因脑电具有高时间分辨率、无创性和低成本等优点，情感脑机接口已被广泛应用于情感识别研究中。然而，高维脑电特征导致基于少量样本的情感脑机接口应用面临分类过拟合、计算复杂度高和实时性差等问题。为了解决上述问题，特征选择算法被用于剔除无用和噪声特征，并挑选富含判别信息的脑电特征子集。其中，相较于过滤式和包裹式特征选择方法，嵌入式特征选择方法通过充分考虑特征在学习任务中的重要性，从而挑选更具判别力的特征子集。然而，当前情感脑机接口研究中常用特征选择方法无法克服特征子集内高度冗余性问题，且忽略了多维度情感标签内关联信息，从而无法挑选出对多维度情感识别的最优特征子集，给基于情感脑机接口的情感精准识别带来困难。

本文以上述问题为导向，开展面向情感脑机接口的嵌入式特征选择算法研究。本文的主要工作和创新点如下：

（1）为了克服以往嵌入式特征选择方法所挑选特征子集内冗余信息问题，研究一构建了两种新型特征选择模型，分别是冗余最小化限制正交回归的两阶段特征选择模型（Orthogonal Regression with Minimum Redundancy, ORMR）和全局冗余最小化限制正交回归的特征选择模型（Global Redundancy Minimization in Orthogonal Regression, GRMOR）。基于正交回归和冗余最小化准则，ORMR方法和GRMOR方法可以通过挑选特征权重矩阵中权重值较高的特征，构建表征能力强且非冗余的特征子集，从而克服以往嵌入式特征选择方法所挑选特征子集内冗余信息问题。基于广义幂迭代法(Generalized Power Iteration, GPI)和增广拉格朗日乘子法（Augmented Lagrangian Multiplier, ALM）两种方法，本研究构建了GRMOR模型的优化算法，用于解决GRMOR模型优化过程所面临的斯蒂弗尔流形的二次问题（Quadratic Problem on the Stiefel Manifold, QPSM）；

（2）为了克服以往嵌入式多标签特征选择算法无法同时考虑自变量和因变量的测量误差，以及无法充分利用多标签内全局关联信息两个局限性，研究二构建了一种全局冗余和全局相关最优化约束正交回归的嵌入式多标签特征选择算法（GRROOR）模型。GRROOR模型利用多标签内全局相关性信息约束多标签的潜在语义索引过程和低维子空间构建过程。GRROOR模型通过正交回归模型同时学习多标签数据的局部和全局结构信息，并同时引入自变量和测量误差分析。十个国际公开多标签数据集的实验结果证明了该优化算法的有效性和收敛性；

（3）为了克服颅内容积传导效应导致的脑电特征全局冗余性问题，研究三利用研究一中GRMOR 模型从高度冗余的脑电特征中挑选非冗余且富含信息的紧致特征子集。为了验证GRMOR在情感脑机接口应用中有效性，本研究采集了一套 128 通道的高密度EEG 情感数据集，并结合脑电情感计算领域现有的DEAP和SEED国际公开数据集，联合验证了GRMOR算法性能。本研究通过分析GRMOR模型从三个脑电情感数据中所挑选脑电特征对应空间电极位置，发现了三个数据集上共有脑电电极CP5，为未来情感脑机接口采集设备电极分布研发提供了参考；

（4）为了克服现有的情感脑机接口研究忽略多维度情感内关联信息问题，研究四基于研究二中GRROOR模型从全局角度分析多维度情感标签内关联信息。为了降低GRROOR模型在多维度情感特征选择任务中计算复杂度，本研究将GRROOR模型进行了简化，构建了一种面向多维情感识别的嵌入式脑电情感特征选择模型（EEG Feature Selection Method for Multi-dimension Emotion Recognition, EFSMDER）。本研究通过进一步分析不同类型脑电特征对于多维度情感识别任务的重要性，发现了时-频特征对多维度情感识别任务的有效性，该发现将为情感脑机接口中特征提取研究提供借鉴意义。

综上，本文围绕面向情感脑机接口的嵌入式特征选择，针对脑电特征冗余的特点，提出了有监督的最小化冗余约束的ORMR和GRMOR模型；并在此基础上，基于情感多维相关的特点，提出了多维标签全局相关约束的GRROOR模型，实现了面向情感脑机接口的多维度情感特征选择。上述结果为基于脑电信号的情感识别和穿戴式情感脑机接口采集设备研发等相关研究提供了一定的思路。

Due to the advantages of high temporal resolution, non-invasiveness and low cost of Electroencephalogram(EEG), emotional brain-computer interface have been widely used in affective computing researches. However, high-dimensional EEG features lead to the problems of classification overfitting, high computational complexity and poor real-time performance for emotional brain-computer interface researches based on a small number of samples. To solve the problem, various feature selection algorithms were employed to eliminate useless (noisy) features and select EEG feature subsets with discriminative information. Compared with the filter and wrapper feature selection methods, the embedded feature selection method selects a more discriminative feature subset by fully considering the importance of features in the learning task. However, the current feature selection methods, commonly used in current emotional brain-computer interface researches, cannot overcome the problem of high redundancy in the feature subsets, and ignore the correlation information in multi-dimensional emotional labels. They makes the commonly used EEG emotion feature selection methods unable to select the optimal subset of features for multi-dimensional emotion recognition, which makes it difficult for accurate emotion recognition based on emotional brain-computer interface.

To solve the above problems, this paper conducts the research on embedded feature selection algorithms for emotional brain-computer interface. The main work and innovation points of this paper are as follows:

(1) In order to solve the problem of redundant information in the feature subsets selected by previous embedded feature selection methods, two novel feature selection models are proposed: ORMR and GRMOR. based on the orthogonal regression model and redundancy minimization criterion, ORMR and GRMOR can obtain representive feature subsets with non-redundant information by selecting features with higher weighting in the feature weight matrix, which could overcome the problem of redundant information in the feature subsets. In this study, the optimization algorithm of GRMOR is constructed by Generalized Power Iteration (GPI) and Augmented Lagrangian Multiplier (ALM), which could solve the QPSM problem in the optimization process of the GRMOR model;

(2) The existing embedded multi-label feature selection algorithms cannot consider the measurement error of both independent and dependent variables, and cannot fully utilize the global correlation information in the multi-label. To overcome the above two limitations, an embedded multi-label feature selection algorithm (GRROOR) model with global redundancy and correlation optimization constraints on the orthogonal regression is constructed. The GRROOR model first uses the global correlation information within the multi-label matrix to constrain the latent semantic indexing process and the low-dimensional subspace construction process of the multi-label. Second, the GRROOR model simultaneously learns the local and global structural information of multi-label data through an orthogonal regression model, and is able to introduce independent variables and measurement error analysis simultaneously. The experimental results on ten international public multi-label datasets demonstrate the effectiveness and convergence of the optimization algorithm;

(3) To overcome the global redundancy problem of EEG features caused by intracranial volume conduction effects, this study employed the GRMOR model to select non-redundant and informative compact feature subsets from highly redundant EEG features. In order to verify the effectiveness of GRMOR in the application of emotional brain-computer interface, this study collected a set of 128-channel high-density emotional data set(HDED), combined with the existing DEAP and SEED international public datasets in the field of EEG-based affective computing. By analyzing the corresponding the spatial electrode positions of the selected EEG features the GRMOR model, it could be found that the EEG electrode CP5/50the was shared in the three data sets, which could provide a reference for the electrode distribution of future emotional brain-computer interface acquisition equipments;

(4) In order to overcome the problem that the existing emotional brain-computer interface researches ignore the multi-dimensional emotional correlation information, this study introduces this GRROOR model to analyze the multi-dimensional emotional intra-related information from a global perspective. To reduce the computational complexity of the GRROOR model in the multi-dimensional emotional feature selection task, this study simplifies the GRROOR model and constructs an embedded EEG emotional feature selection model for multi-dimensional emotion recognition (EFSMDER). By further analyzing the importance of different types of EEG features for the multi-dimensional emotion recognition task, this study found the effectiveness of time-frequency features for the multi-dimensional emotion recognition task, which could provide a reference for EEG emotion feature extraction researches in emotional brain-computer interface.

To sum up, this paper focuses on the embedded feature selection for emotional brain-computer interface, and proposes supervised ORMR and GRMOR models to minimize redundancy in the selected EEG feature subsets. According to the characteristics of multi-dimensional correlation, a GRROOR model with global correlation constraints of multi-dimensional labels is proposed, and the multi-dimensional emotional feature selection for emotional brain-computer interface is realized. The above results provide useful information for EEG-based emotion recognition and the development of wearable emotional brain-computer interface.