前馈神经网络结构设计研究及其复杂化工过程建模应用

发布时间：2017-12-28 22:24

本文关键词：前馈神经网络结构设计研究及其复杂化工过程建模应用　出处：《北京化工大学》2016年博士论文　论文类型：学位论文

【摘要】：流程工业的飞快发展使得该行业呈现出一些特点：生产过程高非线性、相关变量多维性、工艺高度复杂和流程高度综合等,因此导致采用第一原理即机理分析的方法用于过程建模越来越棘手。近些年,先进的传感技术使得过程数据越来越容易被采集和存储,这些过程数据背后蕴含着重要的过程知识,所以基于数据驱动策略的方法在解决复杂流程工业过程建模的问题中发挥着越来越重要的作用。众多数据驱动的方法中,神经网络技术由于其具有学习、并行计算以及强非线性映射的能力已被广泛成功地应用到建模、控制、优化等诸多领域。前馈神经网络因其结构简单和算法易用等特点受到越来越多的关注。然而传统前馈网络模型并不能够满足复杂流程工业过程建模的要求,由此研究建立性能优越的前馈网络模型对丰富神经网络模型和进一步推动神经网络技术应用于复杂流程工业过程建模具有重大的意义。本文主要从递阶结构设计和双并行结构设计对递阶神经网络和极限学习机两种前馈网络展开研究,最终将其应用于复杂化工过程建模。递阶神经网络是一种善于处理高维数据的网络模型,然而其子网结构设计一直是个难点。极限学习机为近些年来机器学习领域的研究的热点之一,该模型具有快速的学习速度和较好的泛化性能。然而面对带有噪声、共线性等特点的过程数据,极限学习机模型仍存在一些问题：1、噪声处理性能低；2、传统三层结构限制模型性能;3、共线性数据对性能影响大。本文逐一解决上述问题,旨在为复杂化工过程建模特定问题下提供可靠的网络模型,最终取得一些研究成果总结如下：(1)针对递阶神经网络子网设计困难的问题,提出一种基于输入属性空间划分的子网设计方法,进而建立基于输入属性空间划分的递阶神经网络,为多参数输入的复杂化工过程提供可靠的模型。该子网设计方法避免繁琐的专家知识,首先采用先进的可拓聚类算法对输入属性的高维空间进行聚类；然后依据输入属性空间的聚类结果确定递阶神经网络的子网个数；最后依据每个子属性空间的输入属性确定子网的输入。该设计方法能够同时解决子网数目确定以及子网输入属性选取的两个难题,从而提供一种简单有效设计递阶神经网络子网的方法。(2)针对极限学习机处理噪声性能低的问题,提出一种具有递阶结构的极限学习机模型。所提出的递阶极限学习机模型中,原输入变量没有直接作为模型的输入,而是先输入到自联想滤波子网,一方面去除噪声,另一方面对多维输入空间实现降维；随后将自联想滤波子网的隐含层输出数据作为极限学习机的输入,进而有效地避免噪声对模型精度的影响。采用带有噪声的工业数据对模型进行测试,仿真结果验证了该模型的有效性和可行性。(3)针对极限学习机三层网络结构的限制问题,采用基于双并行结构的设计方法增强网络性能。双并行结构能够较好解决极限学习机中的结构限制问题,但会带来另外两个问题：1、增加极限学习机模型复杂性；2、增加共线性信息。为解决第一个问题,通过研究双并行网络结构以及皮尔逊相关系数,提出一种输入输出皮尔逊相关系数导向的双并行极限学习机模型。该模型中利用输入输出属性间的相关系数将输入属性分为正、负两种属性,随后建立正、负属性相互独立的双并行结构。工业数据仿真结果表明与传统的双并行极限学习机以及极限学习机模型相比,所提出的改进双并行极限学习机模型参数数目少、响应速度快等特点。(4)针对极限学习机不能较好处理双并行结构中共线性数据的问题,提出一种基于偏最小二乘学习的稳健双并行极限学习机模型。该模型采用偏最小二乘学习算法代替原来广义逆的学习方法来求取输出权值。偏最小二乘算法一方面能够有效的剔除原输入数据间以及隐含层节点输出数据间的共线性信息,另一方面通过选取隐含变量数目有效避免了隐含层节点数目选取的难题。实验仿真结果表明该模型相对其他模型具有鲁棒性强和泛化性能稳定的特点,为复杂化工过程建模提供可靠模型。
[Abstract]:The rapid development of process industry, the industry is showing some characteristics: the production process of highly nonlinear and multidimensional related variables and process flow is highly complex and highly integrated, thus adopting the analysis method of the first principle is the mechanism for process modeling is more and more difficult. In recent years, advanced sensor technology makes the process more easily by data acquisition and storage process behind these data contains important process knowledge, so based on data driven method in solving complex problems in the modeling process and play an increasingly important role. In many data driven methods, neural network technology has been widely applied to many fields, such as modeling, control, optimization and so on, because of its ability of learning, parallel computing and strong nonlinear mapping. Feedforward neural network has attracted more and more attention because of its simple structure and easy to use algorithm. However, the traditional feedforward network model can not satisfy the requirements of complex industrial process modeling, this study established a feedforward network model to enrich the superior performance of the neural network model and further promote the application of neural network technology has great significance in the process of complex industrial process modeling. In this paper, two feedforward networks, hierarchical neural network and extreme learning machine, are studied from the perspective of hierarchical structure design and dual parallel structure design. Finally, it is applied to modeling complex chemical processes. Hierarchical neural network is a network model that is good at processing high dimensional data. However, the design of its subnet structure is always a difficult problem. The limit learning machine is one of the hot topics in the field of machine learning in recent years. This model has fast learning speed and good generalization performance. However, in the face of the process data with noise and collinearity, there are still some problems in the extreme learning machine model: 1, the noise processing performance is low; 2, the performance of the traditional three level structure constraint model; 3, the collinear data have great influence on the performance. We solve the above problems, in order to provide a reliable model for complex chemical process modeling on specific issues, the final results are summarized as follows: (1) aiming at the difficult problem of the design of hierarchical neural network subnet, proposed a partition of the input attribute space subnet design method based on neural network, and establish a hierarchical partition of the input based on the attribute space, provide a reliable model for complex chemical process input parameters. The subnet design method of expert knowledge to avoid tedious, first uses the advanced extension clustering algorithm on the input attributes of high dimensional space for clustering; then according to the number of sub network hierarchical neural network to determine the clustering results of input attribute space; finally the subnet input is determined based on the input attributes of each sub attribute space. The design method can solve the two difficult problems of subnet number determination and subnet input attribute selection simultaneously, so as to provide a simple and effective method for designing hierarchical neural network subnet. (2) to solve the problem of low noise performance in extreme learning machine, a model of limit learning machine with hierarchical structure is proposed. The proposed hierarchical extreme learning machine model, the original input variables not directly as input of the model, but the first input to the associative filter network, a noise removal, on the other hand, dimensionality reduction of multidimensional input space; then from the hidden layer output data as extreme learning machine input subnet filtering Lenovo then, effectively avoid the influence of noise on the accuracy of the model. The model is tested by the industrial data with noise, and the simulation results verify the validity and feasibility of the model. (3) aiming at the limitation of the three layer network structure of the limit learning machine, the design method based on double parallel structure is adopted to enhance the network performance. The dual parallel structure can solve the structural constraint problem in the extreme learning machine, but it will bring two other problems: 1, increase the complexity of the extreme learning machine model; 2, increase the collinear information. To solve the first problem, by studying the double parallel network structure and Pearson correlation coefficient, we propose a dual parallel extreme learning machine model based on input and output Pearson correlation coefficient. In this model, the input attribute is divided into two attributes, positive and negative, by using the correlation coefficient between input and output attributes, and then a double parallel structure with independent positive and negative attributes is established. The industrial data simulation results show that compared with the traditional dual parallel extreme learning machine and extreme learning machine model, the proposed dual parallel extreme learning machine model has fewer parameters and faster response speed. (4) in view of the problem that the extreme learning machine can not deal with the problem of CO linear data in dual parallel structure, a robust dual parallel extreme learning machine model based on partial least squares learning is proposed. The model uses the partial least squares learning algorithm to replace the original generalized inverse learning method to obtain the output weights. On the one hand, partial least squares algorithm can effectively eliminate the collinearity information between the original input data and the hidden layer node output data. On the other hand, by selecting the number of hidden variables, it effectively avoids the problem of selecting the number of hidden layer nodes. The experimental simulation results show that the model has the characteristics of strong robustness and stable generalization performance compared with other models, and provides a reliable model for complex chemical process modeling.
【学位授予单位】：北京化工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP18

【参考文献】