Myoelectric control techniques for a rehabilitation robot

The work presented in this thesis consists primarily of two projects. The first project was the design and development of a real time myoelectric controller using a pattern recognition scheme. The myoelectric control scheme controlled three degrees of freedom which included elbow flexion and extension, wrist pronation and supination, and hand grasping and releasing for a robotic arm in the Biomechatronics Learning Lab at the Rochester Institute of Technology. According to the knowledge of the author, no work has ever combined these three DOF. The design started with an offline analysis of common windows and features found in the literature. Data was obtained from ten healthy subjects and was tested to find the optimal window and feature scheme which provided for the highest classification accuracy. The highest classification accuracy was 94.92% for a 250ms windowing scheme with three autoregressive features from a fourth order model. The classifier used in all of the testing was a linear discriminate analysis. The real time myoelectric control scheme was implemented in Labview and used an adjacent windowing scheme of 250ms with AR features. The same ten healthy subjects were then used to test the real time myoelectric control scheme. The average classification accuracy during the real time testing was 89.52%. The real time myoelectric control scheme was then adapted to a subject with Central Cord Syndrome. He was first tested with single degree of freedom controllers and obtained classification accuracies of 100%, 95.24%, and 80.95% for elbow, hand, and wrist controllers respectively. When tested with the full three degrees of freedom controller, the subject achieved a classification accuracy of 68.25%. The second project involved the theoretical design of a myoelectric controller based on a time delayed neural network. Advantages to this design were that the EMG used for this control scheme were based on complex reaching motions and allowed for the control of multiple degrees of freedom at once. These two advantages are not currently offered in myoelectric control schemes which typically control one degree of freedom at a single instance in time and are based on repeatable isometric contractions. The algorithm was based on previous work completed by Au and Kirsch. Their optimal parameters used a total delay of 875ms and 125ms as the interval of delay. The total delay suggested by Au and Kirsch is not possible for a real time scheme. This work tested for the feasibility of using the time delayed neural network as a real time myoelectric control scheme by decreasing the total delay. Five subjects were used in this work. All of the time delayed neural networks were trained using multiple types of motion and speeds to make the TDNNs robust. The first subject was tested with different TDNNs that used total delays of 900ms, 600ms, and 300ms, delay intervals of 50ms, 100ms, and 150ms, and hidden layer neurons of 10, 20, 30, and 40. The optimal parameters of a 300ms total delay with a 100ms delay interval, and a hidden layer of 10 neurons resulted in an average error of 15.7° for the first subject. These parameters were then used to test the data from the remaining subjects. Using the optimal parameters an average error of 19.0° for all subjects was obtained. Previous errors reported by Au and Kirsch were on the order of 20°. This work showed that the total time delay could be decreased. The next step for this work would be to implement the algorithm in real time or make attempts to decrease the output error