【作者】 王云峰；

【导师】 袁占亭； 李战明；

【作者基本信息】 兰州理工大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 硫化镍矿的磨矿过程是一个复杂系统,耦合性强,且各个环节间的相互影响大,具有一定的相关性。目前,在磨矿过程中,由于缺乏探测球磨机和水力旋流器内部工作状态的仪器设备,磨矿过程的规律至今仍没有被完全认识清楚,而真实磨矿过程的复杂性,很难用数学模型精确表示其整个过程的特性,甚至有许多控制策略是无模型的。本文针对磨矿过程半经验模型,对磨矿整个过程从数据采集处理和软测量技术、多变量动态过程模型辨识技术及先进控制策略等方面进行了相关的一些探索和研究,同时,针对目前控制理论和磨矿过程应用之间的脱节现象,将现代控制理论经过改造移植到选矿过程控制领域中,研究推出从具体磨矿系统的特点出发,寻求对模型要求不高,但在线计算方便,且对环境的不确定性有一定适应能力的实用控制策略和方法。近年来,随着智能技术的迅速发展,以模糊系统、神经网络等为代表的智能技术显示出了对复杂非线性工业系统强大的处理能力,一系列基于智能控制理论及方法的先进控制技术被不断提出和改进,并在对复杂工业对象的控制问题上取得了一些重大的突破和丰硕的成果。但是,由于目前的智能控制技术基础理论发展还不完善,在应用智能控制方法和技术时存在许多值得改进的地方。本文针对目前有关智能方法的先进控制技术提出一些新的参考方法和改进的应用策略,主要创新点有：1、利用RBF神经网络和磨矿过程中可测量信息来预测磨矿过程将来的过程行为,将RBF神经网络和模糊理论结合起来,通过在RBF神经网络隐层增加模糊化层和模糊规则推理层,使磨矿过程特性和参数变化对控制质量影响减小,实现预测过程的响应和设计希望的响应差别为最小,得到了一种模糊RBF网络控制模型；该模糊RBF网络经过在线动态训练,实现了磨矿分级过程多变量非线性和大滞后系统的PID智能整定控制；该分析过程相对简单,网络学习训练时间少,学习精度高,估计值与分析值拟合良好,系统鲁棒性强；仿真表明这类智能控制器可用于多变量非线性时变系统这类难以建立数学模型的控制系统。2、针对较难控制的大滞后过程对象,提出了动态分支预测转移控制策略,使控制回路在运行过程中始终保持在最佳运行状态,最终提高工业过程设备的运行效率。通过在控制过程中增加对被控对象输入输出之间相关性的跟踪及处理,在基本预测控制算法的基础上再增加一个预测控制变量协调决策层,可在线进行拟合,利用反馈校正的滚动优化策略进行记录及优化,获得了被控对象输入输出之间相关性及相应控制策略的动态分支预测转移控制表,结合系统设定值进行区间控制和约束保护措施,在暂态响应和稳态性能之间取得折衷,使控制效果得到了明显的改善,不但增强了输入控制量的规律性,而且提高了响应的快速性和准确性。仿真实验表明模型的在线辨识精确,可以保证系统的鲁棒性能和预期的控制性能。3、运用RS理论研究了某选矿厂磨矿工艺多维数据的属性约简,在建立相应的RBF神经网络预测模型的基础上,得到了表征磨矿生产过程内在规律的最小知识表达,并应用该模型对选矿生产指标进行了预测。结果表明：磨矿工艺数据可以进行浓缩；生产过程中的经验操作有了相应的理论依据,生产工艺人员或操作工人通过使用该软测量模型进行工艺分析与控制,逐步加深了对生产工艺过程内在规律的认识；应用软测量获取了球磨机和旋流器的内部状态主要关键参数；该模型分析过程相对简单,网络学习训练时间少,学习精度高,仿真结果表明估计值与分析值拟合良好。4、提出了磨矿分级系统分段功能模块并行技术,该技术首先将磨矿分级系统中的每个组成部分看作是一个功能部件模块,在建立各功能部件模块的RBF神经网络模型的基础上,确定了磨矿分级系统中对磨矿浓度和溢流粒度性能影响最大的功能部件模块；接着,建立了磨矿分级一个系列RBF神经网络模型,就如何改进磨矿分级系统模型中某一功能部件模块而获得整个磨矿分级系统性能指标的提高提出了自己的看法；最后对改进前和改进后的磨矿过程有关性能指标改善情况进行了比较验证。5、针对大型浮选设备对矿浆液面自动控制系统的高要求,设计了矿浆液面自动控制系统中的执行器,提出了一种符合现场复杂情况,基于神经元网络技术的软测量方法,在基于过程补余量算法中,引入利用不同形状的反馈凸轮片产生不同非线性来改变控制阀的原有流量特性,结合比例系数得到了具有线性流量特性的调节阀曲线函数的一种软测量模型建模和优化的过程。结合BP神经网络模型,从对控制的稳定性、控制响应的时效性等因素考虑后,确定了这种阀芯的最佳拟合曲线,该设计方法经现场的实际使用验证效果良好。更多还原

【Abstract】 Ore grinding is a complex process, the links of the process have a good correlation, for lack of instruments to measure the internal work state of ball mill and hydro cyclone, the regulation of grinding process is not completely known, it is necessary to understand and grasp the mathematics model in order to carry out the online control of grinding process.At present, internal condition of ball mill and hydro cyclone that can not be directly measured. Application of the Soft-measure Technology, key parameters about quality index of the grinding and classification system be directly determined and correlation model of soft measurement be established based on RS Theory and RBF neural network, which is helpful to production technical personnel and operation workers to technically analysis and control system of grinding and classification.As the intelligence technologies continue to develop, products of these technologies represented by fuzzy system and neural network show the powerful processing capability to complex nonlinear systems. A series of advanced control systems based on intelligent control theories and methods are continuously put forward and improved, great breakthrough and plentiful fruits are derived for the control problem of complex systems. However, due to the immaturity of the foundation theory of intelligent control, there are still many aspects and key points need to be improved when applying the control methods. In view of the combined design of the advanced intelligent control system in future, the main research work of this dissertation can be described as follows:1、A control model of fuzzy RBF neural network is presented, and the future process action of the grinding ore is predicted based on RBF neural network and available information, With two additional layers, fuzzy layer and fuzzy inference rules layer on the hidden layer in RBF neural network, it analyzes how the parameters and process characteristics of grinding ore affect controlling effect by receding horizon optimization methods and minimizing the error between model output and experimental data; A intelligent PID control method is presented based on the fuzzy RBF neural network, which realizes decoupling control of multi variable, nonlinear and time-variation by adjusting parameters of PID controller on-line; The analysis course is briefness, the time of network learning and training is little, learning precision is high, estimate value very close in upon analysis value; The simulation researches have verified the proposed approach which can be control systems where it is difficult to build accurate math model.2、A control technology on predicting dynamic branches for large dead-time processes was put forwarded to keep the control loops being optimal all time and improve efficiency of process equipments as well, by the coordination and decision level of prediction control variables, a form on predicting dynamic branches was formed by treatment of relevance of input and output on controlled objects in control software and optimal tracing of rolling optimal strategy based on feedback correction, the technology makes compromise between transient response and steady state performance. It is shown that the control effect was improved after field practical application. The technology not only emphasizes regularity of input controlled variables, but also improves the accuracy and fast of the response. Simulation example shows that the online identified model is accurate and it can guarantee both desired robustness and control performance.3、Attributes reduction of multidimensional data were given on certain ore plant by using Rough Set Theory, on the basis of RBF neural network prediction model, least knowledge expression is presented about ore production processing’s inherence rule, and the metallurgical performance was built by applying the model. The results showed that: attributes reduction of multidimensional data is applicable on grinding and classification system; the model is assistant for production personnel or operation worker when they use RBF neural network putting up analysis and control, and it make them understand ore production processing’s inherence rule, and it can provide theoretical basis for experiential operating; key parameters of ball mill and hydrocyclone were presented by soft measurement technology; the analysis course is briefness, the time of network learning and training is little, learning precision is high, estimate value very close in upon analysis value.4、Parallel segmentation module technology of grinding and classification system is proposed, Firstly, each component of the grinding and classification system is regarded as functional unit module, RBF neural network of system components was established, the most influential modules on grinding concentration and overflow particle size of the grinding and classification system were identified. Then RBF neural module of one series of the grinding and classification was established, his own opinion is proposed and improved on a functional unit of the grinding and classification so that performance index of the grinding and classification is improved. At last, the performance indexes of the grinding and classification were performed comparatively for the situations with and without the improvement.5、Based on big flotation machine that have highly request for liquid level of automation control.In this paper,we designed control valve of liquid level of automation control from the locate,and gave a soft measurement technique based on neural networks. By using different form of feedback cam, which brought different non-linear flow rate, and which influenced character of flow rate of control valve in different processing of change deal and comparison coefficient, we put out control valve’s function and it satisfied linear character of flow rate. Optimal curve fitting is determined based on replenishment quantity of flotation process and model of BP neural network by using the different shape of valve core curve. By experiment and operation on the locate,it worked well.更多还原