时变脑信号的自适应建模与学习方法研究

Brain signals are usually nonstationary due to neuroplasticity and signal recording noises. Most existing brain-machine interface (BMI) systems use static decoders and can not model the nonstationary changes accurately, thus face difficulties in maintaining a stable performance in a long-term, which has become a serious problem in BMI applications. After deeply analyzing the nonstationary properties in brain signals, we propose to use a novel dynamic multiple-model combination approach for robust decoding from nonstationary signals. By dynamically combining the models according to current brain signals, the assembled model can adapt to changes in signals to achieve the optimal estimation. We further extend the multiple model combination framework considering the properties of brain signals, and study online model updating and learning strategies with evolution computing and brain-inspired models, to maintain the effectiveness of models. The proposed method can adaptively adjust itself along with changes in brain signals, to build reliable and stable BMI systems which are capable for long-term use, thus is of great value.

神经可塑性和环境噪声干扰导致脑神经信号具有时变性、非稳性等特点，当前脑机接口系统中大多采用静态解码模型，对脑信号的时变性无法准确建模，导致系统不能长时间稳定工作，极大限制了脑机接口系统的应用。通过对脑信号时变性成因的深刻剖析，本项目提出多模型动态自适应融合的解决思路，通过多模型间的协同工作实现对脑信号变化的自适应，使不同模型在不同脑电模式下动态融合，获得联合最优的结果；同时针对脑信号本身特点，将多模型融合框架与演化计算、类脑计算模型相结合，研究解码模型随脑信号变化的在线可塑性学习和演化方法，以保证模型长期有效。本项目研究的时变脑信号自适应建模与学习方法，能够动态地适应大脑的变化，从而有望建立高可靠性、长期稳定可用的脑机接口系统，具有重要意义。

本项目的研究目标为通过建立多模型自适应集成建模与在线学习的脑电信号解码方法，使解码模型能够动态地适应大脑的变化，从而提高脑机接口系统中脑信号解码的可靠性和长期稳定性。围绕研究目标，本项目构建了一系列多模型的在线动态集成模型和方法，实现解码器对信号变化的自动化感知和自适应调整，有效提高临床侵入式脑机接口的控制精度14%以上。依托本项目共发表论文12篇，其中IEEE汇刊论文4篇，人工智能顶级会议NeurIPS 3篇，获得国际会议最佳海报/最佳论文2次，共申请专利3项。本项目提出的自适应神经信号解码方法，作为核心方法应用于我国首个临床侵入式运动脑机接口系统，有效解决高龄志愿者神经信号弱、稳定性差情况下的鲁棒控制问题，实现了临床志愿者脑控机械臂吃油条、喝可乐、打麻将等应用，受到国内外广泛关注。