Procesamiento de Señales de Electroencefalograma usando Redes Neuronales para aplicaciones en Sistemas BCI
Brain-Computer Interfaces (BCI) are systems that provide an alternative communication between the human brain and a computer, where electroencephalography (EEG) is the the non-invasive and most viable way to obtain the electrophysiological activity. However, the EEG register presents several importa...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2020 |
| País: | México |
| Institución: | Universidad Autónoma de Querétaro |
| Repositorio: | Repositorio Institucional de la Universidad Autónoma de Querétaro |
| Idioma: | español |
| OAI Identifier: | oai:ri-ng.uaq.mx:123456789/2011 |
| Acceso en línea: | http://ri-ng.uaq.mx/handle/123456789/2011 |
| Access Level: | acceso abierto |
| Palabra clave: | Interfaces Cerebro-Computadora Redes Neuronales Separación de Fuentes Ciegas INGENIERÍA Y TECNOLOGÍA |
| Sumario: | Brain-Computer Interfaces (BCI) are systems that provide an alternative communication between the human brain and a computer, where electroencephalography (EEG) is the the non-invasive and most viable way to obtain the electrophysiological activity. However, the EEG register presents several important challenges for both the instrumentation and the signal processing techniques involved in a BCI implementation. Some of the most relevant drawbacks are due to the low signal-to-noise ratio, the presence of undesirable signals such as ocular, cardiac, and muscular activity, as well as a low spatial resolution due to the distance and high impedance between the sources and the location of the electrodes. Therefore, the raw of EEG signals have very low amplitude, very low signal-to-noise ratio, and considerable noise contamination. Digital signal processing, and machine learning have been included in the preprocessing, feature extraction and classication stages of BCI systems in order to improve the signal-to-noise-ratio and hence, increase their effciency. The rst part of the present dissertation consists of comparing the performance of different preprocessing algorithms to estimate the original sources from the EEG registers; particularly, the preprocessing was made using Blind Source Separation (BSS) algorithms. This kind of spatial lters are based on Second Order Statistic (SOS) or High Order Statistic (HOS) information. Most representative BSS algorithms are (SOBI, SOBIRO, fastICA, and Infomax) were compared using semi-simulated sources, using the Pearson's correlation coecient and the Wavelet Coherence (WC) as metrics. This analysis was conducted in the electrophysiological bands. On the other hand, an analysis of dierent descriptors in time and frequency domain was performed for the extraction of relevant features was made in order to nd the most relevant information and thus reduce the dimensionality of input data. In this stage, the used classier was an Multilayer Perceptron (MLP). Finally, new method built from the combination of a Blind Source Separation (BSS) to obtain estimated independent components, a 2D representation of these component signals using the ContinuousWavelet Transform (CWT), and a classication stage using a Convolutional Neural Network (CNN) has been proposed. A criterion based on the spectral correlation with a Movement Related Independent Component (MRIC) is used to sort the estimated sources by BSS, thus reducing the spatial variance. The experimental results of 94.21% using a k-fold cross validation are competitive with techniques recently reported in the state-of-the-art. |
|---|