脑机接口（Brain computer interface，BCI）是一种实现大脑与外部设备之间直接进行交互的通信系统。其中，运动想象脑机接口是一种主动式BCI范式，主要将运动控制意图映射为不同的运动想象任务。通过提取不同运动想象任务诱发的脑电（Electroencephalogram，EEG）等生理信号中的判别性特征，建立相应的运动想象解码模型，以识别运动控制意图。

近年来，基于脑电信号的运动想象（Motor Imagery Electroencephalogram，MI-EEG）神经解码研究取得了很多进展。但总的来看，目前的运动想象任务是将运动意图映射为大粒度肢体运动想象任务，输出的动作意图类别较少、使得BCI指令集数量受限，应用流畅性以及扩展性不足。结合手语引导的运动想象脑机接口基于语言学基础建立手部运动与BCI指令间的联系，为BCI指令的扩展提供了新的途径。当前运动想象脑机接口研究从时-频-空等多个维度探讨了脑电数据中运动想象相关特征提取的理论与方法。跨频率耦合特征能够为运动想象任务提供更多的全局交互信息，然而，目前研究往往关注于特定脑区或频带间耦合关系，忽略了其他脑区之间的跨频交互作用，并且对脑区和频段的自动选择仍是难点。此外，运动想象脑电特征提取和融合中面临的一大挑战是脑电的个体差异性，因此，当前研究的一个重点方向是个体水平上的运动解码模型。然而，这类研究受限于实验设计中实验时长与脑电采集环境的限制，普遍存在小样本问题。为了获得全面且具有判别性的表征，并进一步提高运动想象BCI识别性能，针对上述问题与不足，本文围绕手语指令运动想象神经解码任务，从小样本数据增广、全局稀疏耦合特征表征以及多域特征融合神经解码模型构建三部分展开研究，主要创新工作如下：

（1）针对MI-EEG解码研究中的小样本问题，本文提出了基于变分模态分解的脑电数据自适应增广方法。首先，使用基于聚类的方法获得样本子集，并选取聚类中心用于表征样本子集，以数据驱动的方式自动筛选合适的样本进行数据增广，充分利用原始样本的分布信息。其次，为了克服增广数据存在信息量不足等局限，引入变分模态分解算法，重组包含原始脑电数据时频空动态特性的模态分量，充分保留原始数据的关键特征，实现脑电数据的有效增广。实验结果表明，本章的增广方法能够提升手语指令运动想象任务的解码精度，并且优于对比方法。

（2）针对MI-EEG全脑空间信息提取问题，本文提出了跨节律耦合特征表征模型。首先，通过多层耦合网络捕捉全脑全频段的跨频交互信息，获得手语运动想象任务的多域全局特征表示。然后，构建了一种基于l₁范数和惩罚权重的脑网络的稀疏化策略，利用耦合强度变化来引导网络稀疏优化，自动保留跨频域耦合网络中的关键核心连接的拓扑结构。实验结果表明，本章提出的脑网络稀疏化策略能够对网络拓扑结构进行有效优化，赋予多域全局稀疏耦合特征更好的表征能力，提高手语运动想象任务的解码准确率。

（3）针对MI-EEG解码模型多域综合特征挖掘问题，本文构建了双分支多域特征提取和融合结构。一方面，通过空频和时频双分支特征提取模块，深入挖掘运动想象脑电数据中全局和局部特征，有效提取不同层次的多域特征表征信息。另一方面，引入跨域注意和监督对比学习策略，通过注意力机制有效融合不同层次的特征信息，同时约束融合特征的分布，增强融合特征的判别能力，从而获得更为有效的域间共享特征。利用残差模块保留多域特征视图间的互补信息，增强特征表示的完备性，实现多域信息的综合特征挖掘。所提的方法在四类手语运动想象任务的分类准确率达到45.42%±4.11%，优于对比方法。

本研究为MI-EEG多域特征分析及解码模型提供了新的研究思路，有望进一步促进精细肢体运动想象脑机接口的发展，在辅助控制设备方面具有重要的应用前景。

Brain-Computer Interface (BCI) is a communication system that realizes the direct interaction between the brain and external devices. Motor imagery brain-computer interface is an active BCI paradigm, which mainly maps the motion control intention to different motor imagery tasks. By extracting the discriminant features of EEG or other physiological signals induced by motor imagery tasks, decoding model identify the motor imagery task and decipher the motor control intention.

In recent years, a lot of progress has been made in neural decoding of motor imagery based on EEG signals. However, in general, the current motor imagery task is to map motion intention to large-grained limb motor imagery task, and the output of motion intention categories is few, which makes the number of BCI instruction set limited, and the application fluency and scalability are insufficient. The brain-computer interface of motor imagery guided by sign language establishes the connection between hand movement and BCI instruction on the basis of linguistics, and provides a new way for the extension of BCI instruction.The current research on the motor imagery BCI discusses the theory and method of extracting motor imagery related features from EEG data from multiple dimensions such as time-frequency-space. Cross-frequency coupling features can provide more global interaction information for motor imagery tasks. However, current studies tend to focus on coupling relationships between specific brain regions or frequency bands, ignoring cross-frequency interactions between other brain regions, and it is still difficult to automatically select brain regions and frequency bands. In addition, one of the major challenges in the feature extraction and fusion of EEG in motor imagery is the individual variability of EEG, so a focus of current research is motor decoding models at the individual level. However, such studies are limited by the length of experimental duration and EEG collection environment in the experimental design, which generally has the problem of small sample size. In order to obtain a comprehensive and discriminative representation and further improve the performance of motor imagery BCI recognition, in view of the above problems and deficiencies, this paper focuses on the neural decoding task of sign language instruction motor imagery, including three parts: small-sample data augmentation, global sparse coupling feature characterization, and multi-domain feature fusion neural decoding model construction. The main innovations are as follows:

To solve the problem of small sample size in MI-EEG decoding, an adaptive augmentation method of EEG data based on Variational Mode Decomposition(VMD) is proposed. Firstly, cluster-based method is used to obtain the sample subset, and the cluster center is selected to characterize the sample subset. In a data-driven way, appropriate samples are automatically screened for data augmentation, and the distribution information of the original samples is fully utilized. Secondly, to overcome the limitations of insufficient information content in the augmented data, VMD algorithm is introduced to reassemble the time-frequency and space dynamic characteristic components of the original EEG data, fully retain the key features of the original data, and realize the effective augmentation of the EEG data. Experimental results show that the augmentation method in this chapter can improve the decoding accuracy of sign language instruction motor imagery tasks, and is better than the contrast method.

To solve the problem of extracting spatial information from the whole brain of MI-EEG, a cross-rhythm coupled feature representation model is proposed. Firstly, multi-layer coupled network is used to capture the cross-frequency interaction information of the whole brain and the whole frequency band, and the multi-domain global feature representation of the sign language motor imagery task is obtained. Then, a sparse brain network strategy based on l₁ norm and penalty weights are constructed, and the coupling strength changes are used to guide the sparse optimization of the network, and the topology of key core connections in the cross-frequency coupled network is automatically preserved. The experimental results show that the sparse brain network strategy proposed in this chapter can effectively optimize the network topology, engender better representation of multi-domain global sparse coupling features, and improve the decoding accuracy of sign language motor imagery tasks.

To solve the problem of multi-domain comprehensive feature mining in MI-EEG decoding model, a two-branch multi-domain feature extraction and fusion structure is constructed. On the one hand, through the dual branch feature extraction module of space frequency and time frequency, the global and local features in the motor imagery EEG data are deeply mined, and the multi-domain feature representation information of different levels is effectively extracted. On the other hand, cross-domain attention and supervised contrast learning strategies are introduced to effectively integrate feature information of different levels through the attention mechanism, and at the same time restrict the distribution of fusion features, to enhance the discrimination ability of fusion features, to obtain more effective inter-domain shared features. The residual module is used to preserve the complementary information between multi-domain feature views, enhance the completeness of feature representation, and realize the comprehensive feature mining of multi-domain information. The proposed method has a classification accuracy of 45.42%±4.11% in the four types of sign language motor imagery tasks, which is superior to the comparison method.

This study provides a new research idea for MI-EEG multi-domain feature analysis and decoding model, which is expected to further promote the development of brain-computer interfaces for fine limb motor imagery, and has important application prospects in auxiliary control devices.