【作者】 徐鸿飞；

【导师】 李炜；

【作者基本信息】 兰州理工大学 ， 控制理论与控制工程， 2012， 硕士

【摘要】 丁苯橡胶是一种人工合成橡胶，因其良好的性能可在许多方面替代天然橡胶，已被广泛应用于人们日常生活及军工领域。丁苯橡胶生产的难点在于对聚合转化率的时实监控。目前，国内丁苯橡胶生产企业多采用实验室人工方法对其分析控制，该方法因存有严重的时滞问题，不仅控制效果不佳还会带来人力，物力等资源的浪费。软测量技术的发展为上述问题的解决提供了一条良好途径。然而，目前采用该技术对聚合转化率进行研究的文章还不多见。就丁苯橡胶生产而言，虽工艺复杂，但工作点相对稳定，辅助变量间存有较强的非线性，企业对聚合转化率预测也有较高精度要求。因此，针对该具体对象，本文以核函数思想为基础，提出了3种丁苯橡胶聚合转化率预测的软测量建模方法。其一是基于KPCA–LSSVM的丁苯橡胶聚合转化率软测量方法。考虑实际工况复杂性和企业对预测精度的要求，首先采用具有较强非线性特征提取能力的核主元分析（KPCA）对数据进行前期处理，并将其结果作为具有小样本、良好泛化能力最小二乘支持向量机（LSSVM）的输入，建立了丁苯橡胶聚合转化率软测量模型。其二是基于多核技术对丁苯橡胶聚合转化率软测量方法。在采用KPCA对数据进行前期处理的基础上，以此为输入建立了丁苯橡胶聚合转化率在线预测的径向基函数（RBF）神经网络模型。考虑RBF存在单一核难以全面精确描述复杂问题的缺陷，将具有时频突出局部表征能力的小波核引入，构造了高斯核及小波核的混合核预以弥补。其三是基于核函数的PLS丁苯橡胶聚合转化率软测量方法。考虑偏最小二乘（PLS）算法非线性处理能力的不足，核函数的引入可以提高其非线性处理能力，分别建立了单核和混合核函数的丁苯橡胶聚合转化率PLS预测模型。上述三种软测量建模方法经工业数据仿真研究，结果表明均能满足企业对丁苯橡胶聚合转化率预测指标的要求。KPCA用于复杂数据的前期处理可以为后续模型提供更为精准的数据信息；小波核具有更好的局部表征能力；而基于混合核函数的PLS建模方法，不但可以较好的描述对象的复杂特征，去除噪声，而且同时考虑输入输出间相关性，在对同批数据仿真结果中，基于核函数的PLS的预测精度和效果最好，更适宜建模预测。更多还原

【Abstract】 As a kind of synthetic rubber, Styrene-butadiene Rubber(SBR) has been widelyused in our daily life and military areas, because its good performances can be insteadof natural rubber in many ways. The real-time monitoring of polymerizationconversion rate is a difficulty for SBR production. At present, laboratory analysismethods are used to control the index in many domestic enterprise. As its serious delayproblem, not only the effects don’t very well but also lots of resources are waste. Thedevelopment of the soft measurement technology provides a good approach for thesolution the problems.Articles are not much on polymerization conversion rate by the technology. Justthe production of SBR, although the process is complex, its working points are relativestability, there are strong relevance between the secondary variables and the precisionprediction for polymerization conversion rate. According to the object, this paper putforward three kinds of soft sensor modeling method based on nuclear functionthoughts.First of all, a soft sensor for SBR Polymerization Conversion Rate Based onKPCA-LSSVM Model. Considering the complexity of actual working condition andthe prediction accuracy requirement of enterprise, it used kernel principal componentanalysis (KPCA), which has a strong ability of nonlinear feature extraction, to processthe data firstly, took the result as input of the least squares support vector machines(LSSVM) model, which has these characteristics, such as small sample and goodgeneralization ability, etc, and established the model for SBR polymerization conver-sion rate.Secondly, these soft-sensing issues studied based on more kernels technology forpolymerization conversion rate of SBR. The KPCA was used to process the data firstly.And the results were taken as input of the radial basis function (RBF) model. Conside-ring the defect that RBF having a single kernel was difficult to describe complexproblems accurately, the Wavelet kernel having a ability of time-frequency and localcharacterization was introduced, and a mixture kernel (MK) was constructed byGaussian and Wavelet kernels to remedy for the defect.Finally, this paper provides soft-sensing for polymerization conversion rate ofSBR based on kernel function PLS models. considering the complexity of actualworking condition and the disadvantages of partial least squares (PLS) algorithm for its nonlinear processing power, and the kernel function introduced could increase itsnonlinear processing power. Then PLS models with single or mixed kernel functionwere created separately and used to forecast the SBR polymerization conversion rate.The simulation results showed that the three kinds of models could all meet theenterprise requirements. KPCA used for complex data processing can provide moreaccurate data information for the subsequent model. Wavelet nuclear has a good localcharacterization ability. The PLS modeling method based on hybrid kernel function,not only can better description of the complex object characteristics, remove noises,but also consider the correlation between the input and output at feature extraction. Inthe same group of data simulation results, models based on the kernel function havehighest prediction accuracy and best effects. Show that these kinds of models are moresuitable for the object modeling.更多还原