Clasificación de señales EEG basada en representaciones bidimensionales y redes neuronales convolucionales

It is challenging to design an electroencephalogram (EEG) signal representation that efficiently represents input information to a convolutional neural network (CNN). CNN has been successfully applied to naturally represented two-dimensional data, such as images, however EEGs are time series. In thi...

Descripción completa

Detalles Bibliográficos
Autor: Edgar Hernández-González
Tipo de recurso: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2020
País:México
Institución:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:español
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/2001
Acceso en línea:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/2001
Access Level:acceso abierto
Palabra clave:info:eu-repo/classification/Design an electroencephalogram/Señales de electroencefalograma
info:eu-repo/classification/Convolutional neural network/Red neuronal convolucional
info:eu-repo/classification/Mental calculation/Cálculo mental
info:eu-repo/classification/cti/1
info:eu-repo/classification/cti/12
info:eu-repo/classification/cti/1203
info:eu-repo/classification/cti/120304
Descripción
Sumario:It is challenging to design an electroencephalogram (EEG) signal representation that efficiently represents input information to a convolutional neural network (CNN). CNN has been successfully applied to naturally represented two-dimensional data, such as images, however EEGs are time series. In this work, a new two-dimensional representation of EEG signals was developed, which makes it possible to take advantage of the success of CNNs in images. The model classifies EEG signals from motor imagery or mental calculation; the preprocessing consisted of applying the common average reference (CAR) and an 8- 30Hz band pass filter. Two representations were proposed which are spectrograms with the short-time Fourier transform (STFT) and scalograms with the continuous wavelet transform (CWT). Two classifiers were used: CNN-2D and CNN-2D + LSTM. The neural networks used, in addition to automatically extracting features, also classify the EEG signal. Unlike other models, in this work the same pre-processing, the same number of channels and the same network architecture was used for all subjects in each data set. The accuracy results were 71.34 % for BCI IV-2a (four classes), 80.71 % for BCI IV-2a (two classes), 73.82 % for BCI IV-2b, 83.57 % for BCI II-III and 82.10 % for mental calculation. In addition to obtaining competitive results with the state of the art, the training and prediction time of the proposed CNNs was very low.