Hannes Prosthesis Control Based on Regression Machine Learning Algorithms

“The world around us is by and large shaped to be operated by hands and arms: our homes, our work-places, the means of everyday transportation”. Not only hands allow us to interact with the physical world, but they also help in greeting and talking. The loss of an upper limb incredibly lowers the quality of life, leading to severe impairment in daily-living operational functionality as well as to psychological damages. Since the beginning of civilization, humans started searching for improvements in amputees life. Initially, the adopted solutions were based on replacing the part of the body missing by means of mechanical and passive devices. Along with the technology evolution and electronic improvements, prostheses became mechatronic systems controlled by means of human intentions. The remaining neurons inside the stump are still able to carry information through little electrical signals that travel from the brain to the residual limb muscles in a fraction of a second. When the signal reaches the stump, it is withdrawn by artificial electrodes and it is used for making the active prosthesis move. Nevertheless, the EMG signals conversion into prosthesis movement is still a challenging problem. Nowadays the most widespread control strategy for the upper limb prosthesis is based on two-muscles contraction which leads the amputees to an unnatural feeling of control. Impaired people must learn a new way of thinking to make the prosthesis perform the desired movements. Since the clinically available myoelectric devices provide limited ability in the control of multiple joints at the same time, the movements of the prosthesis appear to be clumsy. Therefore, the control is narrowed to the single DoF and the switching from one joint movement to another is usually implemented through simultaneous contraction of the two antagonist muscles (co-contraction). The latest research studies on supervised machine learning algorithms, based on pattern recognition, have demonstrated the possibility of making the control more natural by extracting crucial features from several EMG signals of the residual limb. The thesis aims to apply and test the pattern recognition algorithm for controlling concurrently multiple-joints of Hannes by using up to six EMG signals withdrawn at the forearm level. Hannes is the poly-articulated hand, developed by IIT and INAIL, able to restore over 90% of lost functionality in people with trans-radial amputation. Five different supervised machine learning algorithms are compared: Artificial Neural Network, Non-linear Logistic Regression, Regularized Least-Squares, Support Vector Machine and Linear Discriminant Analysis. Ten healthy subjects and four amputees are involved in the study to collect sEMG signals for the comparative analysis of the foregoing classifiers. The data is exploited to build different models in order to assess the performances in predicting the users gestures. Although the promising results, pattern recognition algorithms are not still applied to the everyday life of the amputees due to their low robustness in scenarios different from their training. Often, the different placement of the electrodes could confuse the algorithm, significantly lowering the prosthesis usability. Additional developments will surely improve the robustness of the algorithm by using high-density sensors for gathering more and more information or by exploiting data coming from extra-sensors (i.e. IMU, haptic feedback, …).