厦门大学学报（自然科学版）2025，Vol.64Issue(6)：992-1004,13.DOI:10.6043/j.issn.0438-0479.202411001
基于多分支特征生成对抗网络的脑电信号去噪方法
An EEG denoising method based on multi-level feature generation adversarial network
摘要
Abstract
[Objective]Electroencephalography(EEG)is a widely used method for detecting brain electrical activities.It offers advantages of non-invasiveness,high temporal resolution,and cost-effectiveness.Nevertheless,EEG signals are susceptible to being affected by physiological artifacts during the recording process due to their low amplitude,which may mislead the data analysis and seriously affect the interpretation of results.Currently,most deep learning methods for EEG denoising typically employ a single-level structure,thereby suffering from limited feature representation and loss of signal details.[Methods]To tackle this issue,we proposed a multi-level feature generation adversarial network(MFGAN)for EEG denoising.To begin with,we construct a multi-level generator with the following steps:At the first level,we utilize the convolution neural network(CNN)to learn the shallow features of the target signal.At the second level,we use a Transformer-based encoder to obtain the contextual features from the EEG signal.At the third level,we deepen the CNN structure to acquire the deep detailed features of the signal.Furthermore,a feature self-filtering fusion module(FSFM)based on the attention mechanism is designed in the generator for extracting the multi-level features.FSFM applies a self-attention mechanism to filter feature information after channel interaction and a channel attention mechanism to capture feature information of each channel.These two mechanisms run parallel to each other,effectively fusing the aforementioned feature information and eliminating noise components.[Results]This method was qualitatively and quantitatively validated using the EEGdenoisenet and MIT-BIH arrhythmia datasets,with the aim to evaluate the denoising performance of MFGAN on electrooculography(EOG),electromyography(EMG),electrocardiography(ECG)and hybrid artifacts(EMG and EOG)in EEG signals.A systematic evaluation was conducted using three performance indicators,including correlation coefficient(CC),signal-to-noise ratio(SNR),and relative root mean square error(RRMSE).In qualitative experiments,we compared the denoising effects of different methods under ECG and EOG noise conditions.Results indicate that MFGAN can greatly reduce peak overflow and waveform distortion,thereby producing samples that more closely resemble pure EEG signals.In quantitative experiments,the proposed method outperformed the previous optimal method across three evaluation metrics.Specifically,in the EMG single noise environment,the proposed method led to a 0.99％increase in CC,a 4.26％increase in SNR,and a 6.72％decrease in RRMSE compared to the suboptimal model.Besides,compared to the suboptimal model the CC increased by 2.1％,the SNR increased by 6.49％,and the RRMSE decreased by 8.69％in the hybrid dual-noise environment.To gain a better intuitive understanding of the differences in denoising effects between different methods,we also visualized the quantitative results that highlight the superiority and potential of MFGAN in EEG denoising.[Conclusion]The proposed model is based on the generative adversarial network(GAN)structure,which integrates CNN,Transformer,and advanced CNN to sensitively capture shallow details of the signal,understand contextual dependencies and enhance the level of feature extraction to reveal the hidden deep patterns.In addition,introducing attention mechanisms to optimize the feature fusion further promotes the denoising performance.Experimental results show that MFGAN outperforms current mainstream denoising techniques and excels in diverse noisy environments.Consequently,this study not only offers an effective solution to the problem of inadequate feature extraction in conventional single-level architectures,but also presents a dependable data processing approach for neuroscience research.关键词
脑电信号/伪影消除/生成对抗网络/多分支生成器/注意力机制Key words
electroencephalography/artifact elimination/generation adversarial network/multi-level generator/attention mechanism分类
信息技术与安全科学引用本文复制引用
缪佳伶,何倬利,王德清,颜佳泉,陈海兰,冯慧斌..基于多分支特征生成对抗网络的脑电信号去噪方法[J].厦门大学学报（自然科学版）,2025,64(6):992-1004,13.基金项目
福建省发树慈善基金会资助研究专项(MFK23006) （MFK23006）
闽江学院科教联合专项(MJKJ24006) （MJKJ24006）
福建省自然科学基金面上项目(2023J01807) （2023J01807）
福建省自然科学基金引导性项目(2021H0054) （2021H0054）
集美大学科研启动基金(ZQ2021028) （ZQ2021028）
福建省工业引导性(重点)项目(2023H0001) （重点）
