基于神经网络模型的高剪切湿法造粒过程模拟和放大预测

发布时间：2018-04-01 10:14

本文选题：高剪切湿法造粒　切入点：关键参数定义　出处：《天津大学》2016年博士论文

【摘要】：造粒是一种被广泛应用于不同工业领域的单元操作。造粒过程通过联结作用将初级粉料聚集在一起生成颗粒产品。造粒过程的影响因素有很多,包括配方特性、设备几何特性、操作参数等。大多数影响因素与产品特性间呈高度非线性关系,这使得对造粒过程机理的研究变得相当复杂。目前对造粒过程的研究只是定性的描述了某一特定造粒过程所处的机理区域,以及可能出现的造粒现象,尚不能对造粒产品的特性给出定量的预测。人工神经网络模型因其优秀的信息处理能力,在模式识别、预测控制等许多不同领域取得了广泛的应用。其善于处理高度非线性关系的特点尤其适合造粒过程的建模工作。本文借助人工神经网络模型(Artificial Neural Networks)方法对高剪切湿法造粒过程建立了一种工业上普遍适用的半经验模型来对造粒产品的粒度分布参数进行定量的预测。文章的主要内容包括以下几个方面:通过对造粒过程机理研究的相关文献进行调研,以及实验观察,确定影响造粒过程的主要参数,并对这些参数进行了归纳分析。定义了总相对扫料体积(RSVtotal)、相对理论液体可用度(RTLA)、液体注入参数(LIF)以及斯托克斯粘性系数(Stv)四个描述造粒过程的关键参数。对造粒过程建立了统一的标准化描述方法,使其适用于不同的配方和设备。同时,关键参数定义法可以降低问题描述的维度,减少实验及建模过程的工作量。另外,无量纲的关键参数消除了对设备尺寸的依赖,满足了工业放大的需求。本工作在小试车间尺度到工业生产尺度等不同尺寸的设备上进行了造粒实验,记录了每批实验的配方特性、相关操作参数及产品粒度分布参数。测定了计算造粒过程关键参数所需的原材料的基本特性。计算出每批实验的关键参数值,并以此为输入对造粒过程的粒度分布参数进行预测。人工神经网络模型的建立。利用MATLAB软件,以关键参数为输入建立了造粒过程的人工神经网络模型,对造粒产品的粒度分布参数进行预测。利用隐藏节点可变的神经网络考察了网络结构对预测结果的影响,最终选取了最优的网络结构(Network topology)。对不同传递函数(transfer function)和不同训练方法(training algorithm)的效果进行了比较,并选取最优组合。利用MIV(Mean Impact Value)法对各关键参数对造粒产品粒度分布参数的影响进行了评价。针对粒度分布宽度参数预测效果稍差的现象采取了两种建立人工神经网络模型常用的方法进行优化,并对可能的原因进行了分析。最后,在不同的配方、不同的设备尺度及不同的操作条件下对模型的预测效果进行了验证。
[Abstract]:Granulation is a kind of unit operation which is widely used in different industrial fields. The granulation process aggregates primary powder together to form granulation products through the process of granulation. There are many factors influencing the granulation process, including the properties of formula, the geometric characteristics of equipment, and so on. Operating parameters, etc. Most of the influencing factors are highly nonlinear with the characteristics of the product, This makes the study of the mechanism of granulation process quite complicated. At present, the study of granulation process only qualitatively describes the mechanism region of a particular granulation process and the possible granulation phenomenon. It is not possible to give a quantitative prediction of the characteristics of granulation products. Because of its excellent information processing ability, the artificial neural network model is used in pattern recognition. Predictive control has been widely used in many different fields, such as predictive control, which is good at dealing with highly nonlinear relations. It is especially suitable for modeling the granulation process. In this paper, artificial Neural network method is used to model high shear. A semi-empirical model of wet granulation process is established to quantitatively predict the granularity distribution parameters of granulation products. The main contents of this paper are as follows: the mechanism of granulation process is analyzed. Research on the relevant literature, As well as experimental observations to determine the main parameters that affect the granulation process, These parameters are summarized and analyzed. Four key parameters are defined to describe the granulation process: total relative sweep volume, relative theoretical liquid availability, liquid injection parameter (LIFand) and Stokes viscosity coefficient (Stv). The process establishes a unified standardized description method, At the same time, the key parameter definition method can reduce the dimension of problem description and reduce the workload of experiment and modeling process. In addition, the dimensionless key parameters eliminate the dependence on the size of the equipment. The granulation experiments were carried out on equipment of different sizes from workshop scale to industrial production scale, and the formula characteristics of each batch of experiments were recorded. The basic properties of raw materials needed to calculate the key parameters of granulation process were measured. The key parameters of each batch of experiments were calculated. The parameters of granularity distribution in granulation process are predicted. The artificial neural network model is established by using MATLAB software and key parameters as input. The particle size distribution parameters of granulation products are predicted. The influence of network structure on the prediction results is investigated by using the neural network with variable hidden nodes. Finally, the optimal network structure is selected and the effects of different transfer functions and training algorithms are compared. The influence of key parameters on granularity distribution parameters of granulation products was evaluated by using MIV(Mean Impact value method. Two artificial neural systems were adopted to predict the grain size distribution parameters. The network model is optimized by common methods, The possible causes are analyzed. Finally, the prediction effect of the model is verified under different formulations, different equipment scales and different operating conditions.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TQ029

【相似文献】