铝土矿浮选过程粗选矿浆pH值软测量模型及应用

【作者】 杜金芳；

【导师】 唐朝晖；

【作者基本信息】 中南大学 ， 控制科学与工程， 2011， 硕士

【摘要】 浮选是按矿物表面物理化学性质的差异来分离各种细粒矿物的一种有效方法。粗选是浮选流程的起始环节,其矿浆的pH值直接反映磨机中碳酸钠的添加量,对后面各个流程的药剂添加量、精矿及尾矿品位都有重要影响。现有的pH值测定仪昂贵、耗时、费力,且具有滞后性,结果易受环境影响,无法长期、连续、稳定、准确地进行测量。粗选矿浆pH值的影响因素很多,且其变化过程具有非线性特征,很难从化学反应机理上建立矿浆pH值的软测量模型。泡沫图像包含了大量与生产指标相关的信息,图像特征能够实时反映矿浆pH值的变化,为实现矿浆pH值的实时在线检测提供了可能。论文提取了粗选槽泡沫的颜色、尺寸、纹理等静态特征和速度、稳定度等动态特征,并结合BP神经网络逼近能力较强及RBF神经网络快速跟踪性较好的优点,提出一种基于BP和RBF的混合神经网络软测量模型。由于泡沫图像的特征变量较多且相关性强,直接用于软测量模型将导致计算量大、训练时间久等问题,故采用主元分析对其降维处理来确定软测量模型的辅助变量。同时,由于神经网络隐层节点数及软测量模型的混合系数具有不确定性,本文采用遗传算法对其寻优,并针对传统的遗传算法具有的早熟,最优解附近振荡及收敛较慢等缺点,分别从适应度值标定、维持种群多样性及交叉、变异算子三个方面对其加以改进,提出一种自适应遗传算法。此外,由于浮选现场工况具有时变性,当检测误差超过设定值时,采用阻尼最小二乘法对软测量模型中神经网络的权值和阈值进行在线修正,以提高模型的鲁棒性。现场运行结果表明,建立的软测量模型具有精度高、动态响应快且鲁棒性强的特点,可以取代离线的化学分析,从而降低劳动强度,同时对磨机中碳酸钠的加药量具有很重要的指导作用,并可推广应用到其它类似工业过程。更多还原

【Abstract】 Flotation is an effective method for separating various fine grained mineral particles based on the difference of the surface physicochemical properties. Rougher flotation is the beginning of flotation. The pH value of its slurry directly reflects sodium carbonate addition in mill. At the same time, it has significant impact on reagents addition of other processes, concentrate and tailings grade. Present pH determinator wastes time and human resource and has time delay. Furthermore, the result is easily affected by external conditions. It can not be used continuously and stably.The pH value of rougher flotation is affected by many factors and the change of pH value is a nonlinear process. It is hard to build a soft measurement model for slurry pH value based on chemical reaction mechanism. Froth image includes much information related to production indexes. Its features can reflect the changes of slurry pH value. Therefore, the real time pH value of rougher flotation slurry can be obtained by froth image features.In this paper, the static features such as color, size and texture as well as dynamic features like velocity and stability are extracted by digital image process technology. At the same time, because BP neural network has good function of approximation ability and RBF neural network can trace the function quickly, a hybrid neural network model based on BP and RBF neural network is proposed. Due to the strong correlation between the froth image features, it may lead a large amout of calculation and a long training time if the features are used in the soft measurement model directly. So the features’dimension is reduced by principal component analysis before they are used as instrumental variable of soft sensor model. Owing to the node number of hidden layer in neural network and hybrid coefficient of soft measurement model are not determined, they are optimized by genetic algorithm. Meanwhile, aiming at the disadvantage of traditional genetic algorithm such as local optimum, oscillation around the optimal value and slow convergence speed, an adaptive genetic algorithm is proposed based on the improvement of calibration of fitness, diversity of population and crossover and mutation operator. In addtion, in order to improve the robustness of the model, Levenberg-Marquardt algorithm is used to modify the weights and threshold of neural network on-line when measurement error exceeds the set value.Running results show that the soft sensor model has advantages of high precision, fast response and strong robustness. It can replace the offline manual measurement, reduce labor intensity and guide workers to adjust the sodium carbonate addition in the mill. Moreover, it can be popularized in other industrial process.更多还原