| Summary: | Some electromyogram (EMG) signals include information from limb functions and have been used to control the movement of prosthesis. To develop practical malfunction myoelectically controlled prosthesis using an EMG signal, the best identification model for limb functions has to be established. The thesis presents an investigation and comparison between different identification methods for limb functions based on the use of statistical processing, neural networks, and fuzzy methods. The research focused on the establishment of robust identification models of limb functions in order to obtain objective comparison results. Therefore, some factors affecting the robustness of an identification model were investigated in detail, such as, how to choose the length of signal, extract and select the optimal features, and establish and select a model. The analysis of the power spectrum and eigenvalues of the autocorrelation matric of a surface EMG signal was used to decide the length of signals. A selection method of optimal features was presented through the analysis of the advantage and disadvantage of the Jeffries-Matusita distance measurement and the cluster separation index (CSI). The confidence interval of the recognition rate of a model was used to measure the robustness of the model and decide the size of the training and test sets. The different identification models were established using the optimal features using different pattern recognition methods, based on statistical, neural network and fuzzy methods. During ther modelling process of neural networks, the Bayesian technique was used to avoid the model tending to the idiosyncrasies of the test set. During the fuzzy logic modelling process, methods for extracting fuzzy rules were investigated to establish a good fuzzy identification model for limb functions. In the selection process of model, different selection methods of models were used according to the practical situation. Finally, noise analyses were completed to assess the sensitivity of some methods to noise. The comparisons were made between different identification methods based on the identification results, their confidence inv=tervals and their modelling methods. The best processing methods were concluded
|
|---|
