toolbox package with a very near zero error. As a next step, two tests were performed on the neural network, the first with four movements of the wrist with a result of 88.9% accuracy and the second using the six movements of the wrist with a result of 92.9% accuracy. In addition, a validation was performed between the training and the tests performed with a regression with Pearson R correlation results for the neural network. The results indicate that deep learning and electronic elements favor the training of a neural network to control the movement of the wrist.