【作者】 李俊峰；

【导师】 戴文战；

【作者基本信息】 东华大学 ， 机械设计及理论， 2010， 博士

【摘要】 铁谱技术以磨损产生的微观颗粒为研究对象,通过分析机械设备中各种类型磨损颗粒的数量及趋势,监测机械设备所处的状态,进而预测机械设备可能发生的故障。在铁谱技术中,磨粒图像是反映机械设备内部零部件磨损状况的重要信息载体,磨粒特征分析是监测对象实际状态十分丰富而又有效的方法。磨粒识别是铁谱分析的核心环节,识别的正确与否,直接关系到磨损状态诊断的正确性。随着计算机技术的迅猛发展,将计算机视觉技术、专家系统、人工神经网络等引入铁谱分析技术中,实现磨粒识别的智能化已成为铁谱技术研究领域中的热点和难点问题。由于图像具有灰色特性,如图像像素的灰度值、图像的边缘、图像的噪声及图像分割阈值等等,近年来,灰色系统理论受到图像工程领域研究人员的广泛关注和重视,同时探讨它在图像工程中的可行性和有效性逐渐成为一个崭新的课题。本文以铁谱磨粒图像的处理、特征提取及识别为背景,在运用灰色系统理论对磨粒图像进行处理的基础上对以下五个问题进行了研究。1.提高灰色预测模型精度研究研究发现,影响灰色预测模型精度的主要因素有原始数据序列的光滑度、模型的背景值及模型的初值。基于此,分别就上述三个影响因素对提高灰色预测模型精度进行了研究:（1）提出了基于函数x^-a（a＞0）变换和基于多重复合函数变换提高数据序列光滑度的方法,理论上证明了数据序列经过该变换后可以有效地提高其光滑度,且其光滑度优于现有的其它变换函数,进而,总结了提高数据序列光滑度的变换函数的性质;（2）基于数据序列的指数函数特性和积分定义,提出了两种模型背景值构造方法,并对这两种改进灰色模型的拟合和预测精度进行了分析,结果表明这两种改进灰色模型不仅可以用来进行短预测,还可以进行中、长期预测;（3）提出了修正初值和时变初值的模型初值构造方法,给出了利用自适应遗传算法求待辨识参数的方法。为了取得更高的建模和预测精度,可以从序列光滑度、模型背景值及模型初值三个方面同时对灰色预测模型进行改进。2.灰色关联度研究研究发现:目前的灰色关联度模型存在一些缺陷,如灰色关联四公理和灰色关联度的计算方法之间存在着矛盾,序列不同无量纲化计算对关联度的计算结果也会产生影响等。基于此,对均值灰色关联度和T型灰色关联度进行了研究:（1）提出了改进的欧几里德灰色关联,改进的欧几里德灰色关联模型不仅考虑了各点关联系数对其平均值的波动,还考虑了正理想相关和负理想相关,具有平行性、规范性、整体性、偶对称性及接近性;（2）提出了改进的T型灰色关联度,改进的T型灰色关联度能够反映序列的正、负相关关系,具有对称性、唯一性、可比性、接近性及规范性,且对无量纲化处理具有保序性。3.基于灰色系统理论的图像处理算法研究（1）提出了基于绝对灰色关联度和LOG算子的边缘检测算法、基于相关性的图像边缘检测算法及基于级比的边缘检测算法,这些算法不仅具有较强的图像边缘检测能力,可以检测出图像不同方向的边缘,而且能够根据阈值调节检测的边缘细节,定位准确,连续性较好,易于软件编程并行实现,计算量小,其中前两种算法还具有一定的抑制噪声能力,特别对于Gaussian噪声、Speckle噪声及Poisson噪声;（2）提出了基于灰色关联系数的图像噪点检测算法,该算法根据含噪图像与对应均值图像在各像素点的灰色关联系数来识别噪点,利用了整幅图像的信息,特别是噪声的统计信息,可以有效的检测出图像噪点,避免误检;（3）提出了基于加权灰色关联度及基于灰色预测模型的噪点灰度处理算法,这两种算法以含噪图像对应的均值图像或中值图像为基础,分别采用灰色关联度和灰色预测模型来进行噪点处理,可以有效克服周围噪点的影响,更好地改善图像质量,同时图像的保真度较好;（4）提出了基于提升小波和灰色预测模型的图像压缩算法,该算法首先利用提升小波将压缩图像转换到频域,利用零树区分各波段中的重要系数与不重要系数,进而利用hilbert曲线将各波段扫描成一维,最后利用灰色预测模型进行预测编码。实验结果表明该算法有效地提高了图像的压缩率和压缩质量。4.图像保真度评价研究基于相关性指标的特点,提出了基于相关性和离散小波变换的图像保真度评价算法,该算法能够从图像概貌质量和细节质量两个方面对图像进行评价,可以比较准确地反映图像的质量,是一种有效的、具有多尺度分辨功能的、客观的且定量的图像评价方法,更加符合人类视觉系统的特点。5.磨粒图像预处理、特征提取及识别算法研究磨粒图像定性分析包含三个基本问题:磨粒图像处理、特征提取和识别。在磨粒图像处理上,运用本文所提出的图像边缘检测、图像平滑等算法对磨粒图像进行预处理;在特征提取上,提取了正常磨粒等8种磨粒的尺寸参数、边界形状参数、结构特征参数、颜色特征参数及纹理特征参数,共计54个,使得对磨粒分析更加多样化;在磨粒识别上,根据灰色关联度对特征参数进行了精简和优化,确定出面积、周长、等效圆直径、圆形度、长短轴比、欧拉数、畸形度、梯度均值及灰度一梯度相关度9个特征参数来进行磨粒识别;提出了基于灰色关联神经网络的磨粒识别算法,以上述9个特征参数作为神经网络的输入,磨粒类型作为输出,利用灰色关联度来优化神经网络隐含层的神经元数目。该算法可以大大提高神经网络的学习效率,提高磨粒识别的分辨率和准确率,磨粒识别的准确率可达97.5%。更多还原

【Abstract】 Ferrography is one of important technologies in machine condition monitoring and fault diagnosis.It takes wear particle as its research objects.Machine states can be monitored and fault can be diagnosed by analyzing ferrography because wear particle image reflects the information of mechanical equipment’s wearing and tearing. The recognition of wear particle is the core of ferrography for its direct relation to the correctness of monitoring.With the rapid development of the computer technology, some methods have been applied in ferrography technique such as computer vision, expert system,artificial neural network.Since wear particle images have some grey characteristics such as the grey level of image pixel,image edge,image noise and the threshold of image segmentation and so on,the grey system theory has aroused researchers’ attention and its feasibility and validity in the field of image process becomes a brand-new area.In this thesis,grey system theory and its application in ferrographic image processing has been studied by the numbers,which conducts the research of ferrography wear particle image’s processing,feature extraction,image quality assessment as fellows:1.The research of enhancing grey prediction model precisionIt is shown by study that the factors affecting grey prediction model precision primarily are the smooth degree of primary data sequence,the background value and the initial value of grey model,in order to increase grey model precision,the research is conducted separately:（1） Based on function x^-a（a＞0） transformation and based on multiplex composite function transformation,the methods of enhancing smooth degree of data sequence are proposed respectively.It has been proved that the discrete data after transformation can greatly advance its smooth degree.Moreover,the property of transformation function is summarized. （2） The structuring method of background value based on the integral definition and the exponential function are put up,and the fitting and prediction precision is analyzed.The results show that these two kinds of improvement grey model may be used not only in short-term forecasting,medium-term but also in long-term forecasting.（3） The revision initial value and the time-variable initial value are advanced,and the method of its parameter identification based on the adaptive genetic algoritlm is given.In order to obtain the higher modeling and prediction precision,the grey model may be ameliorated from the sequence smooth degree,the background value and the initial value at the same time.2.The research of ameliorating grey correlation degreeThe research has discovered that the existing grey correlation degree has some flaws such as contradiction between the grey correlation four axioms and the computational methods of grey correlation degree,the influence of sequence’s non-dimensionalization against the computed result of correlation.Based on this,the grey average correlation degree and grey T’s correlation degree have been studied, and ameliorating grey Euclid correlation degree and grey T’s correlation degree are proposed separately:（1）The ameliorating grey Euclid correlation degree is given.It considers not only the correlation coefficient fluctuation of each spot to its mean value but also the ideal related and negative ideal related.It is proved that it has parallelism,standardization, integration,even symmetry and appropinquity.（2）The grey T’s correlation degree is supposed.It can reflect the positive and negative relation of sequence and has symmetry,uniqueness,comparability, appropinquity,standardization,and the rank preservation to the non-quantification processing.3.The research of image processing algorithm based on grey system theory.（1） The new three methods of edge detection are proposed according to grey correlation degree and derivative operator,stepwise ratio,higher-dilnension space and correlativity respectively.These algorithlns can detect the edges of different direction and adjust edge detail detected by means of threshold.These algorithms have suppression ability against some kinds of noise such as gaussian,speckle and poisson noise specially.Moreover,these algorithms can locate accurately and the computation load is small.（2） A new method of detection for noise spots is proposed based on grey correlation coefficients.This algorithm that distinguishes noise spots between noise image and mean image according to grey coefficients has used view image information including noise statistical information.It can distinguish the noise spots of image effectively.（3） Kinds of adaptive weighted filter are proposed based on grey correlation degree,grey model,mean image and median image.These algorithms take mean image or median image of noised image as its foundations,and use grey correlation degree and grey forecast model to process noise spots separately.These methods can overcome noise spot’s influence effectively and reduce the image fuzziness,it can preserve integrity of edge.（4） Based on lifting wavelet transform and grey prediction model,an image compression algorithm is proposed.Firstly,this algorithm transforms image to frequency range by using promotion wavelet,and differentiates significant coefficients and insignificant coefficients by using zero-tree in various wave bands. Moreover,it scans various wave bands to one dimension by using Hilbert curve. Finally,it carries on predictive coding by using grey forecast model.The simulation shows that this algorithm can enhance image compression ratio and compression quality effectively.4.The research of the image quality assessment based on the correlativity.Based on the characteristics of wavelet coefficients of image and the correlativity index,a novel image quality assessment is proposed.The algorithm makes full use of perfect integral comparison mechanism of correlativity index,it can not only evaluate the quality of image accurately but also bears more consistency with human visual system. 5.The research of wear particle image pre-processing,feature extraction and recognition algorithm.This thesis makes a review of achievement in the field of ferrography especially in image processing,feature distilling and particle recognition.On the image processing,the wear particle image pre-processing can be done by means of image edge detection and image smoothing which are proposed by this thesis.On feature extraction,about 54 parameters of characteristics are extracted.On wear particle recognition,the characteristic parameters are simplified and optimized by using grey correlation degree to carry on the wear particle recognition.The wear particle recognition algorithm is proposed based on grey correlation neural network.The neural network takes the 9 above-mentioned characteristic parameters and wear particle type as its input and output respectively,and the neuron number of hidden layer is optimized by use of grey correlation degree.The algorithm can optimize the structure of neural network greatly,enhance network’s study efficiency and the accuracy of wear particle recognition.Experiment result shows that the classification accuracy is more than 97.5%.更多还原