【作者】 李杰；

【导师】 黄震；

【作者基本信息】 燕山大学 ， 光学工程， 2017， 硕士

【摘要】 本文从理论以及实验两个层次对遗传以及爬山算法进行了深入的研究,同时将改进的算法运用在提高齿轮缺陷识别率上。本课题提出的利用PCA统计历代个体分布规律并重新构造侵入个体以及自适应调节步长提高爬山算法的收敛速度,对于推进遗传算法的研究具有良好的理论意义和实际应用价值。首先,本文详细的介绍了齿轮分类系统的总体模型,包括图像预处理、特征选择方法、特征降维方法、优劣齿轮的分类方法。认真研究了支持向量机算法的基本原理,并且对支持向量机训练算法中参数的选取加以讨论。其次,针对传统的遗传算法过早收敛的缺点,利用主成分分析统计历代个体分布规律,并在投影空间重新构造若干侵入个体,之后利用投影矩阵将这些侵入个体再次投射到原始空间,有效增加了个体分布范围,减小了局部收敛的可能性。然后,针对传统的爬山算法寻优多元函数时收敛速度慢的问题,本文提出了自适应调节步长的爬山算法。首先,在随机解生成阶段,约束步长采用先缩小后增大的方法,保证收敛到当前波峰,并采用解的差值作为步长生成下一个解,直到当前解优于新解;其次,在新解生成过程中检测相隔固定序号的解的差值的范数,根据范数是否满足一定精度自适应调整步长的约束范围,同时生成新的约束精度;最后,直到新的解满足最终的约束精度时算法结束。最后,用改进的爬山算法对遗传算法的寻优结果进一步进行优化,通过实验对改进的遗传、爬山算法的测试数据进行分析和整理,给出了详细的研究结论。更多还原

【Abstract】 In this paper,the mountain climbing and genetic algorithm are studied from theory and experiment two aspects,and the improved algorithm is applied to improve the gear defect recognition rate.This subject used the PCA algorithm to count the distribution regularity of each generation individual and reconstruct several invaded individual and came up with the method which improves the convergence rate of hill-climbing algorithm by using the adaptive step size,which has good theoretical significance and practical value for advancing the genetic algorithm.First of all,this paper introduces the overall model of gear classification system,including image preprocessing、feature selection methods、feature reduction methods、the classification of gear.The basic principle of SVM algorithm is seriously studied,and the selection of parameters in SVM training algorithm is discussed.Secondly,in view of the premature convergence of traditional genetic algorithm,using the PCA algorithm to count the distribution regularity of each generation individual and reconstruct several invaded individual in the projection space,then using the projection matrix to project the invaded individuals onto the original space again,which effectively increases the individual distribution range and reduces the possibility of local convergence.Then,in view of the problem of slow convergence speed of the climbing algorithm with restricted step size,which can only optimize unit function,this paper proposes a climbing algorithm based on adaptive regulation step size.First of all,in the stage of random solutions generation,generating random solutions in a given constraint range to promise restraining to the present peak of wave,and using the difference value as the step size generates the next solution,step as a solution by,until the current solution is better than new solution;Secondly,detecting the norm of difference value of a fixed sequence number solution during the process of the generating of new solutions adjusting the range of step size adaptively according to whether the norm value meet certain precision and generating new constrainted precision;Finally,the algorithm will not finishuntil the new solution meet the finial constrainted precision.Finally,the paper uses the improved hill-climbing algorithm to optimize results of genetic algorithm.The detailed study result is obtained by analyzing and sorting the experimental data of the improved hill-climbing genetic algorithm.更多还原