【作者】 刘瑞兰；

【导师】 褚健； 苏宏业；

【作者基本信息】 浙江大学 ， 控制理论与控制工程， 2004， 博士

【摘要】 软测量技术是当前过程控制领域研究的热点之一，本论文以实际工业过程为背景，探讨了软测量技术若干问题及解决方法。主要研究工作可以概括如下几个方面， 1．提出了一种改进的模糊神经网络软测量建模方法，采用规则化的平均输出隶属度函数作为模糊基函数进行反模糊化运算；在训练网络时，部分参数采用Levenberg-Marquardt算法来训练，另一部分采用一阶梯度下降法。最后用该建模方法建立了聚合反应中熔融指数的软测量模型，并与完全基于梯度下降的模糊神经网络软测量模型进行比较。结果表明改进的模糊神经网络对初始值的选择不敏感，并且具有很好的收敛性，同时还能达到指定的预测精度，很适合工程应用。 2．提出了一阶TSK模糊神经网络的混合学习算法，算法由三部分组成：基于模糊聚类的网络初始化；基于梯度下降的规则前件的学习算法；基于部分最小二乘的规则后件的学习算法。该混合算法可以根据训练样本的分布自动确定模糊神经网络的初始值，当输入变量个数多时不会出现模糊规则数爆炸现象，训练速度快，模型精度高。仿真例子和工业应用实例均说明了该方法的有效性。 3．提出了基于混合多机理模型的软测量建模方法，混合模型分为两部分，即多个简化机理模型部分和非机理模型部分。多个简化机理模型用来提高模型的鲁棒性和外推能力；非机理模型用来补偿简化机理模型的未建模部分，该部分可以是线性模型，也可以是诸如神经网络的非线性模型。用混合建模方法建立了PTA过程4-CBA含量的软测量模型，通过实际工业数据仿真表明，该方法具有建模样本需求少，精度高，鲁棒性好，易于维护的优点。 4．使用限定记忆部分最小二乘算法在线滚动建立平均粒径的软测量模型。同时考虑到用部分最小二乘算法建模的数据样本都不多，这时如果直接抛弃老样本同样可能会遗失部分信息，针对该问题本论文提出将老样本的部分信息通过方差和均值带到模型中来的思路。工业数据仿真结果表浙江大学博士学位论文明，该方法在线建立的软测量模型精度高，很适合慢时变对象、且训练样本分布不均匀情况下的软测量建模。主导变量与过程变量之间的时序匹配是软测量技术不可缺少的组成部分。时序匹配实际上是确定主导变量相对于每个过程变量的滞后时间。提出了两种无需人为干涉而是直接使用现场采集到的数据来确定滞后时间的方法:最大相关系数法和模糊曲线法。最大相关系数法一般应用于线性和弱非线性对象;模糊曲线法既可以用于线性对象也可以应用于非线性对象。仿真和工业应用表明这两种方法都是适用的。上更多还原

【Abstract】 Soft sensor technology is one of the most important research directions in the area of process control. In this dissertation, several issues and the corresponding solutions about soft sensor technology are discussed based on the real industrial process and the main contributions are described as follows.1. A soft sensor modeling algorithm based on improved fuzzy neural network is presented. The normalized average output membership functions are defined as fuzzy basis functions for defuzzification calculation. In order to improve the property of convergence, some parameters of the fuzzy neural network are trained by Levenberg-Marquardt algorithm, and the others are trained by gradient descent algorithm. Finally, a soft sensor model of melt index in polymer reaction based on the proposed method is established, and the simulation results show that in contrast to the traditional fuzzy neural network the proposed method is not sensitive to initial parameters and possesses good convergence capability and prediction precision.2. A new hybrid learning algorithm is proposed to train the fuzzy neural network based on TSK fuzzy model. Firstly, fuzzy c-means algorithm is applied to initialize the parameters of the fuzzy neural network. Secondly, the parameters of the premise part of the fuzzy rule are learned by the gradient descent algorithm. Finally, the parameters of consequent part are learned by the partial least squares algorithm. The proposed hybrid method can automatically give appropriate initial parameters of the fuzzy neural network and prevent the fuzzy rule number from increasing for high-dimensional systems. The results of simulation and industrial application show that the hybrid learning algorithm has properties of fast convergence and high accuracy.3. A soft sensor modeling method based on hybrid model combining the simplified first principle model and data-driven model is proposed. Several simplified first principle models are used to improve performances of robustness and generalization. The data-driven model, which is either linear function or nonlinear function such as neural network, is adopted tocompensate the non-modeling part of the simplified first principle model. The proposed method is applied to develop a soft sensor for estimating 4-CBA concentration in practical PTA oxidation process. The simulation results by use of real industrial data show that the proposed method has such performances of high precision, strong robustness, easy maintenance and a few samples for modeling needed.4. The partial least squares algorithm with limited memory is applied to model the soft senor on-line to predict the average particle size for the PTA oxidation process. In general, the data window is not long in length to use partial least square algorithm to develop model, some useful information would be lost if old samples are discarded directly. An idea is proposed to introduce the useful information to the model by the variances and means of old samples. The results of simulation show that the soft sensor based on the proposed method has high precision and is suitable for time-varying system with samples which distribution is not uniform.5. Time alignment matching between the primary variables and the process variables is an important part of the soft sensor technology. In fact, the time alignment matching is used to specify parameters that determine how nominal dead times are computed during the input data time alignment process. Two practical methods based on maximum correlation coefficient and fuzzy curve are proposed to determine the dead time according to the data sampled from industrial process without any manually operation. The maximum correlation coefficient based method is suitable for linear systems or weak nonlinear systems, while the fuzzy curve based method is suitable for strong nonlinear systems. Both the simulation and the application to industry show that the two presented methods are meaningful to determine the dead times.更多还原