【作者】 符东旭；

【导师】 熊诗波；

【作者基本信息】 太原理工大学 ， 机械电子工程， 2008， 博士

【摘要】 跳汰选矿是物理选矿的一种重要方法。我国目前入洗原煤的50%左右是用跳汰机分选的。跳汰机作为跳汰选煤的关键设备,其控制包括分层和排料两个方面,控制效果的好坏直接决定了分选效率的高低。由于传统的控制方法是基于精确的系统数学模型的控制,适于解决线性、时不变等相对简单的问题,而对于跳汰机这样复杂的非线性、大滞后系统,建立精确的数学模型极其困难,因此传统的控制方法很难达到良好的控制效果。本文的主要研究工作包括:在详细分析了影响跳汰机分选效果的因素后,指出:尽管分层是排料的前提,但跳汰是一个连续的过程,分层和排料是同时进行的,排料的好坏直接影响着分层效果,分层和排料的共同作用决定了跳汰机分选效率的高低,跳汰机的控制必须将二者结合考虑,而目前的控制系统仍然是简单的、粗放的控制,问题的关键是没有解决跳汰床层分层状态的在线检测。分析了跳汰床层γ射线分层状态检测系统的构成及实现方法,有效解决了分层状态的在线检测问题,为跳汰机有效控制系统的建立确定了基本的、可靠的平台。通过采集分层状态典型样本数据,使用人工神经网络和支持向量机学习算法,建立了跳汰机分层、排料一体模型。提出了把跳汰机分选效果(终极目标-矸石带煤)作为目标函数,即把性能函数直接作为目标函数,而把分层和排料的众多过程变量作为输入,利用人工神经网络的分布并行处理、非线性映射、自适应学习和鲁棒容错等特性,通过实验确定了适合跳汰机特性的网络拓扑结构、隐含层数目、隐含层节点数,并用遗传算法对神经网络的权值和阈值进行了优化。研究了基于统计学习理论的支持向量机算法在跳汰机矸石段的模式识别问题,发现专门针对有限样本情况的支持向量机比人工神经网络具有更大的优越性,表现在测试误差进一步变小。对支持向量机的核函数进行了线性函数、多项式函数、径向基函数和S型函数的对比,并利用最小二乘支持向量机进行了仿真,针对此项研究,发现了径向基函数作为核函数时训练误差和测试误差均较小。由于预测控制注重的是模型功能,而不是结构形式,过程的描述可以通过简单的实验获得,不需要深入了解过程的内部机理。因此,预测控制算法改变了现代控制理论对模型结构较严格的要求,更着眼于根据功能要求,按最方便途径建立多样性模型。这一点对于跳汰机这种理论严重落后于实践的系统尤为适用。从模型算法控制、动态矩阵控制和广义预测控制这三种预测控制算法和基于神经网络的预测控制分析入手,提出了利用事先离线训练好的神经网络预测器,通过预测下一时刻的矸石带煤量实现跳汰机风阀周期预测控制的方法,并利用BP神经网络构建了矸石带煤预测器以及优化器,利用已取得的样本数据进行了仿真,结果表明,矸石带煤有了明显的降低。更多还原

【Abstract】 Ore jigging is one important method in physical preparation. About 50% of raw coal is processed in this method in China. As the key equipment on coal preparation, jig control includes the control of stratification and discharging. The separation efficiency of jig is related directly with its control results. As the conventional control is based on the precise mathematical model and used to solve problems of linear, invariant time system, it is difficult to get the satisfied effect on the complicated, nonlinear and great lag system like jig, for it is hard to find its precise mathematical model. The main work did in this paper includes:After analyzed the influencing factors of jig separation efficiency, it is pointed out that, although stratification is the precondition of discharging, but the jigging process is continuous, the stratifying and discharging is happening at the same time, i.e. the discharging good or bad is affecting the stratifying results, so it’s the combined action of stratification and discharging deciding the efficiency of jig, jig control must be considered with this combination. However the control system we are using now is simple and extensive for the online measurement of jig bed status is not solved yet. Theγray Stratification Status Measuring System is analyzed in details and it shows that the jig bed online measuring problem has been solved and it provided a basic virtual platform for the construction of jig control system. Through taking typical samples from site, the integration model of stratification and discharging is reached by using ANN and SVM.It is pointed out that the Efficiency of Jig should be the final goal function (ultimate objective is the Coal Lost In Waste), i.e., the performance function is the goal function, and many variables in process make up of the inputs of the ANN. As the ANN has the advantages of distributed parallel treating, nonlinear mapping, self-adapting learning and robust, the topology structure, the number of hidden layer and the neuron number of hidden layer are determined. The weights and threshold is optimized through inherit arithmetic.The Support Vector Machine (SVM) arithmetic based on Statistical Learning Theory is studied on pattern recognition of waste segment of jig and it is found that the SVM based on limit samples shows the better results compared with ANN in test error. After compared the SVM kernel with linear, polynomial, RBF and S functions, using simulation of least-square SVM, it is found that the training error and the test error is smaller when the RBF as the kernel.As the predict control pay more attention to the model function rather than structure form, the description of process can be got through simple test and it is unnecessary to realize the inside mechanism of the process. So the predict control algorithm changed the strict request of model structure from the modern control theory. To build multiformity model in most convenient route is established according to the function desire. This method is so suitable to the jig process, which its theory got great behind the practice.Proceeding with model algorithm, dynamic matrix control, generalized predict control and the predict control based on NN analysis, the method is offered of how to use the off-line pretrained ANN to predict the Coal Lost In Waste to achieve the predict control of air valve cycle of jig. The predictor of Coal Lost In Waste and optimizer is constructed using BP Neural Network. The results of simulation using sample data shows that the Coal Lost In Waste is lower.更多还原