【摘要】 脑电波采集过程往往会包含生理噪声和外部噪声，外部噪声目前可以通过滤波进行消除，去除生理噪声通常采用的方法是自适应滤波器、空间滤波和主成分分析，但这些方法需要一定的信号作为先决条件，并且去噪性能有限。近年来，深度学习技术开始运用于脑电波去噪，而且在去噪性能上得到了一定提升。鉴于在生成对抗网络（GAN）的判别器和生成器博弈过程中，生成样本具有逐渐向真实样本逼近的特点，设计一种基于卷积神经网络（CNN）的生成对抗网络模型，用于消除脑电波生理噪声中的肌源性噪声和眼源性噪声，同时在GAN的生成器网络中引入新的损失函数，使去噪后的数据与原始数据更加接近。设计的GAN-1D-CNN模型在去除肌源性噪声上的相关系数达到0.945，在去除眼源性噪声上的相关系数达到0.894。实验结果表明，GAN-1D-CNN模型对脑电波的去噪能力得到了增强，相关性能指标都优于现有基准数据集上的去噪方法。更多还原

【Abstract】 The process of EEG acquisition often includes physiological noise and external noise. At present, the external noise can be eliminated by filtering. The usual methods to remove physiological noise are adaptive filter, spatial filter and principal component analysis, but these methods need a certain signal as a prerequisite, and the denoising performance is limited. In recent years, deep learning has been applied to EEG denoising, and the denoising performance has been improved to a certain extent. In the generative adversarial networks, given the advantage that the generated samples gradually approach the real samples in the game between the discriminator and generator. A generated adversarial network model based on the convolutional neural network is designed to eliminate myogenic artifacts and ocular artifacts in EEG signals, and a new loss function is introduced into the Network to make the denoised data more similar to the original data. The designed GAN-1D-CNN model has a correlation coefficient of 0.945 for removing muscle induced noise and 0.894 for removing eye induced noise. The experimental results show that the GAN-1D-CNN model has enhanced its denoising ability for EEG waves and outperforms the existing denoising methods on the benchmark dataset in terms of relevant performance indicators.更多还原