基于相关向量机的软测量建模技术及应用研究

发布时间：2018-03-29 20:00

本文选题：相关向量机　切入点：核参数　出处：《江南大学》2017年硕士论文

【摘要】：几乎所有的工业生产的最终目标都是获得满足要求的高质量产品,因此在保证安全生产的前提下,质量控制是生产过程的核心。软测量技术的出现虽然一定程度上克服和弥补了传感器以及离线检测的不足,但是工业生产过程要求测量数据实时和精准,逐渐对软测量的发展提出了更高的要求。因此建模方法的改进和优化算法的引入,对简化软测量模型的结构、提高软测量模型的估计精度和提高建模的速度上有重要意义。本文重点研究了两种基于核函数的建模方法—相关向量机和快速相关向量机,以及通过引入优化算法对其性能的相关改进:1、提出了一种组合核函数的软测量建模方法。为了同时得到较强的回归能力和较好的稀疏性,在对相关向量机构造一个组合核函数的同时,又构建了一个综合回归性能和稀疏性的适应度函数,并利用遗传算法优化相关向量机组合核的权系数和核参数。将该方法用于一个双酚A生产流程中裂解回收单元的建模仿真。仿真实例表明,所提方法的估计精度(0.3721)、稀疏性(41)等指标均优于一般的支持向量机组合核模型(0.7327)和GA-RVM单一核相关向量机模型(poly核0.9422、gauss核0.7571)。2、现代化流程工业设备众多、工艺复杂,影响生产过程中各环节技术指标的因素繁多,导致某些质量变量在线检测困难。因此,如何从可检测的生产操作变量中提取有效特征并快速有效的确定模型参数一直是研究的热点。针对此问题,本章利用核主成分分析对软测量模型的多特征输入变量进行特征提取,建立相关向量机软测量回归模型,鉴于核参数对于核主成分分析和相关向量机模型性能的影响,又采用HS算法对KPCA和RVM的核参数进行同时寻优,构建了一个基于HS算法优化的KPCA-RVM软测量模型。仿真结果表明,通过HS算法优化的KPCA-RVM软测量模型估计精度(0.234888)和运算速度(159.57)明显优于HS-SVM算法(0.255969)、GA-RVM算法(0.25423)和HS-RVM算法(0.254186),取得了良好的效果。3、针对应用在现代流程工业中的软测量模型需要满足数据处理量大、估计精度高、实时性强的要求,将快速相关向量机(FRVM)代替相关向量机用于建立软测量回归模型,降低了计算复杂度、减少了训练时间;同时,为了快速准确的确定快速相关向量机的核函数参数,提出了一种具有非线性音调微调概率的和改进优化变量的初始选择方法的改进和声搜索算法用于寻优FRVM的核参数。仿真结果表明,本章提出的改进方法有效的解决了和声搜索算法容易陷入局部最优的不足,并且该方法的估计精度(0.2324)和运行速度(94.76)明显优于基于线性变化PAR的HS算法(0.2815)和固定PAR的HS算法(0.2782)。
[Abstract]:The ultimate goal of almost all industrial production is to obtain high quality products that meet the requirements, thus ensuring the safety of production, Quality control is the core of the production process. Although the emergence of soft sensing technology to some extent overcomes and makes up for the shortcomings of sensor and off-line detection, the industrial production process requires real-time and accurate measurement data. Therefore, the improvement of modeling method and the introduction of optimization algorithm can simplify the structure of soft sensor model. It is very important to improve the estimation accuracy of soft sensor model and the speed of modeling. In this paper, two modeling methods based on kernel function, correlation vector machine and fast correlation vector machine, are studied. And by introducing the correlation improvement of optimization algorithm to its performance, a soft sensor modeling method of combining kernel function is proposed. In order to obtain better regression ability and better sparseness at the same time, a soft sensor modeling method of combining kernel function is proposed. At the same time, a combination kernel function is constructed for the correlation vector mechanism, and a fitness function combining regression performance and sparsity is constructed. The genetic algorithm is used to optimize the weight coefficient and kernel parameters of the combined kernel of correlation vector machine. The method is applied to modeling and simulation of the pyrolysis recovery unit in a bisphenol A production process. The simulation example shows that, The estimated accuracy of the proposed method is 0.3721, and the sparsity of the proposed method is better than that of the general support vector machine combined kernel model (0.7327) and the GA-RVM single kernel correlation vector machine model (Poly kernel 0.9422 gauss 0.7571g 路2). The modern process industrial equipment is numerous and the process is complex. There are many factors that affect the technical index of every link in the process of production, which leads to the difficulty of on-line detection of some quality variables. How to extract effective features from detectable production operation variables and determine model parameters quickly and effectively has been a hot topic. In this chapter, the kernel principal component analysis (KPCA) is used to extract the multi-feature input variables of the soft-sensor model, and the soft-sensing regression model of the correlation vector machine is established. In view of the effect of kernel parameters on the performance of the kernel PCA and the correlation-vector machine model, HS algorithm is used to simultaneously optimize the kernel parameters of KPCA and RVM, and a KPCA-RVM soft sensor model based on HS algorithm is constructed. The simulation results show that, The estimation accuracy of KPCA-RVM soft sensor model optimized by HS algorithm is 0.234888) and the speed of calculation is 159.57). It is obviously superior to HS-SVM algorithm 0.255969U GA-RVM algorithm 0.25423) and HS-RVM algorithm 0.254186.The result is good. 3. According to the requirement of soft sensor model applied in modern process industry, it is better than that of HS-SVM algorithm (0.25423) and HS-RVM algorithm (0.2541866). Meet the large amount of data processing, The fast correlation vector machine (FRVM) is used to build the soft sensor regression model instead of the correlation vector machine, which reduces the computational complexity and the training time, at the same time, the fast correlation vector machine (FRVM) is used instead of the correlation vector machine to build the soft sensor regression model. In order to determine the kernel function parameters of fast correlation vector machine quickly and accurately, An improved harmonic search algorithm with nonlinear tonal fine-tuning probability and an improved initial selection method for optimization variables is proposed to optimize the kernel parameters of FRVM. The simulation results show that, The improved method proposed in this chapter effectively solves the problem that the harmonic search algorithm is easy to fall into local optimum, and its estimation accuracy is 0.2324) and its running speed is 94.76), which is obviously superior to the HS algorithm based on linearly varying PAR (0.2815) and the HS algorithm based on fixed PAR (0.2782).
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18

【参考文献】