结合改进粒子群的非线性盲源分离方法研究  

Nonlinear Blind Source Separation Combining with Improved Particle Swarm Optimization

作　　者：陆建涛 成玮[1] 訾艳阳[1] 何正嘉[1] 

机构地区：[1]西安交通大学机械制造系统工程国家重点实验室,西安710049

出　　处：《西安交通大学学报》2016年第6期15-22,共8页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(51305329);中国博士后科学基金资助项目(2013M532032;2014T70911);教育部高等学校博士学科点专项科研基金资助项目(20130201120040)

摘　　要：针对传统非线性盲源分离(NBSS)算法容易陷入局部最优解从而导致分解精度较低的问题,提出一种基于改进粒子群优化(PSO)的NBSS算法。该方法利用多层感知机(MLP)拟合非线性混合的逆过程,并将分离信号的互信息最小作为优化目标(PSO的适应度),从而实现MLP中参数的优化。然而,标准PSO算法存在粒子早熟从而使待优化问题陷入局部最优解,针对这一问题,对适应度低的一部分粒子进行依概率的杂交和变异,使粒子群体在整个迭代过程中保持多样性,从而有效解决标准PSO算法的粒子早熟问题。仿真和试验结果表明,相比于线性盲源分离算法和基于标准PSO的NBSS算法,提出的算法可以从非线性混合机械信息中提取纯净的独立源信息,并且提高了非线性混合源的分离精度,为机械系统的监测诊断和振动噪声溯源提供科学依据和关键技术。The traditional nonlinear blind source separation（NBSS）algorithms often fall across the problem of local optimal solution to lead a lower separation precision.An NBSS algorithm based on improved particle swarm optimization（PSO）is proposed,where the multilayer perception（MLP）is used to fit the inverse of the nonlinear mixed process,and the mutual information between separated signals is regarded as the optimization objective（Fitness function of PSO）to realize the optimization of parameters in MLP. However,the canonical PSO algorithms usually suffer from particle premature problems and are easy to get into local optimal solution.Thus crossover and mutation operations are applied to the particles with lower fitness according to probability mechanism to efficiently increase the diversity of the particles,and the premature problem of canonical PSO is solved.The simulations and experiments show that compared with the linear blind source separation algorithm and the NBSS algorithm based on canonical PSO,the proposed algorithm enables to extract pure independent source information from mechanical information with nonlinear mixing and improve the separation precision of nonlinear mixed signals.

关 键 词：非线性盲源分离 粒子群优化 粒子早熟 交叉变异 

分 类 号：TH17[机械工程—机械制造及自动化]