基于多源域脑电信号特征矩阵的选择性深度迁移学习方法研究

The electroencephalogram (EEG) pattern recognition methods can establish a mapping relationship between EEG data and complex brain cognitive activities involving perception, intention or emotion as well as pathological patterns, which is an important way to realize the accurate interpretation of human brain cognitive activities and the intelligent diagnosis of diseases by machine system. However, the low signal-to-noise ratio and non-stationary characteristics of EEG signals result in the poor robustness, generalization and adaptation of the recognition methods in practical applications. From the perspective of practicality, the main goal of this project is to mine electrophysiological knowledge and data characteristics from EEG signals to establish stability and efficient recognition models, the following researches will be carried out: (1) by introducing the sparse constraints of EEG artifacts and combining the multi-channel EEG feature matrix and its structural information, the project proposes a robust feature matrix classification model for noisy EEG signals; (2) based on the theory of stacking generalization, the project develops a stacking method of feature matrices between layers, and establishes a deep stacking network by using matrix classification model as the basic unit;(3) On the basis of optimizing the parameters of deep stacking networks from multi-source domains, the project proposes a selective deep knowledge transfer method, and then builds an adaptive recognition model with the lack of target domain data. This project provides new ideas for accurate identification of EEG signals, which is of great significance in breaking through the bottleneck of information transfer between human brain and machine systems and improving the diagnosis and treatment of disease.

基于机器学习的脑电信号识别方法能够在脑电数据与感知、意图、情感等复杂脑认知活动或病理模式之间建立映射关系，是实现机器系统对人脑活动准确判读以及疾病智能诊疗的重要手段。然而脑电信号的低信噪比与非平稳特点，使得脑电识别方法在实际应用中的鲁棒性、泛化性与自适应性方面始终存在局限。本项目从实用性角度出发，挖掘脑电生理知识与数据特点对其建立稳健高效的识别模型，将开展如下工作：(1)依据脑电特征矩阵及其结构信息，联合脑电伪迹的稀疏约束，建立含噪声脑电信号的特征矩阵分类模型；(2)以堆叠泛化为理论基础，建立层级间特征矩阵的堆叠方式，构建矩阵分类模型为基本单元的深度堆叠网络；(3)基于多个源域的深度堆叠网络的参数求解，建立选择性深度知识迁移方法，构造目标域数据匮乏下的自适应脑电信号识别模型。本项目为准确识别脑电信号提供了新思路，对突破人脑与机器系统间的信息交流瓶颈以及提升疾病诊治水平具有重要意义。

在脑电信号与感知、意图、情感等复杂脑认知活动或病理模式之间建立映射关系，能够对人脑活动的准确判断以及对疾病的智能化诊疗。为实现该目标，本项目研究的主要技术内容包括：1）基于特征信息及结构信息构建具有良好泛化能力的深度网络，提高脑电信号识别精度；2）设计新型基本堆叠单元，建立新型深度堆叠网络，实现鲁棒性脑电特征解码；3）建立选择性脑电模型知识迁移模型，缓解脑电训练数据不足的问题。据此，本项目完成了如下工作：1）搭建了基于虚拟场景的脑电数据采集平台；2）提出了基于深度堆叠泛化原理和迁移学习机制的深度迁移最小二乘分类器；3）提出了一种深度堆叠支持矩阵机；4）提出了分类误差一致性准则，并在此基础上建立了选择性迁移分类方法及自适应知识迁移学习方法；5）提出了一种基于知识利用的支持矩阵机；6）提出一种基于局部和全局结构感知熵正则化的平均教师模型；7）构建了一种集成共空间模型和支持矩阵机的特征解码模块，并在此基础上建立深度堆叠架构。通过本项目的实施，能够有效地提升脑电信号识别方法的泛化性、鲁棒性以及自适应性，为准确识别脑电信号、突破人脑与机器系统间的信息交流瓶颈、普及脑-机接口设备在某些疾病（如癫痫病、运动功能障碍）中的智能诊疗提供新思路与技术手段。