【作者】 王侃枫；

【导师】 黄秀宝；

【作者基本信息】 东华大学 ， 纺织工程， 2009， 博士

【摘要】 在纺织生产过程中,纱线作为半成品,其质量对后续织造工序的效率及最终织物的质量都有很大的影响。因此,纱线的质量控制,尤其是当原料变动、品种翻改时,如何保持成纱质量稳定更是至关重要。传统的纱线质量控制方法大都是事后控制,即当纱线纺制成后再通过抽样测试才获得有关纱线质量信息。一旦发现纱线质量不合格,则已不能对这批纱线的质量提出任何控制方案了。所以,如何能在纺纱开始前预测纱线质量就成了纺纱工厂当前关心的一个问题。这一问题的实质就是建立一个能反映输入原料性能、主要工艺参数、成纱品种等诸多因素与成纱质量相互之间关系的模型用于在纺纱开始前预测纱线的质量,并通过调整原料配比方案和合理设定工艺参数等手段控制最终纱线的质量,从而避免因纱线质量问题而造成后道工序的损失。显然,这样的一个预测模型必然是非常复杂的非线性模型,而且这一模型还必须能适应纺纱生产的动态变化。对这样一个预测模型的建立,传统的数学物理和统计回归建模方法由于存在各自的局限而已无法适应,因此,必须寻求新的现代建模方法。近二十多年来,通过对生物系统及其行为特征的模拟产生了新兴学科计算智能,并已经在不少领域得到了应用,而其中的人工神经网络(ANN)、遗传算法(GA)和遗传规划(GP)等计算智能分支也开始应用于纺织领域的检测与预测等方面。这些建模方法不仅克服了传统数学物理和统计回归方法的缺点,而且可以随着生产技术水平的改进而自适应地调整模型参数从而适应纱线生产和质量的动态变化。因此,本文尝试着将这些建模方法引入到精梳毛纱质量和纺纱性能的预测建模中。本文的主要内容是在总结归纳前人对纱线质量进行数学物理与统计回归模型预测研究成果的基础上,提炼出影响成纱质量的因素,并以此作为模型的输入,再用ANN、GA、GP及其相互组合等计算智能方法进行纱线质量预测模型。全文分为引言和七章专题论述,现分别介绍如下:论文的引言部分对计算智能作了非常简明的介绍,同时提及了论文的选题背景。第一章为文献综述,重点介绍了人工神经网络、也列出了遗传算法和遗传规划等计算智能的分支在纺织质量预测、纤维和织物鉴别、织物疵点检测、服装手感和精梳毛纺织物专家系统等方面的应用研究情况,并根据前人研究工作欠缺处,提出了本文的主要研究内容。第二章回顾了前人在纱线质量和纺纱性能方面应用数学物理和统计回归模型进行预测的研究成果,在此基础上提练出纱线不匀、细节、粗节、毛粒、纱线强度及其不匀、纱线断裂伸长及纺纱过程断头等纱线质量与纺纱性能指标的影响因子,这些因子将被用作以后几章所建的预测纱线质量和纺纱性能模型的输入变量。第三章介绍了本文设计的具有多输入、单输出、一个隐层的多层感知器(MLP)的神经网络预测模型,用于预测纱线不匀(CV%)、细节、粗节、纱线强度及其不匀、断裂伸长、毛粒和断头等纱线质量和纺纱性能指标。神经网络采用Levenberg-Marquardt(LM)算法进行训练。预测结果是:纱线不匀和粗、细节预测精度较好,实测值和预测值之间相关系数的平方(R~2)分别为:0.9408,0.9713和0.8930;纱线强度及其不匀和断裂伸长等指标的预测精度不够理想,实测值和预测值之间的R~2分别为0.7930,0.8082和0.8331;而毛粒和断头率指标的预测精度则较低,实测值和预测值之间的R~2分别为0.6132和0.6670。研究结果表明,多层感知器的预测模型需要进一步优化。第四章详细叙述了本文设计的主从式多群体平行遗传算法并用于改善MLP的预测性能的情况。所建立的模型为MLP-GA模型。其中主遗传算法主要用于优化MLP的结构,即优化输入变量维数和隐层节点数;从遗传算法则主要用于优化主遗传算法中每个MLP结构对应的初始参数,即连接权值和偏置,避免因初始权值和偏置选择不当而使MLP训练过程中出现局部最优的现象。主从式多群体平行遗传算法使MLP最优结构和参数的搜索空间由领域搜索扩大到几乎整个解空间搜索,从而使MLP的最终结构和参数达到近乎全局最优解或满意解,不仅使MLP的预测性能显著提高,而且预测结果非常稳定。预测结果显示纱线不匀及粗细节的预测精度与MLP模型基本相当,分别为0.9464,0.9766和0.9177。纱线强度及其不匀、断裂伸长、毛粒和断头的预测精度(R~2)均有较大幅度改善,并分别提高到0.9404,0.9320,0.9412,0.8733和0.8977。第五章介绍了应用小波神经网络对纱线质量和纺纱性能进行预测。本文利用Morlet小波函数替代Sigmoid函数作为MLP隐层的传递函数,通过Morlet小波基的线性叠加来按拟合网络的输出函数,并由此建立起小波MLP网络(MLP-Wavelet),网络采用梯度下降算法进行训练。预测结果显示小波神经网络的预测精度远高于MLP模型;纱线不匀、细节、强度、断裂伸长等指标略高于MLP-GA模型的预测精度,但毛粒的预测精度不如MLP-GA模型,由于小波神经网络是局部接收场网络,加之隐层神经元输入权值固定,因而与MLP-GA模型相比,小波网络的学习速度和运行效率较高,但由于MLP的结构和参数未经整体优化,最终拟合结果的MLP结构和参数未必是一个全局最优解或满意解,因而,MLP-Wavelet模型的稳定性不如MLP-GA模型。第六章介绍了应用遗传规划建立纱线质量和纺纱性能的预测模型。在简单叙述了遗传规划(GP)的基本原理和方法后,即利用与前几章相同的训练样本对纱线不匀,粗、细节等指标进行了全局最优或满意的遗传程序的搜索,并由此建立了相应的工程经验公式。再应用与前几章相同的测试样本,由工程经验公式计算纱线不匀CV%、纱线细节、及纱线粗节的理论输出值,计算结果与实测值非常吻合,R~2分别高达0.9451、0.9885和0.9357。但与MLP、MLP-GA、MLP-Wavelet模型不同的是工程经验公式不仅展示了高精度的预测结果,还清晰地显现了精梳毛纱的条干质量与精梳毛条的纤维性能、纺纱工艺参数和成纱规格之间存在的非常复杂的非线性关系。第七章为总结与展望。对本文的主要贡献和存在问题作了总结,对需要进一步深入研究的问题进行了展望。更多还原

【Abstract】 The spun yarn quality has great effect on the efficiency of the sequential weaving process and the quality of the final produced fabric in textile production.The traditional method of yarn quality control is phase lag control,i.e.the yarn quality informations are only available after the yarn has been spun and tested.In this case, the mill will be incapable of doing anything to retrieve the results,if the yarn is found to be unqualified.It has become a matter of great concern in spinning mill nowadays that how to predict yarn quality before the yarn spinning process starts.The tool for resolving this problem is to find out a model which relates the yarn quality index to input material properties,spinning parameters and yarn specifications,and can be used for the prediction of yarn quality’ before the yarn to be spun.With the aid of this model,the spinner will be able to do the virtual(simulation)spinning and make the yarn quality required by adjusting the input material’s specifications and spinning parameters.Obviously,such a model as mentioned above must be the one which is characteristic of complicated non-linear and capable of self-adaptively adjusting to fitting the dynamic variation of the spinning.The traditional methods of physical/mathematical modeling and statistical regression modeling are,therefore, unsatisfied due to the limitations possessed themselves respectively.It is necessary to use the modern modeling method for efficiently predicting the yarn quality.Since the 90’s of the last century,the Computational Intelligent(CI)which originates from the simulation of biological system and its behavior feature has been employed in some areas of science and technology.Artificial Neural Network(ANN), Genetic Algorithm(GA)and Genetic Programming(GP)are 3 branches of CI and also have been reported to be applied to products inspection,process control and quality prediction in textile researches area.These modeling methods are capable of not only overcoming the disadvantages of traditional physical/mathematical models and statistical regression methods,but also self-adaptively-adjusting to fit the dynamic variation of yarn production.This thesis will apply these modeling methods to predict the worsted yarn quality.The thesis summarizes,at first,the former achievements of yarn quality prediction based on the physical/mathematical and statistical regression models,so as to draw the factors affecting the yarn quality.Then the new models for predicting the worsted yarn quality will be designed by using ANN,GA,GP and their combination. In the designed new models,the factors drawn will be taken as the input variables. The thesis consists of a preface and seven chapters.The details are as following:The preface gives a brief introduction of Computational Intelligence and the background of the research topic of this thesis to be chosen.The first chapter is presented as the literature review of the current research status of applying ANN,GA and GP to quality prediction,fiber & fabric identification, weaving defect detecting,apparel hand feeling evaluation,worsted fabric expert system and etc.In the chapter,the main research contents of the thesis are put forwarded according to the gaps of the former research work.In the second chapter,the former research achievements of yarn quality prediction based on physical/mathematical and statistical regression models are overviewed and summarized.The parameters are drawn from these models,thereby, all of which have meaningful affects to yarn unevenness,thin places,thick places, neps,yarn tenacity and its variation,yarn elongation at breakage,ends-down during spinning.These parameters will be also taken as the input variables of the models designed in the following chapters.In the third chapter,8 prediction models are built up for predicting the yarn evenness,thin places,thick places,neps,yarn tenacity and its variation,yarn elongation at breakage,ends-down during spinning respectively.The models are the construction of Multi-Layer Perceptron(MLP)with multiple inputs,single output and one hidden layer,and trained by using Levenberg-Marquardt(LM)learning algorithm. The affective parameters summed in chapter 2 are taken as the input variables of the models in this chapter.The prediction results with good accuracy for the yarn unevenness,thin places and thick places are obtained,in which the square of correlation coefficient(R~2)between the measured values and predicted values is 0.9408,0.9713 and 0.8930 respectively.The prediction results for yarn tenacity and its variation,and elongation at breakage are not so good as above,the R~2 values are 0.7930,0.8082 and 0.8331 respectively.For neps and ends-down during spinning,the R~2 values are poor,only 0.6132 and 0.6670 respectively.These results show that the models need to be further improved and optimized.In chapter 4,a detailed description of the MLP-GA prediction models is given.In the models,the designed master-slave multi-deme parallel genetic algorithm is used to improve the model’s performance,in which,the master GA algorithm is used to optimize the construction of MLP,i.e.the input variables and the number of nodes in hidden layer,and the slave GA algorithm to optimize the initial parameters of MLP, i.e.the connecting weight and bias.in order to avoid the local optimum due to the incorrect initial parameters setting.The master-slave multi-deme parallel genetic algorithm makes the searching area of MLP construction and MLP initial parameters to be expanded into global solution space,thus the MLP performance is much improved and the prediction results are stable.It is shown that the R~2 values of yarn unevenness,thin places and thick places are similar to that of MLP model,i.e.0.9464, 0.9766 and 0.9177 respectively and the R~2 values of yarn tenacity and its variation, elongation at breakage,neps and ends-down during spinning are raised up to 0.9404, 0.9320,0.9412,0.8733 and 0.8977 respectively.In chapter 5,the wavelet MLP models(MLP-Wavelet)are used to predict yarn quality and spinning performance.In these models,the Sigmoid transfer function in MLP hidden layer is replaced by Morlet wavelet basis.The model is formed by the linear addition of the Morlet wavelet basis to fit the MLP output functions.The gradient descent algorithm is applied to train MLP-Wavelet model.The prediction results are superior to MLP model.The R~2 values for yarn unevenness,thin places, yarn tenacity,and yarn elongation are slightly better than that of MLP-GA models,but the R~2 values for neps index is lower than that of MLP-GA model.They are 0.9854, 0.9758,0.9312,0.9596,0.9284,0.9624,0.8474 and 0.9094 corresponding to yarn unevenness,thin places,thick places,yarn tenacity and its variation,elongation at breakage,neps and ends-down during spinning respectively.Compared to MLP-GA model,MLP-Wavelet model shows higher training efficiency,but lack of stability.In chapter 6,the GP model is applied to predict yam quality and spinning performance.After giving a short introduction of the principle and method of GP,the same train set as that in former chapters is used to build the genetic programmer and accordant empirical engineering formula.The results predicted by using the formula for yam unevenness,thin places and thick places are very close to measured values when inputting the same test set as that in former chapters.The R~2 values reach to 0.9451,0.9885 and 0.9357 respectively.Besides that,the formula is also capable of showing clearly the complex relationship between the yarn quality,spinning performance,and input material’s properties,spinning parameters and yarn specifications.Chapter 7 gives out the conclusion and prospect.The main contributions of the paper are summed and an outline of the problems which need to be further studied is made as well.更多还原