Development of a myoelectric prosthesis control based on deep learning. Optimisation of the neural network parameters and EMG detection system.

The most commonly used commercially available prostheses for transradial amputees offer a limited number of movements of the prosthetic hand. They are old-design prostheses with only one possible closure movement controlled using a direct control strategy. These prostheses are single-joint devices with a pinch that involves the movement of all the five fingers at the same time, with no option for individual finger movement. The control strategy relies on the remaining muscular activity from the stump in an easy and intuitive way for the user. These devices are very easy to control and durable to be used in everyday activities, but they do not provide the possibility of moving the prosthesis in a more natural and physiological way. Having a multiarticulate prosthetic device that allows a higher number of gestures remaining easy and intuitive to control would be a great achievement for transradial amputees. Some trials have been made to implement the control of a greater number of gestures through direct control strategies but with disappointing results due to the low intuitiveness of the control mechanism that makes these devices annoying to use every day. Only using Pattern Recognition algorithms it is possible to develop a modern prosthetic device, intuitive to control and with an increased number of performable gestures. This thesis presents a custom developed pattern recognition algorithm based on an Artificial Neural Network (ANN) and tests the ability of this algorithm to classify a total of nine hand/wrist gestures using a conventional bipolar EMG acquisition system and a high-density EMG system (HD-sEMG). With respect to the state of the art, one of the main achievements of this thesis project is the introduction of wrist movements among the nine gestures to manage, while all the studies in literature reporting such a high number of gestures have included only hand gestures. The ANN implemented is a MultiLayer Perceptron with a number of input neurons equal to the number of EMG channels; the number of hidden layers and neurons is automatically optimized in the training process; the output layer has nine neurons corresponding to the nine gestures to classify. The first acquisition set up used to obtain input features is composed by six bipolar electrodes evenly spaced along the entire circumference of the forearm 5 cm distal from the elbow. The second one is composed by a grid of 24x4 electrodes spaced 10 mm obtaining seventy-two single differential channels. Both bipolar and HD-sEMG were used to train the classifier to understand if the HD-sEMG sampling the muscular activity from the forearm with a high spatial resolution and from a wider area provides some improvements in the gesture recognition with respect to conventional bipolar system. Several analyses have been conducted to optimize the movement recognition performances for both the two detection systems by tuning the preprocessing, network and postprocessing parameters. Finally, a statistical comparison of the performances of the two detection systems was conducted.