A Bioinspired Feature-Projection-Based Approach to Electromyographic Pattern Recognition Based for High Dimensional Sparse Sensor Data

The paper presents an electromyographic pattern recognition for sensor fusion able to discern motions of hand with a small number of training samples. We propose a learning algorithm able to classify estimate class statistics from a limited training set size. A Wavelet Packet Decomposition performs feature extraction. A sparse Principal Component Analysis projects the features in a lower dimensionality space. Classification is performed through multi-layer Perceptron. We employed sparse Principal Component Analysis because it is insensitive to the curse dimensionality problem differently from standard Principal Component Analysis that fails to capture discriminatory information in low-variance sensor data. The approach mitigates drawbacks of the training data collection as time consumption and acquisition difficulties. The latter are particularly relevant in case of high degrees of user disability, in which long sessions of training become unfeasible due to stress and exertion.