非线性系统即时学习建模方法研究

发布时间：2018-10-12 14:50

【摘要】：流程工业系统具有非线性、强烈时变性和不确定性,单一的全局建模方法往往无法满足系统建模与优化的需求。基于分而治之思想的局部建模为解决这一问题提供了一种有效策略。本文结合信息熵与样本的时段特性,提出一种即时学习建模算法。具体工作归纳如下:相似度准则决定了即时学习建模的精度,然而,常用的相似度准则仅考虑样本之间的相似性,而没有考虑输入变量与输出变量之间的相关性,从而影响了预测精度。因此,本文结合信息熵提出一种改进相似性度量准则的即时学习建模算法。它利用互信息评估输入变量与输出变量之间相关程度,构建潜在建模空间,并在此潜在空间中定义相似度指标。数值仿真和在青霉素发酵过程中的应用表明,与传统的即时学习方法相比,本文所提方法的预测精度明显提高。在流程工业中,过程变量的特征往往随着时间的变化而发生变化,亦即:具有明显的时间特性。而传统的即时学习方法采用全局搜索策略,忽略了数据的时段特性。因此,本文在相似度准则中融入时间信息,提出一种时变系统即时学习建模算法。首先采用聚类算法将数据库样本划分为相应的几个时段,其中聚类使用的相似度准则融入时间信息;然后使用分层搜索策略寻找局部建模样本。在此基础上,提出局部模型更新策略,利用偏移补偿算法矫正模型输出,降低即时学习的在线计算量。最后,数值仿真验证所提算法的有效性。
[Abstract]:Process industry systems are nonlinear, highly time-varying and uncertain. A single global modeling method is often unable to meet the requirements of system modeling and optimization. Local modeling based on divide-and-conquer provides an effective strategy to solve this problem. In this paper, an instant learning modeling algorithm is proposed based on the information entropy and the time characteristics of samples. The specific work is summarized as follows: similarity criterion determines the accuracy of real-time learning modeling. However, the commonly used similarity criteria only consider the similarity between samples, but not the correlation between input variables and output variables. Thus, the prediction accuracy is affected. Therefore, this paper proposes an improved real-time learning modeling algorithm based on information entropy. It uses mutual information to evaluate the correlation between input variable and output variable, constructs the potential modeling space, and defines the similarity index in the potential space. Numerical simulation and application in penicillin fermentation show that the prediction accuracy of the proposed method is much higher than that of the traditional real-time learning method. In the process industry, the characteristics of process variables often change with the change of time, that is, they have obvious time characteristics. The traditional real-time learning method adopts global search strategy and neglects the time characteristic of data. Therefore, this paper presents a real-time learning modeling algorithm for time-varying systems by incorporating time information into similarity criteria. Firstly, the database samples are divided into several periods by clustering algorithm, in which the similarity criterion used in clustering is incorporated into time information, and then the local modeling samples are found by hierarchical search strategy. On this basis, the local model updating strategy is proposed, and the offset compensation algorithm is used to correct the model output, which reduces the on-line computation of real-time learning. Finally, the effectiveness of the proposed algorithm is verified by numerical simulation.
【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP301.6

【参考文献】