Finger movement pattern recognition method using artificial neural network based on electromyography (EMG) sensor

In this study, the EMG signals are processed using 16 time-domain features extraction to classify the finger movement such as thumb, index, middle, ring, and little. The pattern recognition of 16 extracted features are classified using artificial neural network (ANN) with two layer feed forward network. The network utilizes a log-sigmoid transfer function in hidden layer and a hyperbolic tangent sigmoid transfer function in the output layer. The ANN uses 10 neurons in hidden layer and 5 neurons in output layer. The training of ANN pattern recognition uses Levenberg-Marquardt training algorithm and the performance utilizes mean square error (MSE). At about 22 epochs the MSE of training, test, and validation get stabilized and MSE is very low. There is no miss classification during training process. Based on the resulted overall confusion matrix, the accuracy of thumb, middle, ring, and little is 100%. The confusion of index is 16.7%. The overall confusion matrix shows that the error is 3.3% and overall performance is 96.7%.

I. Introduction

Recently, a number of studies in bionics hand have developed significantly. The purpose of the studies is mainly to interpret finger movement. The common used sensor for measuring the activity of muscle is electromyography (EMG). The raw signal from EMG sensor is hard to understand and interpreted by human. Thus, the implementation of pattern recognition method that can interpret the muscle activity or finger movement has important role. This research was focused on the feature extraction and pattern recognition of five finger movement classification using EMG raw signals.