【作者】 周开武；

【导师】 杨慧中；

【作者基本信息】 江南大学 ， 控制理论与控制工程， 2009， 硕士

【摘要】 贝叶斯学习理论使用概率表示各种形式的知识和不确定性,并通过概率规则来实现学习和推理过程,是处理不确定信息的有力工具。本文在学习贝叶斯方法的理论及其应用基础上,详细讨论了贝叶斯方法在数据分类中的应用以及贝叶斯方法在化工软测量建模的应用。本文就以上主要内容进行了深入的研究并取得了以下结果:(1)软测量建模问题中为了提高模型的估计精度,通常需要将原始数据集分类,以构造多个子模型。本文利用朴素贝叶斯分类器简单高效的优点,首先对连续的类变量进行类别范围划分,然后用概率论中的“3σ”规则对连续的属性变量离散。为了消除训练样本中干扰数据的影响,利用遗传算法从训练样本集中优选样本。对连续变量的离散和样本的优选作为对数据的预处理,最后由预处理后的训练样本构建贝叶斯分类器。通过对UCI数据集和双酚A生产过程在线监测数据集的实验仿真,实验结果表明:基于遗传算法优选样本集的“3σ”规则朴素贝叶斯分类方法比其它方法有更高的分类精度。(2)将贝叶斯网络应用于化工软测量建模。在综合考虑生产过程工艺机理的基础上利用领域专家知识构建网络模型,采用加权联合高斯分布函数来近似表达贝叶斯网络模型中的联合概率分布,并给出了贝叶斯网络估计公式。对某企业双酚A生产装置在线采集的数据进行建模,离线估计取得了较好的效果。与支持向量机方法相比,在估计精度相当的情况下,省去了许多过程参数的估计,因此也是一种有效的软测量建模方法。(3)为了改善软测量模型的估计精度,提出了一种基于贝叶斯分类算法和关联向量机的多模型软测量建模方法。采用贝叶斯分类器对样本数据集进行分类,并对不同类别的输入数据分别建立关联向量回归机子模型,用“切换开关”方式组合作为最终的软测量模型输出。将该方法应用于双酚A生产过程的质量指标软测量建模,仿真结果表明:与单模型支持向量机相比,该方法估计精度较高,具有一定的应用价值。更多还原

【Abstract】 Bayesian Learning Theory represents various knowledge and uncertainty with probability. The learning and inference are realized by probabilistic rules. Therefore, it is a strong tool dealing with uncertain information. This thesis mainly studies the basic point, application of Bayesian Learning Theory. The application of data classification and the soft sensor model based on Bayesian method are mainly studied as the key problem.This dissertation concentrated on the research work listed below and achieved some creative results.(1) Constructing sub-models can increase estimation accuracy in soft sensing modeling, and the construction of multi-model is based on the classification of the original data set. Among the methods of data classification, Naive Bayesian classifier has been widely applied because of its simplicity and efficiency. The continuous class variables are firstly divided into several categories, then the "3σ" rule based on probability theory is proposed to discretize the attributes. In order to eliminate the interferences from the training sample, the optimal sub sample set is selected from the training sample set by genetic algorithm. Finally the preprocessed training sample is used to build the Bayesian classifier. Both UCI data sets and the on-line monitoring data sets from the process of production for Bisphenol-A (BPA) are made experiment, and the simulation results show that it is possible to reliably improve the naive Bayesian classifier by using data discretization and selected as part of data pre-processing.(2) A new approach based on Bayesian network applied to chemical soft sensor is proposed. The network model is based on knowledge of the field experts and the mechanism of process, and a weighted combination of several normal distribution functions is used to approximate the joint probability distribution in Bayesian network, and then the estimated formula for Bayesian network is been given. The parameters of the model are estimated by processing real time data from a productive plant for Bisphenol A, and the model based on Bayesian network shows good results. Compared with support vector machine, the Bayesian network saves a lot of the estimated process parameters and has considerable accuracy. It is an effective method for soft sensor modeling.(3) In order to improve the estimation accuracy of the soft sensor model, a new nonlinear multi-modeling method based on Bayesian classify algorithm and relevance vector machine is proposed in the paper. The algorithm classifies the inputs by Bayesian classifier, and then trains each class by different relevance vector regression machines, and obtains the final result by the“Switch”way. The proposed algorithm is used for a soft sensor model for the bisphenol-A productive process. The experimental results indicate the proposed algorithm is superior compared with the single model of SVM and has certain application value.更多还原