摘要
对于船舶辐射噪声信号的盲信号分离(BSS)问题,由于常用的最小互信息(MMI)方法需要估计输出信号的高阶累积量,这对于非高斯、非平稳的船舶辐射噪声来说信号估计的精度将会降低.为此,本文验证了最大熵(ME)方法在处理此类复杂信号时能作为最佳对比函数的条件,并在此基础上用高斯混合模型来估计信号的概率分布,提高了信号概率密度估计的精度;同时在算法的迭代过程中使用自然梯度下降法代替随机梯度下降法,提高了算法的收敛速度.通过对船舶辐射噪声信号的盲分离实验,证明了此分离算法是有效的.
When sources are non-Gaussian and non-stationary such as ship-radiated noise, the blind signal separation(BSS) is not efficient because the higher-order cumulants of the sources can not be estimated accurately in Minimum Mutual Information (MMI) approach. Maximum Entropy (ME) approach was demonstrated to be a better contrast function when the p.d.f. of sources is complex. And the Gaussian mixture model was presented to estimate the p.d.f. of the sources based on this ME approach. Moreover,a natural gradient descent algorithm instead of stochastic gradient descent was introduced to speed the convergence. The experiment of separating ship-radiated noise demonstrates that the ME contrast function based on Gaussian mixture model is successful.
出处
《上海交通大学学报》
EI
CAS
CSCD
北大核心
2004年第12期1962-1965,1971,共5页
Journal of Shanghai Jiaotong University
基金
国家自然科学基金资助项目(30170274)
国防科技重点实验室基金资助项目(51444100304JW0301)
关键词
盲信号分离
最大熵
最小互信息
高斯混合模型
船舶噪声
辐射噪声
blind signal separation (BSS)
maximum entropy (ME)
minimum mutual information (MMI)
Gaussian mixture model
ships noise
radiated noise
