Auto-Encoder based Deep Learning for Surface Electromyography Signal Processing

摘要

Feature extraction is taking a very vital and essential part of bio-signal processing. We need to choose one of two paths to identify and select features in any system. The most popular track is engineering handcrafted, which mainly depends on the user experience and the field of application. While the other path is feature learning, which depends on training the system on recognising and picking the best features that match the application. The main concept of feature learning is to create a model that is expected to be able to learn the best features without any human intervention instead of recourse the traditional methods for feature extraction or reduction and avoid dealing with feature extraction that depends on researcher experience. In this paper, Auto-Encoder will be utilised as a feature learning algorithm to practice the recommended model to excerpt the useful features from the surface electromyography signal. Deep learning method will be suggested by using Auto-Encoder to learn features. Wavelet Packet, Spectrogram, and Wavelet will be employed to represent the surface electromyography signal in our recommended model. Then, the newly represented bio-signal will be fed to stacked autoencoder (2 stages) to learn features and finally, the behaviour of the proposed algorithm will be estimated by hiring different classifiers such as Extreme Learning Machine, Support Vector Machine, and SoftMax Layer. The Rectified Linear Unit (ReLU) will be created as an activation function for extreme learning machine classifier besides existing functions such as sigmoid and radial basis function. ReLU will show a better classification ability than sigmoid and Radial basis function (RBF) for wavelet, Wavelet scale 5 and wavelet packet signal representations implemented techniques. ReLU will illustrate better classification ability, as an activation function, than sigmoid and poorer than RBF for spectrogram signal representation. Both confidence interval and Analysis of Variance will be estimated for different classifiers. Classifier fusion layer will be implemented to glean the classifier which will progress the best accuracies’ values for both testing and training to develop the results. Classifier fusion layer brought an encouraging value for both accuracies either training or testing ones.