基于尖峰自组织递归RBF神经网络的SVI软测量研究

发布时间：2018-01-15 23:06

本文关键词：基于尖峰自组织递归RBF神经网络的SVI软测量研究　出处：《北京工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：污泥膨胀是制约我国城市污水处理厂发展的瓶颈,由于污泥膨胀成因众多,机理复杂,各种诱发因素之间相互影响,难以建立其精确的数学模型。污泥体积指数(Sludge Volume Index,SVI)是污泥膨胀的关键参数,其值大小表征污泥沉降性能,被广泛用来描述污泥膨胀的程度。因此,为了实现SVI的实时测量,文中提出了基于尖峰自组织递归RBF(self-organizing recurrent RBF,SR-RBF)神经网络的SVI软测量模型,获得了SVI的实时预测值,实现了污泥膨胀的预测。该论文主要研究工作包括以下几点:1.基于污泥膨胀机理分析和运行数据,获得了一组适用于SVI软测量模型的辅助变量。辅助变量的选取是SVI软测量模型的关键步骤,文中通过研究污泥膨胀理论和诱发因素,分析污泥膨胀机理模型,总结出与SVI相关性较大的12个变量;同时,采用主元分析法对初步确定的12个变量分析,最终确定了SVI软测量模型的辅助变量,由Qin、BOD、COD、DO、pH、TN组成。2.针对递归RBF神经网络结构难以在线调整的问题,设计出一种尖峰自组织递归RBF神经网络。通过大脑皮层信息传递模式和生物尖峰神经元模型,提出一种结构增长修剪机制,实现了递归RBF神经网络的结构调整;同时,提出一种自适应梯度下降法对网络参数进行训练,提高了递归RBF神经网络的性能。非线性系统建模的实验结果表明:提出的自组织机制能够在线优化递归RBF神经网络结构,神经网络预测精度较高。3.建立了一种基于尖峰自组织递归RBF神经网络的SVI软测量模型,解决了SVI的在线测量问题。为了实现SVI的在线测量,将尖峰自组织递归RBF神经网络应用于设计的软测量模型,由于尖峰自组织递归RBF神经网络结构能够在线调整,并且采用二阶LM算法对神经网络参数进行调整,使得提出的SVI软测量模型精度较高,且收敛速度较快,实验结果验证了所建立SVI软测量模型的有效性。4.开发了一种SVI软测量平台。该SVI软测量平台主要包括数据库模块、登录模块、数据处理模块、模型训练及仿真模块及结果查询模块。首先,利用VS2010软件完成操作界面的设计,具备网络模型、自定义模型初始参数及查询建模结果等功能。其次,基于Matlab软件和Mysql数据库编写了后台运行程序,实现软测量模型的计算和用户信息的存储功能。最后,通过用户信息管理、数据处理、神经网络模型选择等模块间信息传输,实现SVI预测值的输出并显示,达到污泥膨胀识别可视化的目的。
[Abstract]:Sludge bulking is a bottleneck restricting the development of municipal wastewater treatment plants in China. Due to the numerous causes of sludge bulking and complex mechanism, various induced factors affect each other. Sludge volume index (Sludge Volume index) is the key parameter of sludge bulking, and its value indicates sludge settling performance. It is widely used to describe the extent of sludge bulking. Therefore, in order to achieve real-time measurement of SVI. In this paper, a SVI soft sensor model based on spike self-organizing recurrent RBF(self-organizing recurrent SR-RBF neural network is proposed. The real-time prediction value of SVI is obtained, and the prediction of sludge bulking is realized. The main research work in this paper includes the following points: 1. Based on the analysis of sludge bulking mechanism and operation data. A set of auxiliary variables suitable for SVI soft sensor model is obtained. The selection of auxiliary variables is the key step of SVI soft sensing model. The sludge bulking theory and inducing factors are studied in this paper. The model of sludge bulking mechanism was analyzed, and 12 variables with great correlation with SVI were summarized. At the same time, the auxiliary variables of SVI soft sensor model were determined by using principal component analysis method. TN composition. 2. Aiming at the problem that the recursive RBF neural network structure is difficult to adjust online. A spike self-organizing recursive RBF neural network is designed and a pruning mechanism of structural growth is proposed by means of the cerebral cortex information transfer model and the biological spike neuron model. The structure adjustment of recursive RBF neural network is realized. At the same time, an adaptive gradient descent method is proposed to train the network parameters. The experimental results of nonlinear system modeling show that the proposed self-organizing mechanism can optimize the recursive RBF neural network on line. The prediction accuracy of neural network is high. 3. A SVI soft sensor model based on spike self-organizing recursive RBF neural network is established. In order to realize the on-line measurement of SVI, the peak self-organizing recursive RBF neural network is applied to the designed soft sensor model. Because the peak self-organizing recursive RBF neural network structure can be adjusted online, and the second-order LM algorithm is used to adjust the neural network parameters, the proposed SVI soft sensor model has higher accuracy. The experimental results verify the validity of the established SVI soft sensor model. 4. A kind of SVI soft sensor platform is developed. The SVI soft sensor platform mainly includes database module. Login module, data processing module, model training and simulation module and result query module. First, the use of VS2010 software to complete the design of the operation interface, with a network model. Custom model initial parameters and query modeling results and other functions. Secondly, based on Matlab software and Mysql database to write a background running program. Finally, through the user information management, data processing, neural network model selection and other modules of information transmission, the output and display of SVI prediction value. The purpose of visualizing sludge bulking identification is achieved.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：X703;TP183

【参考文献】