| 基于原型网络的小样本脑电伪迹检测方法 |
| |
| 引用本文： | 李景聪,林镇远,潘伟健,吴潮煌,潘家辉. 基于原型网络的小样本脑电伪迹检测方法[J]. 华南师范大学学报(自然科学版), 2022, 54(4): 113-120. DOI: 10.6054/j.jscnun.2022065 |
| |
| 作者姓名： | 李景聪 林镇远 潘伟健 吴潮煌 潘家辉 |
| |
| 作者单位： | 华南师范大学软件学院,广州 510631 |
| |
| 基金项目： | 国家自然科学基金项目62006082广东省自然科学基金项目2021A1515011600广东省自然科学基金项目2020A1515110294广州市科技计划项目202102020877 |
| |
| 摘 要： | 正常脑电信号由于容易受到多种脑电伪迹的干扰而导致信噪比低,为了提高脑电信号的信噪比,使用一种基于度量的小样本学习模型来检测脑电信号中的伪迹,提出了一种基于原型网络的脑电伪迹识别模型(EEG Artifact Prototype Network,EAPNet)。该模型能够学习一个从EEG特征到目标空间的非线性映射,然后计算每个类原型表示的距离,并按此距离进行分类; 仅需较少数量的数据样本进行训练,就能实现对伪迹的准确识别。最后,在公开的脑电伪迹数据集TUAR(TUH EEG Artifact Corpus)中进行了伪迹识别实验,并将EAPNet模型与2个深度学习模型(EEGNet、全连接神经网络(FNN))及7个机器学习模型(高斯贝叶斯模型(Gaussian NB)、随机森林模型(RF)、逻辑回归模型(LR)、套索回归模型(Lasso)、支持向量机模型(SVM)、岭回归模型(Ridge)和最近邻算法(KNN))进行了对比实验。实验结果显示:(1)EAPNet模型是一种高效的伪迹检测方法:在2-way 1-shot、2-way 5-shot、2-way 10-shot任务中,模型的检测准确率分别为69.44%、77.21%、80.01%。(2)在所有对比模型中,EAPNet模型的识别准确率最高。
|
| 关 键 词： | 脑电信号 脑电伪迹 小样本学习 原型网络 |
| 收稿时间： | 2021-08-12 |
A Method of Few-shot EEG Artifact Detection Based on the Prototype Network |
| |
| Affiliation: | School of Software, South China Normal University, Guangzhou 510631, China |
| |
| Abstract: | Normal EEG signals are susceptible to contamination by various EEG artifacts, resulting in low signal-to-noise ratios. In order to improve the signal-to-noise ratio of EEG signals, a metric-based few-shot learning method was used to detect EEG artifacts. A few-shot learning model named EEG Artifact Prototype Network (EAPNet) based on a prototype network was proposed to detect artifacts in EEG signals. The model was able to learn a nonlinear mapping from EEG features to the target space, calculate the distance of each class prototype representation and classify them according the distance. In addition, the EAPNet model required only a small number of examples of each new class to train the model to achieve accurate recognition of artifact signals. Artifact recognition experiments were conducted in the public EEG artifact dataset TUAR (TUH EEG Artifact Corpus). The EAPNet model is compared with 2 deep learning models (EEGNet, Fully Connected Neural Network (FNN)) and 7 machine learning models (Gaussian NB, Random Forest (RF), Logistic Regression (LR), Lasso Regression (Lasso), Support Vector Machine (SVM), Ridge Regression (Ridge) and Nearest Neighbor Algorithms (KNN)) in a comparative experiment. The experimental results showed that the EAPNet model was an efficient artifact detection method, with a detection accuracy of 69.44%, 77.21% and 80.01% for 2-way K-shot(K=1, 5, 10) tasks respectively. Among all the compared models, the EAPNet model had the highest recognition accuracy. |
| |
| Keywords: | |
|
| 点击此处可从《华南师范大学学报(自然科学版)》浏览原始摘要信息 |
|
点击此处可从《华南师范大学学报(自然科学版)》下载全文 |
|
