【作者】 唐云岚；

【导师】 陈英武；

【作者基本信息】 国防科学技术大学 ， 管理科学与工程， 2008， 博士

【摘要】 随着烟草行业卷烟工艺水平的发展,计算机辅助卷烟产品设计系统越来越受到卷烟企业的重视。该系统在综合分析历史数据和经验知识的基础上,建立符合卷烟生产实际的各类模型,并以此为指导进行产品优化设计。本文以国内某大型卷烟企业在该领域的课题研究为背景,重点研究该企业计算机辅助产品设计系统中的核心模块——卷烟产品参数优化设计模块。在研究过程中发现,卷烟产品参数优化设计是一个十分复杂的黑盒多目标优化问题,主要反映在:(1)优化对象“工艺参数”与优化目标“质量指标”之间的映射关系十分复杂,难以建立常规的数学优化模型,对于给定的一组工艺参数取值,只能通过现场实验才能获得准确的质量指标评价;(2)优化目标由相互冲突的多个目标组成,这些目标在大多数情况下不能直接进行优劣关系的比较,目标之间相互冲突,在不降低某一目标性能的情况下不能通过参数优化任意提高其他目标的性能。针对上述问题,本文提出了一种基于神经网络和多目标进化算法混合策略的集成计算智能方法:首先,利用人工神经网络对历史数据进行训练,获得能反应参数优化过程中参数向量空间到目标向量空间非线性映射关系的神经网络模型;其次,将训练好的神经网络模型嵌入到多目标进化算法中,以此作为进化过程中个体的适应度评价函数,使得多目标进化算法可以直接应用于产品参数优化设计过程。具体地,本文的主要研究内容和创新成果概括如下:1、提出了一种集成神经网络和多目标进化算法的产品参数优化设计方法现实世界中,很多产品参数优化设计问题均可归纳为黑盒多目标优化问题。黑盒多目标优化问题具有系统建模困难和多个目标必须协调优化的特点。本文通过结合跨越BP神经网络和改进的非劣解排序遗传算法(NSGA-II),借助不同智能计算方法的优点,互补不足:(1)可以充分利用跨越BP神经网络建模的优点,解决复杂系统建模困难的问题,并为NSGA-II的进化个体提供适应度评价函数;(2)采用NSGA-II解决复杂系统中的多目标优化问题。2、提出了基于跨越BP算法的人工神经网络建模方法复杂系统建模是成功解决产品设计参数优化问题的关键,然而传统BP算法具有收敛速度慢、网络结构选择困难、容易陷入局部极小等缺点。本文从连接方式、结构优化以及优化策略三个方面,对传统BP算法进行了改进:(1)采用基于跨越连接的误差反向传播算法对网络进行训练。有跨越连接的神经网络摒弃了传统神经网络只有前后层相连的拓扑结构,能以更加简洁的结构逼近神经网络的理想状态,加快网络收敛速度。国防科学技术大学研究生院博士学位论文(2)提出了一种BP神经网络结构优化算法。该算法通过引入有方向的均方误差,在有跨越连接的多层前馈人工神经网络结构方程式的基础上,分别导出隐层层数和隐层神经元数判别式。(3)采用基于MOEA和BP混合算法的神经网络建模方法。由于要维持具有一定规模的群体,多目标进化算法必须同时处理搜索空间中的若干点而不像梯度下降法那样只处理单点,从而有助于搜索全局最优点,免予陷入局部最小。这样就可以避免传统BP人工神经网络采用梯度下降法所带来的缺点,同时也确保了良好的收敛速度。3、提出了基于NN和MOEA的卷烟工艺参数优化设计方法卷烟工艺设计主要分为打叶复烤工艺设计、制丝工艺设计和辅料配套工艺设计,它们的本质都是基于各类参数指标关系模型的多目标参数优化过程,且属于黑盒多目标优化范畴。因此,可采用集成神经网络和多目标进化算法的产品参数优化设计方法(ICIA–NN & MOEA)优化求解。在具体应用过程中,结合工艺设计实际提出了基于NN和MOEA的卷烟工艺参数优化设计方法,并将其应用于二次润叶工序工艺参数多目标优化问题,取得了令人满意的效果。4、提出了基于NN和MOEA的卷烟配方参数优化设计方法卷烟配方设计主要分为叶组配方设计和糖香料配方设计,它们的本质都是基于感官质量评价模型的多目标参数优化过程,且属于黑盒多目标优化范畴。与工艺参数优化设计不同的是,配方参数优化设计涉及到感官质量评价问题,这是一个主观性较强的评价过程,难以直接建立类似于工艺参数指标关系模型的单料烟比例与感官质量指标关系模型。针对上述问题,本文提出了基于NN和MOEA的卷烟配方参数优化设计方法,该方法与工艺参数多目标优化设计方法相比,主要有两点不同:(1)借鉴卷烟配方实践中的感官质量评分标准,将感官质量评价结果转换为感官质量得分,实现了非数值型指标向数值型指标的转变;(2)以烟叶化学成分为中间环节,分别建立单料烟比例与烟叶化学成分关系式和烟叶化学成分与感官质量得分关系模型,成功实现了由单料烟比例到感官质量评价的非线性映射。最后,将基于NN和MOEA的卷烟配方参数优化设计方法应用于配方创新和配方维护,取得了令人满意的效果。更多还原

【Abstract】 With the development of the cigarette industry, the computer assistance design system of cigarette product obtains more attention by cigarette enterprises than before. This system would establish some effective cigarette models based on history data and experience, and has been taken to instruct the optimization of product design. This thesis mainly studies the parameter optimization module of computer assistance cigarette product design system. Because the cigarette parameter optimization design is an extremely complex black-box multi-objective optimization question, this paper proposed a kind of new optimization method based on a mixed strategy of neural network and multi-objective evolutionary algorithm. Specifically, the main contents and fruits of this thesis are outlined as follows:1、Research on an integrated design approach based on neural network and multi-objective evolutionary algorithm in product parameter optimization.In the real world, many product parameter optimization design questiones are the black-box multi-objective optimization question. The black-box multi-objective optimization question has the character that the system modelling is difficulty and many goals must be optimized coordinate. This thesis integrates the BP neural network and NSGA-II. First, we would make use of artificial neural networks to train historical data and establish a neural network model that can respond the non-linearity mapping relations of the parameter vector space to the goal vector space; next, in the processing of parameter optimization, we would make use of the neural network model to obtain individual fitness.2、Research on a modeling method of artificial neural networks based on the cross connection BP algorithm.The complex system modelling is the key to solve product design parameter optimization question. However, the traditional BP algorithm has some shortcomings, which includes the slow convergence rate, the network-architecture-choosing difficulty, falling into partial minimum easily and so on. This thesis has made some improvement to the traditional BP algorithmfrom in the above three aspects. First, Neural networks with any kind of connections can always be sorted as cross-connected ones. According to traditional multi-layer feed-forward neural network, we elaborated the concept of completely-fully connected neural network and then put forward a cross-connected multi-layer feed-forward neural network algorithm. It can be theoretical proved that the cross-connected neural network can reach ideal results with more concise framework compareing with the non-cross connected neural network.Next, it is difficult for us to choose the neural network structure. On the basis of the network structure equation of multi-layer feed-forward neural network with cross connection, discriminants of quantity of hidden layers and discriminants of quantity of perceptrons each layer are given. According to the discriminants, a new neural network structure optimization algorithm is proposed.Third, a novel approach, combining MOEA with BP, is presented to evolve the neural network. The multi-objective evolution algorithm can work in search space simultaneous by certain scale population, not like gradient method which only deal with one point, thus is helpful in searching the overall optimum point and ensuring the good convergence rate.3、An approach of cigarette process parameter optimization based on neural network and multi-objective evolutionary algorithmAccording to the cigarette craft practice, the parameter optimization model in cigarette product process design is given, and belongs to the black-box multi-objective optimization question. Based on neural network and multi-objective evolutionary algorithm this thesis introduced the integrated design approach to solve this problem, and proposed a approach of cigarette process parameter optimization. This approach was used to ordering-cylinder, and obtained satisfactory effect.4、An approach of cigarette formulation parameter optimization based on neural network and multi-objective evolutionary algorithmThe cigarette formulation parameter optimization is different with the process parameter optimization question for organoleptic character. The appraisal of organoleptic character is an extremely complex process, and it is difficult to get the relation model directly. Also, the organoleptic character belongs to the non-value index; the appraisal result of organoleptic character cannot use in the neural network modeling directly. In view of the above difficulties, this thesis proposed a parameter optimization method in cigarette product design with organoleptic character.First, on the basis of the organoleptic character grading standard in cigarette formulation practice, we could transform the organoleptic character appraisal result into the organoleptic character score, realized the non-value index to the value index transformation. Next, we take tobacco leaf chemical composition as the middle link, separately established the relational model of tobacco proportion with chemical composition and that of chemical composition with organoleptic character score. Then the relational model of tobacco proportion with organoleptic character was given. In this foundation, according to the common parameter optimization method of cigarette product parameter design, we proposed the cigarette formulation parameter optimization method based on neural network and multi-objective evolutionary algorithm. Finally, this approach was used to the formulation design and the formulation maintenance successfully.更多还原