UBC Theses and Dissertations
Classification of multi-channel EMGs for jaw motion recognition by signal processing and artificial neural networks Fu, Bin
This thesis presents an EMG pattern recognition method to identify jaw movements. Pattern recognition is carried out using backpropagation artificial neural networks (BPN) trained by supervised learning. Different feature extraction methods have been implemented. Results are presented to support the feasibility of the suggested approach. Electromyography (EMG) is electrical activity of muscle. Bruxism is the involuntary and excessive clenching and grinding of teeth. It is one of the reasons that cause serious teeth damage and jaw muscle disorder and currently there is no definitive cure. Knowing actual jaw actions during bruxism will help in designing a more targeted treatment mode. EMG provides information about the neuromuscular activity from which it originates. When performing different muscle contractions, different EMGs are detected. Different muscle contractions are related to various movement tasks. Therefore it is possible to process the EMG to obtain movement classification. The purpose of this study is to design an algorithm to detect jaw motion (simulated bruxism) from the EMG signals. The process consists of feeding the EMG signals to a feature extraction block; and then the extracted features are supplied to an artificial neural network (ANN). The ANN classifies the motion into six categories: left, right and forward with two speeds, fast and slow. The application of this algorithm can provide the basis for a clinical evaluation of bruxism. In this study, three feature extraction techniques have been implemented for comparative analysis: EMG linear envelope (LE), autoregressive modelling (AR parameter estimation) and a hybrid approach of the AR model and discrete wavelet transform (DWT). The performance of the three methods has been evaluated. The ANN structure consists another part of the study. While both linear envelope method and AR model method were able to classify the direction of jaw movement, each confused with the fast and slow speed at the same direction to some extent. To solve the problem, a new ANN structure was designed for the linear envelope method by adding more features like the duration and the integral information of the waveform These features work as bias to the ANN structure and they not only yield high correct classification rates but also strengthen the robustness of the ANN. Finally the classification performance of the algorithm was checked in 10 healthy subjects and reasonable results were obtained. These were best with downsampled LE with the ANN (structure III).
Item Citations and Data