Abstract:

[Current prosthetic control schemes fall in roughly two categories: Classically programmed Direct Control, in which trigger signals switch between grip patterns, or Machine Lear- ning/AI based systems, in which Pattern Recognition is used to switch directly to the desired grip. Both systems have disadvantages, namely the slowness of Direct Control and the limited patient pool of Pattern Recognition based systems. This thesis aims to investigate weather the advantages of both systems can be merged in an Artificial Neural Network based approach to Direct Control. This system should at a minimum keep the robustness of Direct Control, speed up the trigger detection and provide patients the possibility to train the system with their own individual trigger signals. Eleven Neural Nets of three different Architectures are examined. The size of the Nets is constrained by having to run an a microprocessor. The training data is recorded like one would have to do for each patient resulting in a small dataset. Preprocessing with different methods is done to increase the available information for the Neural Networks. The effectiveness of these methods is also analysed. The Neural Networks are implemented in python using the tensorflow library. Their minimum size in the memory of a microprocessor and the required time of execution are calculated, in order to fit the constraint. This thesis shows, that it is possible to train Neural Networks successfully with a small dataset and that the preprocessing of data improves Neural Network performance. The investigated system however fails to match expectations. Constraints limit its perfor- mance and the overall complexity of the system and the training process, as well as the inconsistency of the results render this system not suitable for real world application.]