基于机器人刺激与分类器群对脑负荷的调制与估测的自适应脑-机交互系统研究

The event-related potential with low signal-noise ratio causes low accuracy or speed of detecting human’s intention in the potential-based brain-robot interaction system, which cannot satisfy the need of controlling a robot. The traditional system is hardly to adjust itself to the individual difference of the event-related potential because of fixed parameter, which decreases its ability of adapting to various people and environments. Since mental workload affects the characteristics of brain signal and performance of the interaction system, and our previous study has found and demonstrated that the meaning of image stimulus influences the features of event-related potential. This project uses meaning of robot stimulus to modulate the mental workload, in order to increase the signal-noise ratio of event-related potential and improve the efficiency of recognizing. The project makes the modulation of stimulus meaning, the recalling and other cognitive processes explicit by combining various paradigms and simulating situations of the interaction, e.g., load accumulation and over-load. It tries to understand these processes by analyzing the interconnection of multiple brain areas. This project trains multiple classifiers under various conditions of mental workload to form a classifiers group, combines it with a curve of mental workload to estimate people’s mental state, and designs an adaptive model to adjust the parameters automatically based on the estimated state. This project could improve the accuracy rate, information transfer rate and adaptability of the brain-robot interaction system. In addition, it is an elementary exploration of the process of understanding visual robot meaning in brain, and is a reference of selecting visual stimulus for visual brain-computer interface.

事件相关电位的较低信噪比造成基于此电位的脑-机器人交互系统辨识意图的精度或者速度难以满足控制需求；传统系统的固定参数不利于适应电位的个体差异，导致其对人和环境的泛化能力较差。由于脑负荷影响脑电特征和系统的交互效果，且申请人发现图片刺激的含义丰富度影响事件相关电位的特征，因此针对上述问题本课题利用机器人刺激的含义丰富度调制脑负荷，进而增大事件相关电位的信噪比和辨识效率。通过范式融合、交互中的脑负荷积累与过量脑负荷情况的模拟，和脑区耦合等实验与分析方法突出并解析刺激含义对脑负荷的调制，以及对回忆等其他脑认知过程的影响。本课题通过在多种脑负荷下训练分类器形成分类器群，并结合脑负荷曲线估测脑状态，从而构造基于脑负荷的自适应模型来自动地调整系统参数。本课题不仅有助于提高系统的辨识准确率、信息传输率和适应性，而且初步探索了大脑理解机器人的视觉含义的过程，为视觉脑-机接口的刺激设计提供一定的借鉴。

脑电解码是提升脑机接口性能的重要方面。脑电解码提升包括增强脑电信号质量和先进化解码算法两方面。脑电信号质量在视觉刺激脑机接口中与刺激属性相关，为此本项目从三方面展开研究。首先针对刺激内容的视觉结构采用样本延迟匹配和欧德宝范式融合的方式，发现全局结构通过减小脑负荷可产生特征更强的诱发脑电信号。其次基于刺激内容造成脑电差异的发现，开展结合不同视觉刺激的界面设计，提升了系统的输出速度。最后，在优化视觉刺激的基础上开展视觉、听觉脑负荷冲突研究，发现听觉通路刺激对视觉通路的脑负荷影响，通过主客观数据诱发与验证了低、中、高三种脑负荷，解释了脑机接口在日常生活中出现解码不稳定的原因。.提高解码算法先进性是从分类器群、深度学习和迁移学习三方面展开。对实验诱发的三种脑负荷，进行不同负荷调整权重的自适应分类群算法研究，提高了解码的自适应性。借鉴深度学习在计算机视觉的动态捕捉思想，将脑电自然数据转化为脑电图视频流，从空间与时间共同构建三维卷积核，实现个体内差异的自适应。针对个体间差异性问题，提出基于TrAdaboost的迁移学习算法，根据调整不同个体的分类器权重来适应新的被试解码特点，实现小样本下的高精度分类。本项目从理论到实验，从脑科学到工程计算，解释了事件相关电位与脑负荷的关系，提出自适应个体差异的先进算法，为提高脑机接口性能提供有力推动。