基于仿射传播聚类和高斯过程回归的软测量建模研究

发布时间：2018-07-24 21:47

【摘要】：实际的工业过程往往具有多个不同的工况,为准确地对过程特性进行描述,多模型建模是一种有效的软测量方法。通过对系统进行划分,将复杂系统的建模问题简化为局部建模问题,可以有效地提高软测量模型的性能。多模型建模过程中,聚类方法、建模方法以及融合方式都会对软测量模型的性能产生影响。因此本文从上述三个方面入手,以实际的工业生产过程为背景,对多模型软测量建模方法进行研究。论文的主要研究内容如下:针对具有多工况特性的复杂工业生产过程,对基于仿射传播(Affinity Propagation,AP)聚类的高斯过程回归(Gaussian Process Regression,GPR)软测量建模方法进行研究。采用AP算法对训练样本进行类别划分,得到不同工况下的子数据集,然后建立相应的GPR局部预测模型,最后通过新来样本到各子数据集聚类中心的距离计算得到各局部模型的权重,融合得到最终的预测模型。考虑到过程数据维度较高的情况,提出一种基于改进AP的多模型软测量方法。首先,采用主成分分析(Principal Component Analysis,PCA)方法和差分进化(Differential Evolution,DE)算法对AP算法进行改进,使算法可以避免冗余信息影响的同时,还可以实现参数的寻优,划分得到全局最优的子数据集;然后,建立各GPR局部预测模型;最后,对于新来的样本,利用预测方差计算其隶属于各局部模型的后验概率,以此为权重对各局部模型进行融合,得到全局预测输出。通过对两个标准数据集和污水处理过程数据进行仿真,验证了所提建模方法的有效性,对处理具有高维度特性的工业过程建模问题具有非常实用的参考价值。为解决模型性能随时间推移而老化的问题,提出一种基于增量AP的在线软测量建模方法,对软测量模型和样本数据库进行及时更新。采用AP算法对训练样本进行划分,对于新来的样本,利用即时学习(Just-In-Time Learning,JITL)结合GPR的方法建立各局部预测模型,并进行融合得到在线的预测输出;对新加入数据库的样本,用增量方法对AP算法进行改进,实现其证据矩阵的增量式更新,快速地完成对新来样本的分类和数据库的更新。通过对青霉素发酵过程数据进行建模仿真,验证了所提在线软测量方法的有效性。
[Abstract]:In order to accurately describe the characteristics of industrial processes, multi-model modeling is an effective soft sensing method. By dividing the system and simplifying the modeling problem of complex system into local modeling problem, the performance of soft sensor model can be improved effectively. In the process of multi-model modeling, clustering method, modeling method and fusion method will affect the performance of soft sensor model. Therefore, this paper starts with the above three aspects, taking the actual industrial production process as the background, and studies the multi-model soft sensor modeling method. The main contents of this paper are as follows: for the complex industrial production processes with multi-working conditions, the soft sensing modeling method of Gao Si process regression (Gaussian Process Regeneration based on affine propagation (AP) clustering is studied. The AP algorithm is used to classify the training samples, and the sub-data sets under different working conditions are obtained, and then the corresponding GPR local prediction model is established. Finally, the weight of each local model is obtained by calculating the distance from the new sample to the center of each sub-data cluster, and the final prediction model is obtained by fusion. Considering the high dimension of process data, a multi-model soft sensor method based on improved AP is proposed. Firstly, the principal component analysis (Principal Component) method and the Differential evolution (DE) algorithm are used to improve the AP algorithm, so that the algorithm can not only avoid the influence of redundant information, but also realize the optimization of parameters and partition the global optimal data set. Then, the local prediction models of each GPR are established. Finally, for the new samples, the posterior probability of each local model is calculated by using the prediction variance, which is used as the weight to fuse the local models to get the global prediction output. Through the simulation of two standard data sets and sewage treatment process data, the validity of the proposed modeling method is verified, and it has a very practical reference value for dealing with the industrial process modeling problems with high dimensional characteristics. In order to solve the problem of model performance aging with time, an on-line soft-sensor modeling method based on incremental AP is proposed, which updates the soft-sensor model and sample database in time. The AP algorithm is used to divide the training samples. For the new samples, the local prediction models are established by using the method of Just-In-Time learning (JITL) combined with GPR, and the online prediction output is obtained by fusion, and the new samples are added to the database. The incremental method is used to improve the AP algorithm to realize the incremental updating of its evidence matrix and to quickly complete the classification of the new samples and the updating of the database. Through modeling and simulation of penicillin fermentation process data, the effectiveness of the proposed online soft sensing method is verified.
【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】：TP274

【参考文献】