Performance Evaluation of SRELM on Bio-signal Pattern Recognition Using Two Electromyography Channels

The classification accuracy of pattern recognition is determined by the extracted features and the utilized classifiers. Many efforts have been conducted to obtain the best features either by introducing a new feature or proposing a new projection method to increase class separability. Recently, spectral regression extreme learning machine (SRELM) has been introduced to improve the class separability of the features. However, the evaluation of SRELM was only focused on the myoelectric or electromyography pattern recognition from many EMG channels. In practical application, the user is more convenient with less number of channels. Then, the problem is whether the SRELM would be able to work efficiently for less EMG channels. The objective of this paper is to examine the performance of SRELM for bio-signal pattern recognition using two EMG channels. The EMG electrodes were located on flexor policies lounges and flexor digitorium superficial muscles of ten healthy subjects. Various time domain features were involved with various sizes. SRELM will project these features to more recognize features before being feed to multiple classifiers.  Those classifiers are randomized Variable Translation Wavelet Neural Networks (RVT-WNN), extreme learning machine(ELM), support vector machine (SVM), and linear discriminant analysis (LDA). The performance of SREM was compared to other feature methods, such as LDA, uncorrelated LDA (ULDA), orthogonal fuzzy neighborhood dimensionality reduction (OFNDA), and spectral regression discriminant analysis (SRDA). The experimental results show that SRELM performed well when dealing with different class numbers by classification accuracy of around 95.67% for ten class movements and performed better than SRDA.