【作者】 陈堂功；

【导师】 汪友华；

【作者基本信息】 河北工业大学 ， 电工理论与新技术， 2006， 博士

【摘要】 电磁装置最优化问题的目标函数常常是非线性、多峰值、不可导甚至不连续,并且求解这些问题需要大量计算。连续运动带材横向磁通感应加热装置的优化问题涉及三维涡流场与温度场的耦合计算和全局优化设计问题。本文从分析遗传算法产生早熟现象的原因入手,受生物进化过程中突变和灾难等现象的启发,提出了自适应群体消亡遗传算法并对其性能进行了详细分析。首先从个体的海明距离引出反映群体相似性程度的多样性度量函数,继而导出了随群体多样性自适应变化的个体消亡个数,并随机补充相同数量的新个体以便维持群体规模、抑制早熟现象的出现。该算法提高了遗传算法的寻优能力和效率,可有效改善群体的分布特性,克服早熟现象,并减小了遗传算法对初始群体性能的依赖性。赋予精英策略的自适应群体消亡算法既具有克服早熟现象的能力,又具有较高的搜索速度和收敛率,是一种性能优异的遗传算法。为克服标准遗传算法的早熟现象,受生物基因重组现象和逆转算子的启发,提出了循环移位遗传算法并对其性能进行了详细分析。在标准遗传算法的基础上,以不同的策略和一定的概率,实施循环移位算子,使得个体的基因串在一定程度上发生类似于基因重组的现象,并借以产生新个体,为维持群体的多样性提供新的手段和途径。仿真测试结果表明,不同策略下的循环移位遗传算法的平均收敛代数均比标准遗传算法有较大的提高,说明循环移位算子可有效提高遗传算法的搜索能力,加快遗传算法的收敛速度。利用自适应群体消亡遗传算法和循环移位遗传算法完成了拍合式电磁继电器的优化设计。根据遗传算法的宏观策略和混合遗传算法的设计思想,本文提出了一种将遗传算法和粒子群算法结合组成新的混合遗传算法的新结构,即嵌入式结构,该结构根据粒子群算法的特点,将粒子群算法缩成算子形式,并与标准遗传算法结合,形成新的混合遗传算法。仿真结果表明,该算法在搜索质量和搜索效率方面明显优于其各自组成算法,并在横向磁通感应加热温度场预测神经网络权值优化设计中获得了成功的运用。在上述工作的基础上,对连续运动带材横向磁通感应加热装置的优化设计问题进行了研究。首先利用通用电磁场分析软件对连续运动带材横向磁通感应加热问题进行了数值仿真,并对线圈的形状及其尺寸与出口处热源及温度分布关系进行了对比研究。其次为了提高计算效率,采用了神经网络预测方法,建立了该装置温度场预测的神经网络模型,并利用上述仿真结果作为神经网络训练学习的样本完成了该神经网络的训练。最后建立了以频率、电流及线圈外径与带材宽度比值为设计变量的连续运动带材横向磁通感应加热问题的数学模型,并利用上述设计的循环移位遗传算法和神经网络对该问题进行了优化设计,并获得了预期的结果。更多还原

【Abstract】 The objective functions of optimal design of electromagnetic devices are often highly non-linear, abrupt, multi-hump, discontinuity or non-differential, and almost all the problems need huge computation that makes the traditional optimization techniques be incapable. The optimal design of transverse flux induction heating (TFIH) devices for continuously moving strip involves three-dimensional coupled problem of eddy field and temperature field and global optimization design problem.Based on the analysis of the reason for the premature convergence causing in genetic algorithm, and enlightened by the phenomenon of mutation and disaster in the biology evolution procession, an adaptive population disappearance genetic algorithm (APDGA) was presented and discussed. The diversity measure function was presented to describe the similar extent of population using the Hamming distance of each individual, and then educed disappear number of individual which vary from the diversity of population adaptively, the same number of new individuals are added to the population randomly to maintain the suitable population size and the diversity of population. Compared with standard genetic algorithm (SGA), APDGA can observably improve the distribution character of population, maintain population diversity, restrain premature phenomenon, decrease the dependence of SGA to the character of initial population, and has much highly convergence ratio. With elitist strategy its convergent speed has great increase. APDGA is a high performance Genetic Algorithm.Inspired by the phenomenon of genetic recombination and inversion operator, a cyclic shift genetic algorithm (CSGA) was presented and discussed with different strategy. Based on the SGA, implements cyclic shift operator with different strategy to make the gene code of individual reformed, and generates new individual. From the results of simulation, the average convergent generations of CSGA are small than that of SGA. CSGA can effectively suppress the premature phenomenon, and enhance the convergent speed and rate.APDGA and CSGA were successfully used in the optimal design of electromagnetic relay. According to the macro strategy of GA and design idea of Hybrid Genetic Algorithm (HGA), the new structure of HGA– embed structure was discussed. The structure takes the PSO as a operator, and add it to the SGA to form new hybrid genetic algorithm called HGAPSO. From the simulation results, HGAPSO has better searching quality and efficiency than that of SGA or PSO, and is applied successfully to the temperature neural network (NN) prediction in transverse flux induction heating (TFIH).The paper discussed the optimal design problem of TFIH devices for continuously moving strip finally. Firstly the mathematical simulation of the problem was done by Ansoft software, and founded the relation between the form and size of coil and the heat resource distribution at the device outlet. Secondly built the neural network model of temperature prediction to save the calculative time and difficulty of the work, and fulfilled the NN sample train by taking the simulation results as the samples. Finally founded math model of optimal design of TFIH device for continuously moving strip with three variables as frequency、current and ratio between coil outside radius and strip width, and obtained the optimal results of the math model by HGAPSO and the NN.更多还原