Sie sind auf Seite 1von 4

1

CHAPTER 1 INTRODUCTION

Objectives The primary objective of the presented research is to enhance the accuracy of the Electromyography (EMG) pattern classification by exploring various kinds of feature extraction methods. The temporal approach is good for the classification, based on the fact that significant temporal structures exist for different motion types [36]. The spectral approach is also proposed and tested in the pioneering works [29][33][36][40][42][43]. Both approaches are practical ways to classify the EMG patterns with good performances. Designing and implementing efficient EMG feature extraction methods in the prosthetic hand control based on these two approaches are the central part of this work. There rises an interesting question: How much useful information is contained within the frequency domain? As we will show in this work, the energy information has a good capability for the classification. Is it possible that most information in the features from frequency domain is just an alternative way to represent the energy information? If the answer is positive, then the temporal approach is a suitable and efficient way to extract the energy information. It can be implemented in a more computation efficient way since no transformation from time domain to frequency domain is necessary. On the other hand, it is encouraging to extract feature values hiding completely in the frequency domain. Using the frequency information together with the energy information will lead to a higher accuracy. It will increase the possibility to obtain a good

2 classification result by using a short length of EMG signal at the initial phase of motion. This research will provide valuable exploration for the designing of future real-time prosthetic hand control system. The secondary objective is to apply a neural network classifier named Mahalanobis distance-based simplified ARTMAP network (DFAM) [49][50] on EMG pattern classification. DFAM is a powerful and efficient neural network that can deal with stability-plasticity dilemma very well [51]. It represents the distribution of the input samples in an efficient way by using the covariance matrix and Mahalanobis distance. This work verified the performance of DFAM specifically for the EMG classification. The third objective is to introduce a methodology to analyze and process the EMG signal. It includes describing the characters of the EMG signals by using time-frequency analysis, revealing their structure properties by a set of measurements, designing the corresponding feature extraction methods, and creating the feature vectors ready for the classifier. The classification results are compared to evaluate the various kinds of feature extraction methods. Original Contribution This thesis designs and implements a number of temporal and frequency feature extraction methods. The extracted feature vectors are sent to a classifier to evaluate the performance of these feature extraction methods. The successful classification rate is significantly improved in this research, comparing to previous research work on this area [49]. Some experiments have the successful classification rates of more than 97%. This thesis provides the details of the algorithm, the mathematic equation for each experimented feature set, and the detail experimental results analysis. We believe that these are

3 practical and useful resources for the future work in this area, and for the feature extraction of other biomedical signals which have a similar nature as the EMG. Using the moments of power spectrum as features is original introduced in this work. It has good performance and is proved to be a suitable method for EMG feature extraction. The discrete prolate spheroidal sequences (DSPP) [16][17] are employed in multiple windows method, which belongs to the temporal approach. Using DSPP in temporal approach is a new idea. It is not only a new feature extraction method, induced by the fact that the DSPP play a key role for the good performance of the multi-taper method [19][34]. The multiple trapezoidal windows method is also a new method in this area. Its performance is surprisingly good while the implementation is so simple. The thesis provides a lot of experimental results. The extent to which the frequency information is useful for the EMG classification is revealed to some degree by the experiments. Part of the thesis is about the employed classifier. The Euclidean distance-based simplified ARTMAP network (ESAM) is discussed as originally proposed in previous paper [49]. Its detail algorithm is introduced as well as its good performance. Thesis Outline Chapter 1 states the major objectives of this work, the original contribution of this work and short descriptions for each chapter. Chapter 2 introduces the physiology background on generating and detecting of EMG signals. This is followed by an overview of the prosthetic hand control project and description of how the EMG pattern classification is used for this project [46]. The general processes of the classification are briefly introduced as well as two basic EMG feature extraction approaches

4 the temporal approach and frequency approach. Specifically, the feature extraction methods those are proposed and experimented in this thesis are briefly discussed. Chapter 3 gives a detail introduction of the classifier used in this work. The Simplified ARTMAP Network (SFAM) [2] based on Adaptive Resonance Theory (ART) [1][3] is discussed after a background introduction about pattern classification. There are two generalizations of SFAM. One is Euclidean distance based SFAM and the other is Mahalanobis distance based SFAM. We illustrate the nature of the distance based SFAM by applying the networks on two popular benchmark problems: circle-in-square [4] and two-spiral [7]. Chapter 4 discusses the temporal approach for EMG feature extraction. The multiple windows methods are proposed and tested. The experiments results are provided for an evaluation of these methods. Chapter 5 discusses the frequency approaches. The conceptual and mathematical background for the Fourier theory, power spectral density estimation, and the Thomsons theory are introduced. The wavelet transform are introduced as an addition to the spectral approach. We discussed how these theories are used for the time-frequency analysis of the EMG signal and how the feature extraction methods are designed based on the analysis. The feature extraction algorithms based on the short-time Fourier transform and the short-time Thomsons transform are presented as well as the corresponding experimental results. Chapter 6 summarizes all the experiments results. The classification accuracies for each proposed methods are compared and discussed. Finally a conclusion is reached that the efficient way to encode the EMG signals is by STTT.

Das könnte Ihnen auch gefallen