STATISTICAL PATTERN RECOGNITION TECHNIQUE FOR IMPROVED REAL-TIME MYOELECTRIC SIGNAL CLASSIFICATION

The authors in this paper propose a statistical technique for pattern recognition of electromyogram (EMG) signals along with effective feature ensemble to achieve an improved classification performance with less processing time and memory space. In this study, EMG signals from 10 healthy subjects and two transradial amputees for six motions of hand and wrist is considered for identification of the intended motion. From four channels myoelectric signals, the extracted time domain features are grouped into three ensembles to identify the effectiveness of feature ensemble in classification. The three feature ensembles obtained from multichannel continuous EMG signals are applied to the new classifiers namely simple logistic regression (SLR), J48 algorithm for decision tree (DT), logistic model tree (LMT) and feature subspace ensemble using k-nearest neighbor (kNN). Novel classifiers SLR, DT and LMT, select only the dominant features during training to develop the model for pattern recognition. This selection of features reduces the processing time as well as memory space of the controller for real-time application. The performance of SLR, DT, LMT and feature subspace ensemble using kNN classifiers are compared with other conventional classifiers, such as neural network (NN), simple kNN and linear discriminant analysis (LDA). The average classification accuracy with SLR is found to be better with feature ensemble-1 compared to the other classifiers. Also, the statistical Kruscal–Wallis test shows, the classification performance of SLR is not only better but also takes less time and memory space compared to other classifiers for classification. Also the performance of the classifier is tested in real-time with transradial amputees for actuation of drive for two intended motions with TMS320F28335eZdsp controller. The experimental results show that the SLR classifier improves the controller response in real-time.