基于多视角脑电信号深度模型知识迁移的下肢运动意图在线预测方法研究

The recognition of movement intention based on EEG signals has become an important means of limb rehabilitation training and effectiveness evaluation. However, because of spatial distribution characteristics of the brain regions activated by the lower-limb movement intention, automatic EEG recognition method cannot meet the needs of clinical rehabilitation applications. From the perspective of functional segregation and integration in the brain, this project proposes a collaborative learning and transfer learning framework based on the EEG deep features from different views, and builds an online prediction model for lower-limb movement intentions. The following researches will be carried out: (1) according to the activation pattern of the local brain region and the global brain network, the project proposes two feature extraction methods based on EEG time-space-frequency high-dimensional space and dynamic brain functional network respectively, and builds a deep feature representation model of EEG by the multi-view collaborative learning; (2) by introducing the distribution similarity of EEG data between subjects and combining the sparse constraint of EEG data from the target subject, the project proposes a model knowledge transfer learning framework for deep neural network; (3) based on the stack generalization theory, the project establishes a stacking method between layers of deep neural networks, and constructs an online prediction model for non-stationary EEG time series. This project provides new ideas to break through bottlenecks in terms of robustness, generalization and timeliness of EEG recognition methods. It is of great significance for studying the treatment mechanism of lower limb rehabilitation training and improving the effect of auxiliary diagnosis.

基于脑电信号的运动意图识别已成为肢体康复训练和效果评估的重要手段。然而由于下肢运动意图所激活脑区空间分布复杂，使得自动化脑电识别方法在康复临床应用上受到局限。本项目从脑功能分离与整合机理呈现的角度出发，根据不同视角下的脑电深度特征建立协同学习及迁移学习框架，构建下肢运动意图在线预测模型，将开展如下工作：(1)针对局部脑区和全局脑网络的激活模式，分别提出基于脑电时-空-频高维空间和动态脑功能网络的特征抽取方法，构建基于多视角协同学习的脑电深度特征表示模型；(2)依据受试者间脑电数据的分布相似性，联合目标受试者脑电数据的稀疏约束，建立深度网络模型知识迁移学习框架；(3)以堆叠泛化理论为基础，建立深度网络层间堆叠方法，构建非平稳脑电时间序列的在线预测模型。本项目为突破脑电识别方法的鲁棒性、泛化性和时效性瓶颈提供了新的思路，对研究下肢运动康复治疗作用机制以及提升辅助诊疗效果具有重要意义。

基于脑电信号的肢体动作意图解码已成为肢体功能康复训练和效果评估的重要手段。为实现该目标，本项目研究的主要技术内容包括：1）学习不同视角的脑电特征表示，构建协同学习框架，实现精确的脑电信号识别；2）设计受试者间脑电信号的迁移模型，提高脑电识别模型的泛化性能；3）建立含监督信息的深度网络层间堆叠策略，提升脑电信号预测模型的时效性。本项目完成了如下工作内容：1）建立了面向下肢运动意图识别的脑机交互实验平台；2）提出了一种基于脑电特征结构信息的深度矩阵学习模型；3）提出了一种基于核范数正则化的脑电深度特征学习框架；4）设计了一种协同学习策略，学习不同视角下脑电数据结构一致性信息；5）分别提出自适应多模型知识迁移的支持矩阵机方法以及基于双重监督机制的脑电信号领域泛化方法，用于跨受试者脑电信号识别；6）分别提出含监督信息的层间堆叠策略以及自适应多层迁移策略，简化模型的优化过程。通过本项目的实施，能够增强脑电识别模型的鲁棒性、泛化性以及时效性，为脑机交互在肢体功能康复中的应用提供新的思路和方法。