Improvement in Available Methods for Simultaneous and Proportional Control using the Kernel Technique for Unsupervised Myoelectric Intention Estimation of Individual Fingers

Recent advances in the field of biomedical engineering has prompted modern research to focus on challenges of human machine interface. This paper provides an improvement in unsupervised learning methods already available for estimating myoelectric intention of individual fingers using the kernel technique. The unsupervised methods which have been improved upon for simultaneous and proportional intention estimation are NMF and NMF-HP. These methods are called semi unsupervised algorithms as models are evaluated blindly using only the target finger. The algorithms implemented with kernels are named as kNMF and kNMF-HP. The kernel technique increases the feature matrix for the NMF and NMF-HP models and improves the performance of these algorithms. The algorithms were analyzed in terms of signal to noise ratio using the strength of the signal of the activated finger and the levels of other fingers not activated. Significant improvements were seen through the implementation of the kernel matrix on the parameters analyzed. An in-house eight channel signal instrumentation scheme was used to acquire the EMG signals using dry electrodes. In addition, a comprehensive signal filtering scheme was designed in order to remove the acquired EMG signal of noise. Finally, we used the algorithms to successfully drive a robotic hand.