【作者】 张泽琳；

【导师】 杨建国；

【作者基本信息】 中国矿业大学 ， 矿物加工工程， 2014， 博士

【摘要】 随着科技的发展，选煤自动化程度逐步提高，推动了煤炭洗选效率和经济效益的增长。然而国内选煤厂的自动控制还停留在自动启停车、介质密度桶液位控制、自动加药控制、产品灰分实时监测等零散反馈控制阶段，未能对整个生产环节进行实时监测和调控，主要原因在于目前没有任何技术能够实时监测入料煤和产品煤组成信息，进而不能对整个生产环节进行实时控制。为此，论文提出了基于机器视觉的煤质快速分析方法，力图在线测定煤的粒度组成、密度组成和平均灰分。论文以太西无烟煤为研究对象，建立了一套煤质快速分析实验系统。提出了三种煤粒图像分割方法即非接触煤粒背光图像分割方法、煤堆图像局部分割方法和煤堆图像整体分割方法。非接触煤粒背光图像分割方法主要针对背光散粒煤图像，采用双峰法、面积阈值和孔洞填充算法精确分割不重叠煤粒；煤堆图像局部分割方法为半自动分割方法，结合人工勾勒目标区域和彩色图像分割方法识别煤堆中的目标煤粒区域，其中还涉及到形态学图像处理、分水岭边缘处理、面积阈值区域筛选和煤粒区域最小外接矩形截取等算法；煤堆图像整体分割方法采用对比度受限自适应直方图均衡法、最小值和最大值滤波算法增强图像，进而引入Hessian矩阵和高斯函数的多尺度线性滤波器进行煤粒边缘检测，其效果比传统边缘检测算法更可靠更精确，最终采用双阈值边缘连接和标记分水岭分割算法识别煤粒区域；前两种图像分割方法主要用于精确建立预测模型，而第三种图像分割方法用于煤质快速分析；同时还提出了一种煤堆图像分割效果量化方法，结果表明煤堆图像整体分割方法误差区域百分比为12.76%。建立了一个基于机器视觉的煤粒质量预测模型，将煤粒图像二维信息转化为三维信息。通过对比分析煤粒区域实际大小与图像测量大小，表明图像测量方法精度很高；发现了煤粒区域面积与周长、最小外接矩形长和宽之间的指数关系，并分别采用最小二乘法建立了三个指数关系模型；采用多重线性回归方法（MLR）建立并优化了煤粒厚度预测模型，结合煤粒区域面积和密度提出了煤粒质量预测模型，测试结果表明煤样质量预测相对误差在±6%以内。提出了一种基于机器视觉的煤堆粒度组成快速分析方法。通过对比分析，采用最小外接矩形宽（DB）表征煤粒粒度；通过建立表面煤粒区域所属粒级的概率模型，提出了煤堆表面重叠误差校正方法；采用R-R粒度特性方程探寻煤堆整体粒度分布参数和表面粒度分布参数之间的内在关系，提出了煤堆颗粒偏析误差校正方法；结合上述研究提出了一种煤堆粒度组成预测方法，测试结果表明实验次数越多，预测误差越小，两种误差校正方法能够有效减小粒度组成预测误差，前20次煤堆粒度组成平均预测误差最高为3.79%，最低为0.03%。提出了一种基于机器视觉的煤堆密度组成快速分析方法。本文提取了煤粒表面50个颜色、光泽和纹理特征参数，并对其进行异常点检测和标准化处理；采用箱线图对所有特征参数进行了初步分析，探索其随煤粒粒级和密度级的变化趋势，初步筛选特征参数；随后采用核主成分分析（KPCA）和遗传算法（GA）进一步优化特征参数，结果表明GA特征筛选方法效果更佳；支持向量机分类器（SVM）相对于BP、RBF和PNN神经网络能够更准确的预测煤粒密度级；窄粒级煤粒密度级预测准确率远高于全粒级煤粒，并且窄粒级中煤粒密度级预测准确率随粒级增大而升高；结合上述研究提出了一种分粒级进行的煤堆密度组成分析方法，测试结果表明实验次数越多，预测误差呈降低趋势，前20次各密度级组成平均预测误差最高为8.03%，最低为0.87%。提出了一种基于机器视觉的煤堆各密度级灰分和总灰快速分析方法。通过建立煤粒灰分与密度的二次多项式模型，确定煤粒表面特征参数随灰分的变化趋势应与其随密度级的变化趋势一致；采用遗传算法（GA）筛选特征参数，结合支持向量机（SVM）建立煤粒灰分预测模型；结果表明全粒级煤粒灰分预测模型精度不如窄粒级灰分预测模型，并且灰分预测模型精度随粒级增大而升高，SVM预测模型比BP和RBF神经网络模型更适合用于煤粒灰分信息预测；结合上述研究提出了一种分粒级进行的煤堆各密度级灰分与总灰分析方法，测试结果表明预测误差随实验次数增多呈降低趋势，前20次各密度级平均灰分预测误差最高为3.39%，最低为0.22%，总灰平均误差为0.73%。初步尝试了基于机器视觉的煤质快速分析半工业试验。在神华宁煤太西选煤厂设计开发了一套原煤可选性实时预测系统和一套超纯煤灰分实时预测系统，运行情况表明煤质快速分析方法是可行的，两套设备基本能够满足现场环境实时煤质预测要求；实时预测的入料原煤粒度组成、密度组成和灰分信息绝对误差均在10%以内，并可实时显示可选性曲线；实时预测的超纯煤灰分绝对误差平均值为0.12%；本文研究内容的初步应用表明了煤质快速方法的有效性，后续研究将继续提高系统稳定性和预测精度。更多还原

【Abstract】 Coal separation efficiency and economic benefit were increased with thescientific and technological development and the increasing automation degree of coalpreparation. However, the automatic controls of domestic coal preparation plants werestill stayed as several feedback control phases, such as automatic launching andstopping machine, liquid level control of medium density barrels, automatic dosingcontrol, real-time ash monitoring. The real-time monitoring and controlling of thewhole production processes has not been realized. The main reason is the afunction ofreal-time monitoring of raw coal and products quality. Hence, this paper proposed themethods of fast analysis of coal property based on machine vision, mainly includingsize distribution analysis, density distribution analysis and ash content analysis.Tai-xi anthracite was taken as research object, and a fast analysis system of coalproperty were built for experiments. According to the need of fast analysis of coalproperty, three image segmentation methods of coal particles were proposed,including segmentation method of non-touching coal particle image,local-segmentation method of coal pile image and whole-segmentation method of coalpile image. Segmentation method of non-touching coal particle image mainly aimedto backlit images of non-touching coal particles, and two-peaks method, areathreshold, hole-filling method were used to segment coal particles accurately.Local-segmentation method of coal pile image is a semi-automatic segmentationmethod, combined with drawing the outline of the target region by manual and colorimage segmentation method. Related algorithms also include morphologicalprocessing, watershed edge processing, area threshold method and minimumcircumscribed rectangle interception. Whole-segmentation method of coal pile imageused CLAHE, minimum and maximum filter algorithms to enhance image.Multi-scale linear filter by Hessian matrix and Gaussian function was used to detectthe coal particle edges, and the effect is better than traditional edge detectionalgorithms. Finally double-threshold edge connection and marked watershedalgorithm were taken to identify the coal particle region. The first two imagesegmentation methods were mainly used to establish the estimated models accurately,and the third image segmentation method was used for fast analysis of coal property.Meanwhile, a segmentation effect quantitative method of coal pile images wasproposed, and the error percentage of the above segmentation method is12.76%.In order to estimate the3D information of coal particle from its2D information, a mass estimation model for coal particles was established in this paper. Actual sizeand measured size by image processing were contrasted and analyzed, showing theimage measuring method is accurate. The exponential relationships between area andperimeter, minimum circumscribed rectangle length and breadth were found and threeexponential models were established by least square method. Thickness estimationmodel of coal particles was established and improved by multiple linear regressionmethod, and then mass estimation model of coal particles were proposed with areaand density of coal particles. Test results indicated the absolute errors of estimatedmass of coal samples are less than6%.A fast analysis method of size distribution of coal piles by machine vision wasproposed. Ten size features were contrasted and analyzed, and then the breadth ofminimum bounding rectangle were determined as the best particle sizecharacterization of coal particles. Through establishing the size-fraction probabilitymodel of surface coal particles, a surface overlapping error correction method wasproposed. R-R granularity characteristic equation was used to explore the innerrelationship between equation parameters of the whole coal pile and the surface, andthen a granular segregation error correction method was proposed. Combined withabove researches, an analysis method of size distribution prediction of coal piles wasproposed, and the results indicated the more test times, the smaller prediction errors.The above two correction methods were useful to reduce the prediction errors. Thehighest error of the first twenty estimated results of coal pile size distribution is3.79%,and the lowest error is0.03%.A fast analysis method of density distribution of coal piles by machine visionwas proposed. Fifty color, luster and texture features were extracted and processed byoutlier detection and standardized treatment. Box-plots were used to analyze thevariation tendency of all features with the increasing of size fractions and densityfractions, and then selecting all the features initially. KPCA and GA were used tooptimize the left features, and results indicated GA is more suitable for featureselection. SVM is better than BP, RBF and PNN to predict the density fraction of coalparticles. The prediction accuracy of narrow size fractions is much higher than thewhole size fraction, and the bigger size fractions, the higher prediction accuracy.Combined with above researches, an analysis method of density distributionprediction of coal piles by each narrow size fraction model was proposed, and theresults indicated the more test times, the smaller prediction errors. The highest error of the first twenty estimated results of coal pile density distribution is8.03%, and thelowest error is0.87%.A fast analysis method of total ash content and ash content of each densityfraction of coal piles by machine vision was proposed. Through establishing thequadratic polynomial model of ash content and density of coal particles, the variationtendency of features with the increasing of ash content should be consistent with thatof features with the increasing of density. GA method was used to select the features,and SVM was used to establish the prediction model of ash content. Results indicatedthe prediction accuracy of narrow size fractions is higher than the whole size fraction,and the bigger size fractions, the higher prediction accuracy. SVM model is betterthan BP and RBF models in ash content prediction. Combined with above researches,an analysis method of total ash content and ash content of each density fraction ofcoal piles by each narrow size fraction model was proposed, and the results indicatedthe more test times, the smaller prediction errors. The highest error of the first twentyestimated results of coal pile density distribution is3.39%, the lo west error is0.22%,and the average error of total ash content is0.73%.Pilot-scale tests of fast analysis of coal property by machine vision were carriedout in a preliminary attempt. An online washability prediction system of raw coal andan online ash content prediction system of ultra-pure coal were designed andestablished in ShenHua Ningmei Taixi coal preparation plant. Test results indicatedthe fast analysis methods of coal property are feasible and the above two systems areable to satisfy the prediction requirements of coal preparation plant basically. Thereal-time prediction absolute errors of size distribution, density distribution and ashcontents are all less than10%and the washability curves is able to show in real time.The real-time prediction absolute error of ultra-pure coal ash contents is0.12%. Theprimary applications show the availability of analysis methods proposed in this paper.System stability and prediction accuracy should be improved and enhanced in futureresearch.更多还原