【作者】 黄江帅；

【导师】 王永骥；

【作者基本信息】 华中科技大学 ， 控制理论与控制工程， 2009， 硕士

【摘要】 脑机接口是一种实现大脑直接与外界环境进行沟通并进行控制的新技术。随着多通道神经元信号采集技术与计算机控制技术的日益成熟,从大脑皮层神经元群体活动中提取运动信息的解码算法是整个脑机接口系统实现脑信号与外界环境联系的关键部分。本文针对脑运动神经系统的建模与辨识问题,深入研究了从大脑运动皮层神经元脉冲序列信号中提取关于生物具体运动行为信息的解码算法。本文的实验对象是猴子运动大脑皮层脉冲数据与它的肢体运动的姿态的建模关系,对大脑运动皮层神经元脉冲序列信号的建模的工作主要是集中在对采样的各个方向运动的脉冲数据进行方向分类以及轨迹拟合。本文首先介绍了液体状态机(Liquid State Machine,LSM)这种神经网络模型。液体状态机是一种新型的回归神经网络,采取的神经元是Leaky-Integrate-and-Fire(LIF)模型神经元。由于它能够直接接受和处理脉冲,脉冲序列不再需要像频率编码的处理模式一样转变成频率序列,使得以这种方式处理能够大幅度提高建模精度。实验结果表明对于脑神经脉冲序列的处理它是一种合适的计算模型。为了提高实验的建模精度并且有效地避免液体状态机在随机生成时带来的问题,本文还用粒子群算法对液体状态机的中间回路的连接权值进行优化,以提高其分类能力,实验结果证实,采用PSO优化后的LSM的分类能力有了大幅度提高。最后针对本文中问题的独特性,本文提出了改进型的液体状态机并且应用它来对本文的实验数据进行分类与建模研究。从分类问题的分类精度和轨迹拟合可以看出本文提出的改进型更能够适用于本文的特定实验。更多还原

【Abstract】 Brain-Machine interface is a kind of new technology that makes the brain communicating with and even controlling the outside devices a reality. As the advance of the multi-channels neural signals sampling and computer control technology, the decoding of the neural activities of neurons, which are in the form of spike trains, is the key for the communication of the brain and outside devices. Aiming at the modeling of the spike trains, we take a deep look into the ways of drawing neural information from the neural signals recorded from the motor cortex. The research target of this experiment is a monkey whose brain is under some special surgery so the neural signals of the motor cortex can be recorded by some electrodes. The main purpose of the modeling of motor cortical signals is the directional classification and track fitting of the spike trains. As when the signals are recorded, the monkey is doing some arm-movement, and we just want to know the relationship between the movements—the move direction and move trajectory—and the spike trains.The computing model we adopt is the Liquid State Machine (LSM). It is a kind of recurrent neural network but the neuron model is Leaky-Integrate-and-Fire model so it can receive and deal with the spikes directly. Doesn’t like the traditional way of dealing with the spikes, the LSM never transforms the spikes anymore, so it can get a better precision relatively. The experimental result shows that LSM is a proper way to model the spike trains.Also we adopt the Particle Swarm Optimization (PSO) to the circuit of LSM and we find that the separation property of the circuit is greatly enhanced after the amelioration of the circuit with the PSO.In order to apply the LSM to our experiment and get the higher model precision, we improve the LSM in a special way and we find that after the improvement we get higher modeling precision.更多还原