毕业生知识库

作为日常生活中的关键生理信息，情绪在许多神经认知科学以及精神疾病 的诊断中起着至关重要的作用。与在情绪识别中容易伪装的发声、面部表情、手势和身体姿势等行为信号相比，脑电信号(electroencephalogram,EEG)难以伪装且具有高时间分辨率，基于EEG的情绪识别因其高准确性和可靠性而得到广泛 研究。然而，由于EEG信号个体间分布差异大且具有时间非平稳性，不满足传统深度学习假设的独立同分布要求，因此情绪识别很难对新的个体或者同一个 体的新时间采集的数据保持良好的分类精度。
本研究着力于将最新的迁移学习算法和动态网络结合，并将两者的组合作为 基本EEG信号解码模型的框架，充分利用已有数据的知识，设计出高效的EEG 情绪识别模型,实现高精度的跨个体以及跨时间的EEG情绪识别，并尝试在抑郁 诊断的临床应用中验证模型的效果。本研究的主要工作内容包括以下两个方面：
(1)本研究构建了基于EEG信号的多源动态情绪识别模型，引入了一个动态 特征提取器，通过自适应样本实现多源域适应，缓解多源域间的分布冲突。同时， 结合最大均值差异和最大分类器差异的域适应算法设计出的三段式对抗性训练 机制用于拉近源域和目标域之间的距离，共同促进特征提取器挖掘域不变和情 感可区分的特征。该模型在SEED和SEED-IV数据集中进行了全面的实验评估， 并与代表性方法进行了比较，验证了所提出方法的优越性，证明了提出的模型为 多源场景下基于EEG的情绪识别提供了更有效的迁移学习途径。此外，基于模 型对EEG信号频段和通道对情绪识别贡献度进行分析，验证了模型的有效性。
(2)EEG信号抑郁诊断是情绪识别的临床应用，精神疾病已被证实与情绪紧 密相关。但是现有的EEG信号抑郁诊断算法很少关注个体独立问题，忽略了现 实场景中个体间EEG信号极大的差异性，因此个体独立的抑郁检测仍然具有挑 战性。本研究基于在EEG信号情绪识别中表现良好的模型框架提出了一种结合 了动态卷积特征提取器、对抗域适应算法和关联域适应算法的抑郁诊断模型。对 抗域适应算法用于实现边缘分布域适应，关联域适应算法用于实现条件分布域 适应。在公开数据集和真实数据集中都进行实验验证，证实了所提出的跨个体的 抑郁诊断模型的临床应用价值。

Emotion is a crucial aspect of daily life and plays a significant role in various neu rocognitive sciences and the diagnosis of psychiatric disorders.However,recognizing emotions accurately from behavioral signals,such as vocalizations,facial expressions, gestures,and body postures,can be challenging,as these signals can be easily disguised. In contrast,electroencephalogram(EEG)physiological signals have high temporal reso lution and are difficult to be disguised,making them a promising tool for emotion recog nition.Despite the high accuracy and reliability of EEG-based emotion recognition, these signals exhibit large inter-individual differences and temporal non-smoothness, which do not conform to the traditional deep learning assumption of independently and identically distributed training and test sets.As a result,it is difficult to maintain high classification accuracy for new individuals or the data from different sessions of the same individual.
To address these challenges,this study proposes a framework for an efficient EEG based emotion recognition model that combines the latest transfer learning algorithms and dynamic networks.This framework leverages existing data to achieve high accuracy across individuals and time and aims to validate the model in clinical applications for diagnosing depression.The main objectives of this study are twofold:
(1)This study presents a multi-source dynamic emotion recognition model based on EEG signals using a dynamic feature extractor to achieve multi-source adaptation by adapting samples and alleviate distribution conflicts between domains.To close the gap between source and target domains,a three-stage adversarial training algorithm is designed by combining the domain adaptation algorithm of maximum mean difference and maximum classifier difference.The model aims to jointly promote the feature ex tractor to mine domain-invariant and emotion-distinguishable features.The proposed algorithm is comprehensively evaluated on the SEED and SEED-IV datasets,demon strating its superiority over the representative methods used.The results confirm that the proposed model provides a more effective transfer learning pathway for EEG-based emotion recognition in multi-source scenarios.The contribution of EEG signal fre quency bands and channels to emotion recognition is analyzed,and the interpretability of the model is verified.
(2)EEG-based depression diagnosis is a clinical application of emotion recogni tion since psychiatric disorders have been closely linked to emotions.However,existing EEG depression diagnosis algorithms often ignore the great variability of EEG signals among individuals in realistic scenarios,making individual independent depression de tection still challenging.In this study,we propose a subject-independent depressiondiagnosis model that combines a dynamic convolutional feature extractor,an adversarial domain adaptation algorithm,and an associative domain adaptation algorithm based on a model framework that has shown good results in EEG signal emotion recognition. The adversarial domain adaptation is used to implement marginal distributed domain adaptation,and the associative domain adaptation is applied to implement conditional distributed domain adaptation.The model is validated on both public and real datasets, confirming the clinical application of the proposed cross-subject depression diagnosis model.