【作者】 刘金平；

【导师】 桂卫华；

【作者基本信息】 中南大学 ， 控制科学与工程， 2013， 博士

【摘要】 泡沫浮选是以表面化学为基础,根据矿物粒子表面疏水性的不同来选别有用矿物,是最重要的矿物分选方法。尽管泡沫浮选已有百年应用历史,但由于实际工业浮选流程长、过程影响因素多、内部机理不明确,矿物浮选过程自动监控仍然难以有效实施,造成矿物资源回收率低且指标波动频繁。鉴于工业过程视觉监控具有检测速度快、结果客观和不干扰浮选生产的优点,近年来,基于机器视觉的矿物浮选过程优化控制是进一步提高选矿自动化水平和矿产资源回收率的发展趋势。在基于机器视觉的矿物浮选过程监控中,研究合适的泡沫图像处理和分析方法以获取与生产工况密切相关的泡沫表面视觉特征参量是进行后续浮选过程建模与自动控制的基础。然而,由于浮选泡沫表面表现出来的随机堆积性、无背景性、形状不规则性等特点,常用的图像处理、分析方法难以有效实现泡沫表面视觉特征的准确测量与分析。为了能对这些形状各异、大小不一的、随机堆积的矿化气泡进行特征量化描述,本文将概率论、统计学习、模式识别等方法应用到图像信号处理和分析中,根据特定变换域的泡沫图像随机场所表现出来的统计分布特性建立合理的图像统计分布模型,为后续图像的分析和理解提供有效的先验知识；然后,基于所建立的统计模型,有效解决泡沫图像表面颜色、气泡大小和表面随机纹理等视觉特征难以准确表征问题,并成功应用于矿物浮选过程的泡沫状态分类与生产工况智能识别中,实现浮选生产工况的机器鉴别与自动评价。论文主要研究工作及创新点如下：(1)针对浮选泡沫图像噪声大,泡沫表面视觉特征不能准确提取的难题,提出一种时空信息联合的图像序列多尺度几何变换去噪方法。该方法首先通过收集大量未受噪声干扰的泡沫图像样本进行图像统计分布建模,建立了泡沫图像多尺度几何变换域系数的统计分布模型；然后,以所建立的图像统计模型为先验知识,采用贝叶斯最小二乘估计方法获得基于帧内信息的泡沫图像去噪结果；最后,根据图像序列时空相关的信息,通过加权处理帧间图像去噪信号,实现了时空信息融合的图像序列最优无噪图像信号估计,解决了常规图像去噪方法中经常遇到的图像细节系数与图像噪声难以区分的难题。该方法在提高图像去噪效果的同时,极大限度地保持了图像边缘和表面纹理细节,为泡沫图像特征分析与理解提供了高质量的处理信号。(2)针对泡沫图像因色偏严重而难以实现泡沫表面真实颜色的准确测量问题,提出一种基于图像空间结构统计分布的最优泡沫图像颜色自动校正方法。基于已有的图像光照估计方法,首先深入分析了图像边缘响应统计分布特征与图像最优光照估计方法间的关系,建立了基于图像边缘响应统计分布的图像入射光照最优估计模型；然后,根据入射光照的估计结果将发生色偏的泡沫图像自动校正到标准参考光照下的颜色表示。在进行泡沫图像光照估计时,以Ciurea和Funt建立的包含11346帧图像的Gray-ball标准光照数据库为训练样本,通过统计学习获得基于图像边缘响应分布的图像最优光照估计方法选择的混合高斯(MoG)模型,实现了图像最优光照估计方法的自动选取。实验表明,该方法能自动根据泡沫图像边缘响应的统计分布特点实现图像最优光照估计,进而有效地对泡沫图像颜色进行颜色校正,为泡沫表面颜色特征的准确提取与浮选生产工况的客观鉴别奠定了基础。(3)针对矿物浮选过程加药操作缺乏有效评价方法的问题,根据气泡大小分布随药剂操作的改变而动态变化的特点,提出一种基于泡沫大小动态分布特征自适应学习的浮选过程药剂操作健康状态统计模式识别方法。首先根据泡沫图像局部区域像素的统计分布特点和矿化气泡的几何边缘特性,提出一种改进的泡沫图像分割方法,解决了因泡沫表面矿物粒子随机粘附造成矿化气泡上表面高亮点分散而引起泡沫图像的严重过分割问题；然后通过核密度估计获得浮选泡沫大小的累积分布函数(CDF)；再采用无监督的最远邻聚类(FNC)学习方法获得各典型药剂操作状态下气泡尺寸统计分布特征集；最后,根据测试时间段的浮选气泡大小分布的动态变化特点,采用贝叶斯推理获得对应的药剂操作健康状态识别结果,并自动根据浮选生产工况的波动情况对各典型药剂状态下的气泡统计分布特征集进行在线修正。所提出浮选药剂操作健康状态的统计识别方法,能实时跟踪泡沫大小分布的动态变化,进而根据气泡大小分布的变化情况实现浮选药剂操作健康状态的自动识别与客观评价,为实现浮选生产过程的加药量优化控制奠定基础。(4)为了能进一步根据浮选泡沫表面纹理的细微差别实现浮选生产工况的自动鉴别与评价,提出一种基于泡沫图像多尺度多方向纹理表征的浮选生产工况综合分类与识别方法。鉴于二维Gabor基函数具有与绝大多数哺乳动物的视觉皮层简单细胞的感知域模型相似的性质,将Gabor小波变换应用到泡沫图像的多尺度多方向视觉特征提取中。分别分析了各分解子带上的图像实部谱(RGFR)、虚部谱(IGFR)、幅度谱(AGFR)和相位谱(PGFR)的统计分布特征,建立了RGFR、IGFR、AGFR和PGFR的边缘分布和图像相邻像素值的联合统计分布模型；分别采用t Location-Scale分布和Gamma分布模型来拟合RGFR, IGFR与AGFR的边缘分布,并计算各Gabor小波卷积谱的联合分布特征参量共同作为泡沫图像的表面纹理特征参量集；最后,利用所提取的泡沫纹理特征对浮选工业生产状态进行无监督的模糊聚类分析与有监督的生产状态识别。实验效果表明,该纹理特征提取方法有效地获取了各种浮选状态下泡沫表面纹理的细微差别,基于该纹理特征参量的浮选状态识别准确率高。(5)以中国铝业有限公司中州分公司的铝土矿浮选过程监控为例,将所提出的图像建模、分析和识别方法应用到矿物浮选过程监控中。在铝土矿浮选现场设计并搭建了浮选泡沫图像采集、处理软硬件平台,实时提取了包括泡沫颜色、气泡大小及其分布、表面纹理等浮选泡沫图像表面视觉特征,根据泡沫表面视觉特征与生产工况的关系,获得了最佳精选泡沫图像的局部纹理特征区间,并最终实现了整个浮选流程中关键浮选槽泡沫状态的机器鉴别与客观评价。应用表明,监控系统所获得的泡沫特征曲线为生产工人提供了明确的工况信息,并给出具体的操作建议,避免了工人操作的盲目性,有效提高了矿物回收率并稳定了选矿生产指标,为浮选过程优化控制奠定了基础。图89幅,表10个,参考文献211篇。更多还原

【Abstract】 Froth flotation is the most important mineral separation technology, which is used to concentrate the valuable minerals from the source ores according to the different surface hydrophobicities of the mineral particles based on the principle of surface hydrophobicities. Though one hundred years have elapsed, the automatic monitoring of the flotation process is still hard to be put into effect, resulting in low recovery of the mineral resource with fluctuating product indexes, sine the practical industrial flotation process usually consist of long process circuits and great amount of influence factors with unclear inner mechanism. In view of the merits of the industrial vision monitoring system, which responses fast and provides objective and non-intrusive monitoring of the froth states, machine vision based flotation process monitoring is recently expected as a promising tool to improve the online detection, measurement and control means and ultimately to achieve the optimal control the flotation process by both the scientific researchers and industrial engineers. In the machine vision based flotation process monitoring, researching the proper froth image processing methods and extracting the effective bubble features closely relating to the production condition of the flotation process are the prerequisites for subsequent flotation process modeling and optimization. However, the traditional image processing and analysis methods are difficult to be used in the froth image analysis and feature extraction due to the random accumulation of the mineralized bubbles, which have diverse morphological structures without background between each other. In order to quantitatively delineate the randomly accumulated bubbles with various shapes and sizes and obtain their comprehensive visual features, which do not depend on the single bubble in the froth surface, the methods such as probability theory, statistical learning and pattern recognition are applied to the froth image analysis and processing. The statistical distribution models of the froth image are established in the image transform domain according to the statistical distribution profile of the random signal characteristics for the subsequent image analysis and understanding to provide effective prior knowledge. The statistical characteristics of the froth image are studied by establishing reasonable mathematical models for image distribution feature description, which effectively solve the problems of accurate extraction of the forth surface color, bubble size and bubble surface texture. The extracted froth image statistics are successfully applied in the froth state classification and the intelligent production condition recognition in the mineral flotation process. The main researches and contributions are as follows:(1) A spatio-temporal information fused froth image denoising method based on multi-scale geometric transformation of the froth image sequences is proposed, aiming at solving the problem of serious noises on the froth image, which may lead to inaccurate extraction of the froth surface features. Firstly, a great many of froth images are collected for statistically modeling the froth image coefficients in the multi-scale transform domain. Then, the recovery of the clean froth image signals is obtained by Bayesian least square estimation based on the intra-frame information with the established statistical model of the froth images as the prior knowledge. At last, by weighting the inter-frame recovery signal, it restores the uncontaminated image signal optimally based on temporal and spatial fused image sequence information. It solves the problem of the confusion of the image details and image noises, which is frequently existed in the commonly used image denoising methods. This method can remove the image noise while effectively keeps the edge curves and surface textures of froth image simultaneously, which provides high-quality image signals for the subsequent froth feature extraction.(2) An adaptive froth image color correction method is presented based on the statistical distribution of the spatial structures of the images, since the froth images are prone to have the color cast. The correlationship of the edge response distribution features of the froth images and the corresponding optimal illumination estimation method is analysed in advance, which results in the establishment of the optimal scence illumination estimation model based on the statistical distribution of the edge responses of the froth images. Then, the forh image color will be corrected to the comparative color space under a canonical illumination according to the estimated results of the incident illumination color. During the illumination estimation, a well known image database for color constancy research including11346images with the known illumination built by Ciurea and Funt is used as the training samples to establish the distribution model of the image edge responses. After building the Mixture of Gasussian classification model based on the relation of the statistical distribution features of the images with their corresponding optimal illumination estimation method, it achieves the automatic selection of the optimal illumination color estimation methods. The experimental results demonstrate that this method is capable of achieving the optimal estimation of the illumination of the image and effectively correcting the froth image color, which paves the way of accurate color feature extraction of froth images and automatic identification of the flotation production conditions.(3) An intelligent recognition method of the health states of the reagents operation is presented based on the adaptive learning of the dynamic distribution features of the froth bubble size distribution according to the fact of the bubble size distribution of flotation froth varies with the dynamic changes of the reagent operation, for lack of the effective method to monitor and assess the health states of the reagent operation. Firstly, an improved froth image segmentation algorithm is proposed based on the local pixel gray-level distribution of the froth images in combination with geometric characteristics of the bubble edge curves. It solves the problem of image over-segmentation resulting from the dispersed highlights caused by the light reflections of the mineral particles on the bubble surface. Then, the cumulative density functions (CDF) of the froth images are fitted through the statistics of the segmented region area with the kernel density estimation. The statistical distribution feature sets of the bubble size under the typical operational conditions are learned by unsupervised furthest neighbor clustering (FNC); subsequently, the current health state of the flotation process in the test time period is inferred by the Bayesian rules according to the dynamic change of the bubble size distribution; what is more, the bubble size distribution features extracting under the typical dosage addition conditions are updated timely according to the disturbance of the operation conditions. The proposed health state recognition method can track the dynamic change of the bubble size distribution and achieve the automatic recognition and evaluation of the reagent operation effectively, which lays a foundation for the realization of the optimal control of reagent addition in the flotation process operation.(4) A kind of multi-scale and multi-orientation texture features of froth images based automatic operational condition classification and recognition method is proposed for the ultimate purpose of automatic identification and evaluation of the flotation production conditions according to the subtle change of the forth surface texture. The Gabor wavelet transformation is used to decompose the froth image into in advance in view of the two dimensional Gabor basis function can simulate the responses of the simple cells in the visual cortex of the most mammal brains. The convolution images including the real part (RGFR), imaginary part (IGFR), amplitude part (AGFR) and phase part of the Gabor filter responses (PGFR) are statistically analyzed respectively, whose marginal distribution and joint distribution features are both taken into account in this work. And then the distribution profiles of RGFR (IGFR) and AGFR are characterized by t Location-Scale and Gamma distribution respectively and the cumulative distribution features of each joint distribution of the Gabor filter responses are also calculated. Both the marginal distribution feature parameters and the joint distribution features of the Gabor filter responses at each sub-band are taken into account to construct the froth image texture feature parameter vector. At last, the extracted froth image texture variables are used to unsupervised clustering analysis and supervised recognition of the industrial production states. The experiment results demonstrated that this method can extract the distinctive froth texture in various flotation states and achieve high recognition rates of the flotation production states by these froth image texture parameters.(5) The proposed image statistic analysis methods are applied in a bauxite flotation plant in the Zhongzhou branch of China limited aluminum corporation, where the flotation froth image acquisition device is designed and mounted by deploying the corresponding software system for image processing and process monitoring. The visual features of the froth image are extracted, including froth color, bubble size with size distribution and the froth surface texture, and so on. The optimal texture feature interval of the forth image is obtained in accordance with the relation of the froth image features and the production conditions. Consequently, the vision system can offer the real-time measurement and objective evaluation of the flotation process. The application results indicate that the extracted froth feature curves can offer operator useful production information with operation adjustment advice, which avoid the blindness of manual operation of the flotation process. To summarize, the established system with the proposed method improves the flotation performance efficiency and lays a foundation for the optimal control of the flotation process. There are89figures,10tables and211references in this dissertation.更多还原