Master thesis presentation
Evaluation of post-processing strategies for simultaneous pattern recognition based myoelectric prosthetic control
Presented by Joel Falk-Dahlin
Examiner: Bo Håkansson
Supervisor: Max Ortiz
Multi-functional prostheses are developed to improve the amputee’s quality of life. Myoelectric control is a field that has shown great potential when it comes to controlling these artificial limbs. This approach makes use of the electrical signals that are generated by muscle contractions to create an intuitive way for the user to control the prosthesis. Modern pattern recognition algorithms applied to the myoelectric signals have proven capable of distinguishing between multiple user intentions. For example, patterns observed when the user tries to open the amputated hand can be used to open the artificial hand, creating a very intuitive control. Modern systems can recognize the different intentions with very high offline accuracy (>90%), but some of the motions are still misclassified. The effect of the misclassified motions can be reduced by using different post-processing algorithms in real-time control.
A new myoelectric control system that allows the user to control several movements simultaneously is currently developed at Integrum AB. To further develop this system, this work has focused on the need of post-processing algorithms that work together with the simultaneous control.
The real-time misclassifications have been studied to identify key problems with the simultaneous control approach. Two post-processing algorithms used for individual movement control, Majority Vote and the Decision-Based Velocity Ramp, have been modified to work with the simultaneous control system. The effect of using Majority Vote has been evaluated in a simulation of the system, and the effect of using the Decision-Based Velocity Ramp has been evaluated in a series of real-time tests.
Results showed that the Majority Vote cannot improve the real-time classification accuracy significantly but it can be used to achieve the same performance with reduced computational cost. This algorithm was modified to work with simultaneous control and two different versions have been suggested in this work. Subjects reached targets more directly when they used the Decision-Based Velocity Ramp, which suggests that controllability of the virtual limb was increased. The Decision-Based Velocity Ramp can be used the same way as it is with individual control, and still improve the controllability of the simultaneous control system.