Application of biosignal-driven intelligent systems for multifunction prosthesis control

University of Technology, Sydney. Faculty of Engineering and Information Technology.Prosthetic devices aim to provide an artificial alternative to missing limbs. The controller for such devices is usually driven by the biosignals generated by the human body, particularly Electromyogram (EMG) or Electroencephalogram (EEG) signals. Such a controller utilizes a pattern recognition approach to classify the EMG signal recorded from the human muscles or the EEG signal from the brain. The aim of this thesis is to improve the EMG and EEG pattern classification accuracy. Due to the fact that the success of pattern recognition based biosignal driven systems highly depends on the quality of extracted features, a number of novel, robust, hybrid and innovative methods are proposed to achieve better performance. These methods are developed to effectively tackle many of the limitations of existing systems, in particular feature representation and dimensionality reduction. A set of knowledge extraction methods that can accurately and rapidly identify the most important attributes for classifying the arm movements are formulated. This is accomplished through the following: 1. Developing a new feature extraction technique that can identify the most important features from the high-dimensional time-frequency representation of the multichannel EMG and EEG signals. For this task, an information content estimation method using fuzzy entropies and fuzzy mutual information is proposed to identify the optimal wravelet packet transform decomposition for classification. 2. Developing a powerful variable (feature or channel) selection paradigm to improve the performance of multi-channel EMG and EEG driven systems. This will eventually lead to the development of a combined channel and feature selection technique as one possible scheme for dimensionality reduction. Two novel feature selection methods are developed under this scheme utilizing the ant colony arid differential evolution optimization techniques. The differential evolution optimization technique is further modified in a novel attempt in employing a float optimizer for the combinatorial task of feature selection, proving powerful performance by both methods. 3. Developing two feature projection techniques that extract a small subset of highly informative discriminant features, thus acting as an alternative scheme for dimensionality reduction. The two methods represent novel variations to fuzzy discriminant analysis based projection techniques. In addition, an extension to the non-linear discriminant analysis is proposed based on a mixture of differential evolution and fuzzy discriminant analysis. The testing and verification process of the proposed methods on different EMG and EEG datasets provides very encouraging results