State Key Laboratory for Manufacturing Systems Engineering, School of Mechanical Engineering, Xi'an Jiaotong University, Xi'an, Shaanxi, China.
Med Biol Eng Comput. 2019 Jun;57(6):1297-1311. doi: 10.1007/s11517-019-01960-9. Epub 2019 Feb 9.
Improper selection of the number and the amplitude of noise channels in noise-assisted multivariate empirical mode decomposition (NA-MEMD) would induce mode mixing and leakage in the obtained intrinsic mode functions (IMF), which would degrade the performance in applications like brain-computer interface (BCI) systems based on motor imagery. A measurement (ML-index) using no prior knowledge of the underlying components of the original signals was proposed to quantify the amount of mode mixing and leakage of IMFs. Both synthetic signals and electroencephalography (EEG) recordings from motor imagery experiments were used to test the validity. The BCI classification performance using NA-MEMD with the optimal parameters selected based on the ML-index was compared with the performance under the non-optimal parameter condition and the performance using the conventional filtering method. Test on synthetic signals demonstrated the ML-index can effectively quantify the amount of mode mixing and leakage, and help to improve the accuracy of extracting the underlying components. Test on EEG recordings showed the BCI classification performance can be significantly improved under the optimal parameter condition. This study provided a method to quantify the amount of mode mixing and leakage in IMFs and realized the optimization of the parameters associated with noise channels in NA-MEMD. Graphical abstract One of the synthetic multivariate signals comprised four components oscillating at different rates (middle column). Noise-assisted multivariate empirical mode decomposition (noise-assisted MEMD) was used to extract different components. Mode mixing issue occurred under the non-optimal parameter condition (left column). The issue was alleviated under the optimal parameter condition (right column) which can be obtained with the proposed method in this study.
在噪声辅助多元经验模态分解（NA-MEMD）中，如果噪声通道的数量和幅度选择不当，会导致所得到的固有模态函数（IMF）中出现模态混合和泄漏，从而降低基于运动想象的脑机接口（BCI）系统等应用中的性能。本文提出了一种使用原始信号的潜在成分的先验知识的测量（ML-index），以量化 IMF 的模态混合和泄漏程度。使用合成信号和运动想象实验的脑电图（EEG）记录来测试其有效性。使用基于 ML-index 选择的最优参数的 NA-MEMD 与基于非最优参数条件的性能以及使用传统滤波方法的性能进行了 BCI 分类性能比较。在合成信号上的测试表明，ML-index 可以有效地量化模态混合和泄漏的程度，并有助于提高提取潜在成分的准确性。在 EEG 记录上的测试表明，在最优参数条件下，BCI 分类性能可以显著提高。本研究提供了一种量化 IMF 中模态混合和泄漏程度的方法，并实现了 NA-MEMD 中与噪声通道相关的参数的优化。
