基于K近邻的运动想象分类中的噪声效益

Noise Benefits in Motor Imagery Classification Using K-nearest neighbor

关于脑电信号中的噪声处理问题一直是脑-机接口(BCI)领域的重点研究方向,通常认为噪声是有害的,所以针对脑电信号中的噪声处理往往以降噪或消噪为主。但是根据随机共振(SR)的思想,在非线性系统中噪声往往能增强信号处理,而脑电信号恰好具有非线性的特征,因此提出运用高斯噪声提高运动想象脑电信号的识别率。通过在脑电信号中加入独立的高斯噪声,将原始训练集与添加噪声的训练集串联起来增加训练样本量,考虑训练样本量增加与否和噪声加入的阶段(训练或/和测试);通过共空间模式(CSP)和小波包变换(WPT)提取分类特征,并用K近邻(KNN)算法进行分类。实验结果表明,只要加入适当强度的噪声,均可提高系统的分类准确率,出现随机共振现象;增加训练样本量的同时在训练集和测试集中加入适当强度相同的噪声,系统最大平均分类准确率相比不加噪声时增加9.28个百分点;K近邻算法的最大平均分类准确率相比决策树(DT)和支持向量机(SVM)而言整体更高,体现出K近邻算法的优越性和可靠性。

The problem of noise processing in EEG signal has always been a key research direction in the field of brain-computer interface(BCI),and it is usually considered that noise is harmful,so the noise processing in EEG signal often focus on noise reduction or cancellation.However,according to the principle of stochastic resonance(SR),noise can enhance signal processing in nonlinear systems,and EEG signals have nonlinear characteristics.Therefore,Gaussian noise is used to improve the recognition rate of motor imaginary EEG signals.Adding independent Gaussian noise to EEG signals,the original training set connects with the noise-added training set to increase the training sample size.Considering whether training sample size increases or not and the stage of noise addition(training or/and testing).The classification features are extracted by common spatial pattern(CSP)and wavelet packet transform(WPT),and classified by K-nearest neighbor(KNN).The experimental results show that the accuracy of classification system can be improved and appear stochastic resonance as long as add the appropriate noise intensity.When increasing the training sample size and adding the appropriate noise of the same intensity to the training set and test set,the maximum average classification accuracy of system is 9.28 percentage points higher than that without noise.The maximum average classification accuracy of K-nearest neighbor is higher than that of decision tree(DT)and support vector machine(SVM),which reflects the superiority and reliability of K-nearest neighbor.