Abstract: In order to extract meaningful information from noise-contaminated Electroencephalogram (EEG) signals with the characteristics of non-stationarity, non-linearity and low signal-to-noise ratio (SNR), a new EEG de-noising method is proposed based on nonlinear multiscale representation in this paper. First, the singularity locations of EEG signal are detected. Secondly, EEG signal is processed by the nonlinear mulitscale representation (NMR) algorithm which uses nonlinear prediction operator constructed by polynomial cell-average interpolation in the vicinity of the intervals containing singularities while adopts linear prediction operator in other intervals. Next, the de-noised signal is obtained by reconstructing transform coefficients which are processed by threshold value at each scale．The efficiency of the proposed approach has been demonstrated by comparison with Garrote threshold, wavelet transform using hard threshold, soft threshold and adaptive threshold on both synthetic data and real BCI Competition IV Data Set 1. Experimental results show that this algorithm has a certain practicality and can be used to eliminate the noise of EEG signal in the brain-computer interface (BCI) system.

Key words: EEG, nonlinear multiscale representation, BCI, wavelet transform

摘要： 为了有效地从混有噪声的非平稳、非线性、低信噪比的脑电信号中抽取出有用信息，提出了一种新的基于非线性多尺度表示的脑电信号消噪方法。首先检测出脑电信号奇异点的位置，其次在跳跃奇异点的附近区间采用多项式单元平均插值构成的非线性预测算子，而在其他区间采用线性预测算子，对脑电信号进行非线性多尺度表示，然后在各个尺度上对变换系数进行阈值处理，重构处理后的系数得到去噪后的脑电信号。采用仿真数据和实际脑电数据BCI Competition IV dataset 1对所提方法进行实验测试，并与其他现有方法进行比较分析。实验结果表明，所提方法的消噪效果优于Garrote阈值、小波硬阈值、软阈值、自适应阈值，具备一定的实用性，可用于脑-机接口系统中脑电信号消噪。

关键词: 脑电信号, 非线性多尺度表示, 脑-机接口, 小波变换