【作者】 任会峰；

【导师】 阳春华；

【作者基本信息】 中南大学 ， 控制科学与工程， 2012， 博士

【摘要】 铝土矿浮选是“选矿-拜耳法”氧化铝生产工艺中的一个重要工序。浮选过程中,矿浆pH值能够直接反映生产工况和碳酸钠药剂的消耗量,而浮选矿浆具有流动性差、温度高、强腐蚀性等特点,容易堵塞接触式的pH值检测装置。通常选矿厂主要采用离线化验的方式监测pH值,滞后一个小时,且不能连续监测,难以指导生产操作。实际生产过程中,泡沫图像特征是矿浆pH值的重要指示器,为此,研究基于泡沫图像的铝土矿浮选pH值实时检测具有重要的理论意义和实际应用价值。论文分析铝土矿浮选工艺机理的基础上,依据泡沫图像与浮选矿浆pH值的关联性,提出泡沫颜色、形态和纹理特征提取方法,建立基于泡沫图像特征与多工况子模型组合的pH软测量模型,实现铝土矿浮选pH值的实时检测,为浮选过程优化控制创造条件。论文主要研究工作及创新性成果如下：(1)基于铝土矿浮选pH中和过程的复杂反应机理,从铝土矿微粒可浮性、浮选药剂活性和矿浆中离子组成三个方面分析pH值对浮选工况、指标的重要影响。引入理想混合连续搅拌釜式反应器的分析思想,从宏观和微观两个方面,基于质量守恒定律和电荷平衡守则研究pH值与调整碳酸钠加入量的相关性,以及浮选pH中和过程的非线性特性。通过分析大量不同pH值条件下的浮选泡沫图像,确定主要的泡沫图像特征,从浮选机理上证明泡沫图像是pH值的指示器。(2)针对铝土矿浮选泡沫特征提取难点,研究泡沫颜色、形态和纹理特征提取方法。采用全变差修复模型去除泡沫顶部亮点,通过色彩空间变换实现设备无关性的泡沫颜色鲁棒计算,有效解决泡沫顶部亮点干扰及设备老化导致泡沫颜色失真问题。提出自适应粒子群优选谷底检测阂值的形态特征提取方法,该方法将粒子群算法的惯性权重和加速因子设置为全局最优点适应度的函数优选谷底检测阈值,采用局部灰度极小选择边界检测模板,根据多角度逻辑规则比较分割泡沫图像,获取气泡形态特征,该方法能够避免矿化气泡形态各异,现场光照不均对泡沫图像分割产生的不利影响。针对表层泡沫沿前进方向径向拉长的特点,在灰度共生矩阵上增加方向性差异,通过多角度融合计算空间灰度共生矩阵的二阶统计量描述泡沫纹理,很大程度上提高了纹理计算的准确性,为建立pH值的软测量模型提供数据基础。(3)针对pH中和过程的非线性特点,提出一种代价约束的稀疏多核最小二乘支持向量回归机方法,建立浮选pH值局部软测量模型。将多核最小二乘支持向量回归机的原始优化问题转化为二阶锥规划形式,定义核函数代价因子约束复杂核函数的权重,以确定最优的组合核函数,这种对不同特性核函数进行凸组合的方法,提高了模型的学习能力和泛化能力,同时,采用改进最近邻山峰聚类和施密特正交化方法约简核矩阵,有效降低计算复杂度,避免多核学习中核函数的代价问题,再根据支持向量的数目以及活动核函数的类型评估多核学习的总代价,减少变量存储空间和计算时间。(4)针对局部模型难以反映整个浮选工况的问题,提出一种多工况子模型模糊加权组合的全局建模方法。采用特征加权模糊多类分类支持向量机划分浮选工况,在每个工况区间分别建立泡沫图像与pH值的关系子模型。根据泡沫图像不同特征对工况反应的灵敏度差异,提出基于信息增益的泡沫图像特征重要度分配方法,利用泡沫图像对工况的模糊隶属度作为子模型的权重组合多个子模型实现浮选过程全局建模,这种基于泡沫图像的集成建模方法,有效地增强了模型在不同工况条件下的鲁棒性。(5)以工业现场铝土矿泡沫浮选过程为研究对象,开发浮选过程监控系统,实现了泡沫图像特征的在线提取与矿浆pH值的实时检测。该系统能够实时监控各种工艺参数条件下泡沫外观特征及pH值的变化趋势,及时评估生产状态并更新专家规则,为调整碳酸钠药剂加入量提供操作指导。更多还原

【Abstract】 Flotation is an important part of alumina production process using ’bayer process with dressing’. Slurry pH is a direct reflection of the production condition and the consumption of sodium carbonate. However the pulp with poor mobility, high temperature and high corrosion, easy to plug the contact pH test device, while off-line tests are adopted by most of concentrators, which leads to a long lag time, and is difficult to continuously detect. Studies have shown that froth image is the indictor of pH, so researching real-time detection of pH values using froth images is great theoretical significance and practical application value.On the basis of the analysis of bauxite flotation process mechanism, considering the correlation of the bubble images and the pH value of pulp, froth color, size, shape and texture features are extracted, and the soft sensor of pH is established made up of multiple sub-model using image features, which is applied on froth flotation industry successfully. The thesis research and innovation results are as follows:(1) For the complex reaction mechanism of the pH value neutralization process of bauxite flotation process, the importance and the significant impact of pH on flotation performance is discussed from three aspects of bauxite particles floating, the activity of flotation reagent and ion composition. According to the law of conservation of mass and charge balance from the macro and micro aspects, it is proved that pH is the basis of adjustment of sodium carbonate and there is a typical non-linear characteristic in pH neutralization process based on ideal mixed continuous stirred tank reactor. A number of froth images are analyzed to choose proper image features, which shows that froth image is an indicator of pH.(2) Considering the difficulty to extracting froth features for bauxite flotation process, indicative function to pH of froth image, the feature extraction of bubble color, size, shape and texture is researched. As the interference of highlight and equipment aging causes color distortion, an image inpainting method is used to remove the top highlight of the bubbles, and then robust calculation with device-independent of froth color is realized based on color space conversion. Around the mineralized bubbles of various shapes, uneven scene illumination adversely affect the bubble image segmentation, the inertia weight and acceleration factor of particle swarm optimization is set to the function of global optimum fitness to detect threshold of valley edge. The local gray minimal is used to select boundary detection template, and then the froth image is segmented according to logic rules to obtain the bubble shape characteristics. In view of the radial elongated features along the forward direction for the surface froth, while the traditional spatial gray level co-occurrence matrix ignores the directional difference, which leads to inaccurate texture calculation, a multi-angle-fused spatial gray level co-occurrence matrix is proposed to calculate second-order statistics to describe texture. The image features provide plentiful data for modeling of pH soft-sensing.(3) In view of the non-linear characteristics of the pH neutralization process and the characterization to pH of the bubble images, the local soft-sesing model of pH is set up based on multi-kernel least squares support vector regression machine. Multiple kernels with different characteristics are convex combined to improve learning ability and generalization ability of the model. Focusing on the ignorance of cost and computational complexity of multiple kernel learning, a novel cost constraint multi-kernel learing idea is raised. The original optimization problem of multi-kernel least squares support vector regression machine is converted into the form of a second order cone programming, and then a cost factor of mult-kernel learing is difined to to determine the optimal combination of kernel functions. To reduce the computational complexity, an improved nearest neighbor peaks clustering algorithm and Schmidt orthogonalization method are adopted. The total cost of multi-kernel learing is evaluated according to the number of support vectors and active kernel, which saves the variable storage space and computing time.(4) Aiming at the deficiency of reflecting the entire flotation conditions for the local model, a global modeling using multiple working condition sub-models fuzzy combination is achieved. The flotation process is divided into three working conditions based on feature weighted fuzzy multi-classification support vector machine, and sub-model of bubble image features and pH value is established in each range of conditions. Distribute the importance metric of image features according to sensitivity differences in response to operating conditions based on information gain. The fuzzy membership of the bubble image on the working conditions is used as the combination weight of multiple sub-models to realize global soft sensing model. Effective results using industrial data of off-line froth image are achieved on the recognition of flotation performance and the soft-sensing of pH.(5) Taking bauxite flotation process in industrial field as a research case, a process monitoring system based on froth image analysis is developed, which could extract froth features on line and detect the pH values of flotation slurry in real time. That the flotation performance can be evaluated and expert rules can be updated according to froth image and pH achieved by soft-sensor, provides operational guidance to adjust the sodium carbonate amount.更多还原