【作者】 刘伟华；

【导师】 隋青美；

【作者基本信息】 山东大学 ， 控制科学与工程， 2011， 博士

【摘要】 煤矿井下生产作业过程中会产生大量粉尘颗粒,它们能够较长时间悬浮在空气中,会在重力的作用下慢慢地沉积在工作面的底板、巷道壁、机电设备的表面等。若机械运转、局部通风、人员走动、放炮则会使已沉降的粉尘再次飞扬形成二次扬尘,骤然增加作业场所内空气中的粉尘浓度。悬浮着的粉尘会危害矿工的身体,导致尘肺病。此外,煤矿中的粉尘达到一定浓度(在爆炸下限至上限浓度范围内),遇明火还有可能发生爆炸,爆炸产生的冲击波会使积尘扬起,导致产生更为严重的后果。此外,煤炭生产已经实现了机械化,部分煤矿已经实现了生产自动化,大量的贵重设备和精密仪器应用在井下,煤尘颗粒的粘附会加速机械设备的磨损,缩短精密仪器的使用寿命。因此需要监测粉尘的变化,并依据检测结果指导井下作业环境中的抑尘、防尘、降尘和除尘。但传统的检查方法多为手动操作,只能定时定点对粉尘进行检测,测量结果不稳定、不能反映煤尘浓度的变化规律,实时性差、不能及时有效地指导防尘降尘工作。因此研制煤尘在线检测系统有着重要的理论意义和应用价值,本文研究的主要概括如下：(1)设计了一种基于机器视觉的煤尘颗粒在线检测系统。机器视觉技术因其直观、智能、易与其他系统集成而在工业生产过程中得到广泛应用。基于机器视觉粉尘检测系统可以实现长期、多点、非接触测量,因此本文提出了一种新的基于机器视觉的远程在线微米级颗粒检测方法,基于该方法实现远程监测煤矿井下粉尘。检测系统的功能包括两部分下位机的图像采集和上位机的图像分析。下位机系统实现煤尘样本采集、煤尘图像采集和图像远程传输,上位机对获得的图像进行处理和分析,并给出颗粒分析的报告。(2)提出了一种滤波参数自适应的非局部全变差滤噪算法。获得的粉尘图像经过压缩、传输、解压等一系列的步骤,图像信噪比低、对比度差,需经滤噪提高图像质量。非局部均值去噪的基本原理是遍历整幅图像,通过与其相似像素灰度值加权平均来估计每个像素点的灰度值。图像中不同像素点间灰度值越相似,它们具有的相似权系数就越大。非局部均值方法虽能有效地降低图像噪声并保持图像的细节,但也存在着最优滤波参数选择和运行速度慢的问题。通过分析滤波参数与噪声标准差之间的关系,发现滤波参数的选择对图像去噪的效果影响很大,因此文中给出了定量估计最优滤波参数的方法。此外,结合非局部均值滤波和全变分运算的思想,基于图像像素间的邻域相似性和搜索窗内相似性两个度量,建立了一种非局部均值全变差的滤噪模型。采用Split-Bregman迭代,在提高了运行效率的同时保持了图像的纹理和边缘。滤噪后的图像具有更高的信噪比,更快的运行效率,且去噪后的图像保持了纹理和边缘等细节特征。(3)提出了对比度自适应增强的多尺度Retinex颗粒图像增强算法。图像采集系统采用了暗场照明的方式,但所取得的图像由于环境的原因仍会出现背景不均匀的现象。若未加处理直接增强图像,则会将本应属于背景的区域错误地增强为目标,导致后续图像分析出现误差。多尺度Retienx算法是依据人眼视觉感知特性将图像分为照射分量和反射分量,通过将图像与多尺度高斯函数卷积估计照射分量,再从原图像中减去该照射分量,即可去掉光照对图像的影响,得到反射分量。经典Retienx算法能有效地压缩图像的动态范围,改善图像的质量,但仍存在着图像变灰的问题。为了增强图像的对比度,采用归一化的非完全Beta函数增强Retienx处理后的图像。根据实验采集图像的特点,确定了最优对比度时的非线性函数的两个参数的值。经过试验证明,增强后颗粒图像具有连续均匀的背景和更高的图像对比度。另外,通过对颗粒粒径测量的实验也佐证了处理后的图像能够显著减少因光照而导致的粒径测量的误差。(4)提出了基于改进微粒群算法的二维最大熵的自适应图像分割方法,获得了连续一致的图像分割结果。虽然图像分割的方法很多,但阈值法因其简单有效而被广泛应用。熵是信息量的表征,利用图像中各个像素的点灰度值及其区域灰度均值生成二维直方图,它所对应为二维熵原理,熵最大时的阈值为最优阈值。二维最大熵图像分割方法的核心内容是利用点灰度和区域灰度均值信息提取图像的有用信息,忽视了边界和噪声点,使得在图像信噪比低时也能取得好的分割效果。为了提高寻求最优阈值的速度,采用改进的微粒群算法寻找最优阈值,但微粒群算法存在过早收敛和微粒趋同的问题,因此引入的多群共享向量和概率学习变量的改进算法。多群共享向量实现多个独立的群之间的信息的共享,保留每代粒子中每一维的当前全局最优值,改进了“维数灾”的问题。为了解决因为微粒停滞导致优化陷入僵局的问题,引入学习概率变量,改进“颗粒趋同”的问题,即以一定的概率对几代内未更新的微粒重新初始化。基于改进微粒群算法选取的阈值稳定,寻优速度快,实现了图像的连续一致的自适应分割。(5)提出了一种基于凹点搜索和匹配的颗粒分离重叠颗粒,提高颗粒检测的准确率。煤尘颗粒形状不规则,边界凹凸不平,重叠颗粒的数目也较多,分水岭和腐蚀膨胀的方法不适用,因此提出了基于凹点搜索和匹配的方法。首先计算每个区域所包含的颗粒个数,判定是否为重叠区域。然后,计算点到不同长度弦的距离乘积表征曲率,并设定曲率阈值和夹角阈值,据以选取重叠颗粒的凹点。根据凹点和颗粒个数的对应关系确定不同的重叠类型和不同的匹配规则,用Bresenham画线连接匹配凹点分离颗粒。基于凹点搜索和匹配的颗粒分离方法可以实现不同重叠程度、多个粘连颗粒的分割。凹点的判别主要是简单的平方、相加和开根号等运算,算法的复杂度低,运行速度快,同时避免了重叠颗粒的过度分割。运算过程中无需多次腐蚀与膨胀,因此能够保持不规则颗粒图像的边缘,使得分离后的变形小,保证了后续颗粒分析的正确性。(6)最后设计仿真实验系统模拟煤矿井下的粉尘测量环境,验证检测结果的准确率和重复率。先采用标准微粒作为实验样本验证系统测量的准确性,然后采用真实的煤尘颗粒作为样本来验证系统的可重复性。仿真实验证明系统能够直观地监控煤尘浓度的变化,处理并分析颗粒图像。最后给出煤尘粒径的分布报告,并且颗粒粒径检测的准确率超过94%,重复率超过98%。更多还原

【Abstract】 Lots of coal dust are produced during the process of production operation in all coal mine’s working places. Coal dust can be suspended in air for a long time, and they will be slowly settled on those places under the gravitational attraction, such as the floor of mining face, walls of the tunnel, surface of electromechanical equipment. Dust concentration in the air will increase suddenly in workplaces, when settlement dusts were raised again, if machine operation, local ventilation, miner moving and blast firing. Suspended dust would endanger the body of miners and even lead to pneumoconiosis. In case of coal mine dust reaches a certain concentration (in the range of upper and lower range explosion concentration) and open flames, an explosion may be occur, explosion shock wave will kick up dust and result in more serious consequences. In addition, the mechanization of coal production has been achieved, some coal production automation has been achieved, a lot of expensive equipment and precision instruments used in mine, the adhesion of coal dust particles will accelerate wear of machinery and equipment, shorten life of precision instruments. For the above reasons, it is important to monitor changes of coal dust and guide reaching curb dust, falling dust, dust proof, removing dust. However most of traditional inspection methods are manually operated, and dust can only be detected on the fixed time and area, which will lead to uncertainty measurement, can not reflect the variation of dust concentration. Because the detection is not real-time, it can not effectively guide the work of reach curb dirt. Thus, development on-line detection system of coal dust may produce important theoretical significance and application value, and the major contributions in this dissertation are following:(1) A new remote on-line coal dust particle detection system based on machine vision is designed. Machine Vision has been widely used in the industrial production process because of its intuitive, intelligent and easy to integrate with other systems. Dust detection system based on machine vision can monitor particle in a way of long-term, multi-point and non-contact, so a new remote online micron particle detection method based on machine vision has been developed in this paper, which made it possible to realize remote monitoring coal dust. The chief function of single-chip microcomputer are collecting particle, obtaining image and transmitting image. Dust samples were collected using a low-cost particle collector sensor with a quick way, image of micron dust particles were obtained using image particle collector, then the image were transmitted to the particle analysis system in remote control room by cable.(2) A new adaptive non-local mean total variation algorithm is proposed to denoise and improve image quality. Obtained particle images with low signal to noise ratio and poor contrast after a series of steps, such as compression, transmission, decompression, so it is necessary to filter noise and improve image quality. The principle of Non-local means denosing is estimating gray value by weighted average of gray value between similar pixels by traversing the entire image. The more similar between two pixels, similar weights will be larger. Although non-local means method effectively reduced image noise and preserved image details, but there are problems of how to determine optimal filter parameter and slow speed. By analyzing the relationship between filter parameter and standard deviation of noise, we found that filter parameters would impact on the effect of denosing, so a quantitative estimate optimal filter parameters was given. In addition, a new denoise model based on similarity weight of pixel neighborhood and search window is proposed, which combined non-local mean filter with total variation algorithm. In order to improve the operational efficiency, Split-Bregman iteration was used. Experiments show that our proposed algorithm has quite good ability of noise suppressing as well as edge and texture preserving, operational efficiency is also faster.(3) Adaptive contrast enhancement algorithm based on multi-scale Retinex for particle image is proposedIn order to obtain high contrast, multi-scale Retienx is used to enhance image. Dark-field illumination is designed in our system, but image background is uneven occasionally because of environmental, certain parts of the image background are darker and others are lighter. If we enhance images directly without handling uneven background, then some background will be mistaken as particles, which will lead to measurement errors in following image analysis, the principle of Multi-scale Retienx is simulate human visual perception characteristics, it divide image into light component and reflection component. Light component of image is estimated by convolution with the multi-scale Gaussian, and reflection component can be calculated by subtracting light component from the original image. Classic Retienx algorithm can effectively compress the dynamic range of images, improve image quality, but there is a gray-out problem. In order to enhance image contrast, a normalized nonlinear Beta function is used to enhance image processed by Retienx. According to the features of obtained images, two parameters of nonlinear function are determined. Experiments prove that processed images have more uniform background and higher contrast. Furthermore measurement errors caused by illumination are significantly reduced.(4) Image segmentation with 2-D maximum entropy based on improved particle swarm optimization is proposed to separate particle from particle in consistent manner. Although there are many methods of image segmentation, threshold method is widely used for its simple and effective. Entropy is the characterization of information.2-D maximum entropy consider of both gray information and spatial neighbors using the 2-D histogram of the image. Maximum entropy criterion is used to determine the optimal threshold. For sake of extraction useful information, image segmentation method with 2-D Maximum entropy make use of the gray value and average regional gray value, which ignores the boundaries and noise points, image can also achieve good segmentation results when the signal to noise ratio is low. However, the time-consuming computation is an obstacle for this method to be used in real time application systems. In order to improve the speed of finding the optimal threshold value, an improved PSO algorithm is used, but there are two problems premature convergence and particles convergence, so multi-group shared vector and probability learning variable are introduced. Multi-swarm shared vector is used to improve the problem of "dimension disaster" by information sharing among different swarm, which will retain current global optimal value of multiple independent particles in each generation each dimension. In order to solve the problems of "particle convergence" leaded by particle stagnation, learning probability variable is the introduced, which is a certain probability of the re-initialization if particles were not updated in several generations. Experiments show the proposed algorithm is effective and threshold stability, it can separate particles from image adaptively in a consistent manner.(5) An algorithm of powder particle automatic segmentation is proposed, which makes use of automatic finding and pairing concavity points to improve the accuracy of particle detection. Watershed and corrosion expansion method is ineffective in separating overlapping coal dust particle, because of particle shape irregular and uneven borders, in addition, the number of overlapping particles is greater. So a new algorithm based on the concave point search and matching is put forward. Firstly, each region is selected to determine overlapping or not according to the number of particles contained in it. Then, curvatures is calculated by the product of distance form point to chord with different length, concave points of overlap particles are decided according to angle and curvature threshold. Then, different kinds of matching rules are given by comparing the numbers of concave points with the numbers of particle. Segment lines are drawn based on Bresenham and overlapping particles are separated after obtaining matching concavity points. Because most of calculations are simple operations, such as power, square root and sum, and algorithm is effectively, low complexity, it can avoid over segmentation of overlapping particles. Without corrosion and expansion operations during operation, so edge and shape of irregular particles can be maintained, which will ensure the accuracy of particle analysis in following particle analysis.(6) In order to verify the accuracy and the repetition rate of the detecion system, simulation system is designed to simulate environment of dust measurements under coal mine. Standard particle and coal dust particle are used as sample to verify accuracy rate and repeat rate.Experimental results show that this system can directly monitor changes of dust concentration, process and analysis particle image. Finally, particle size distribution report is given, accuracy rate and repeat rate of particle size is greater than 93% and 98% respectively.更多还原