【作者】 王常青；

【导师】 张勇传； 周建中；

【作者基本信息】 华中科技大学 ， 系统分析与集成， 2012， 博士

【摘要】 随着科学技术的不断发展,旋转机械设备的结构越来越复杂,造价越来越高,各部分之间的联系更加紧密,因而发生故障的风险也在逐渐加大。及时准确的提取表征旋转机械设备运行状态的信息,从而捕捉其中的故障信息并加以识别判断,对发现旋转机械设备的异常,提高旋转机械本身运行的可靠性有着重要的意义。轴心轨迹是旋转机械故障诊断的重要手段之一,反映了转子旋转时轴上任一点在其旋转平面内相对轴承座的运行轨迹,它包含了机组的各种故障信息,因此,其形状特征对判断旋转机械转子轴系故障非常重要。从这一问题入手,可将故障诊断问题转化为图像处理、分析与识别问题,传统的方法多集中在以傅里叶变换、小波变换和矩为基础的特征提取的研究上,而图像的特征提取寻求的最佳效果是用较少的数学描述来表达图像中重要的信息,将目前数字图像处理与分析的先进理论与方法引入轴心轨迹的特征提取与分类识别,无疑将为旋转机械故障诊断提供一个新的研究思路。论文的主要工作及创新性成果如下：(I)针对旋转机械轴心轨迹自动识别的研究需求,深入研究了数字图像模型的描述、基本运算、图像的变换、图像分割,以及目标的表示与描述方法和图像的模式识别理论,为轴心轨迹自动识别研究打下了基础。(2)在对轴心轨迹的几何特征如面积、周长、圆度和离散指数等进行分析的基础上,针对传统的链码不具有旋转、平移和尺度不变性的问题,提出了改进的链码直方图。然而,采用单一的改进链码直方图对轴心轨迹特征进行提取,椭圆形和外“8”字形存在特征向量相同的情况,给分类识别增加了难度。因此,本文更进一步的提出了基于边界描述的改进链码直方图和形状数融合的轴心轨迹特征提取方法,同时引入概率神经网络学习几何特征和轴心轨迹类型之间的映射关系,从而将训练好的识别器用于轴心轨迹识别。此外,对改进链码直方图与其他几何信息融合的特征提取方法进行了研究,实验结果表明,将改进链码直方图和形状数融合的特征提取方法优于其他融合形式,具有较高的识别率。(3)针对旋转机械振动信号常常伴有噪声导致轴心轨迹曲线不光滑的情况,利用脉冲耦合神经网络模型源于哺乳动物视觉皮层神经细胞的研究,并且不需要训练的固有属性,提出了脉冲耦合神经网络时间序列与轴心轨迹的圆度融合的特征提取方法,得到用于进行轴心轨迹图像识别的特征向量二值图像。此外,提出了基于脉冲耦合神经网络信息熵时间序列的轴心轨迹特征提取方法,将两种方法提取的特征向量分别通过概率神经网络和径向基函数神经网络进行分类识别,实验结果表明脉冲耦合神经网络时间序列与轴心轨迹的圆度融合的特征提取方法能够更精确的表达轴心轨迹的原始信息,而概率神经网络在多类别识别方面的能力优于径向基函数神经网络。(4)利用Contourlet变换能够在任意尺度上实现任意方向的分解,并且擅长描述图像中的轮廓信息,仅用少量系数即可有效的表达图像中的边缘轮廓的特点,提出了将小波变换与Contourlet变换融合的轴心轨迹特征提取方法,将小波变换对于点奇异性的良好稀疏性与Contourlet对线性特征的良好检测性相结合,精确的描述轴心轨迹的形状信息,为轴心轨迹的自动分类提供一种新的特征向量,也为旋转机械故障诊断自动化提供了新的思路。引入支持向量机对特征向量进行分类,使得轴心轨迹的自动分类过程变得更加快速、有效。更多还原

【Abstract】 With the continuous development of science and technology, the structure of the rotating machinery and equipment grows more and more complicated, increasingly close ties between the different parts lead to the increasingly high manufacturing cost, etc., the risk of failure is also gradually increased. It is necessary to sample the state information of the machinery timely and accurately, then identify the working state of the machinery, thus to improve the reliability of the rotating machinery itself.Shaft orbit is an important means for rotating machinery fault diagnosis. It is synthesized by vibration signal, and reflects various fault information of rotating machinery, therefore, the shape characteristics of shaft orbit is very necessary for diagnosis shaft fault of rotating machinery. Based on these, the fault diagnosis problem can be transformed into the problems of image processing, analysis and identification. Traditional method is mainly concentrated in the research on the Fourier transform, wavelet transform, and moment-based feature extraction. But image feature extraction is to seek the best results with less mathematical description to express important information in the image. Therefore, it is undoubtedly effective to introduce a method of digital image processing for the feature extraction and identification of shaft orbit.The main contents and innovative results are listed as follows:(1) Considering the needs of automatic identification for the rotating machinery, the thesis establishes the foundation for shaft orbit automatic identification research, with digital image model description, basic computing, image transformation, image segmentation, target representation and description, and image pattern recognition theory.(2) By analyzing shaft orbit geometric characteristics, such as area, perimeter, roundness, and the dispersion index, the thesis put forward a method of modified chain code histogram. The method avoids the shortcomings in traditional way as rotation, translation and scale sensitive. However, it is not good for shaft orbit in ellipse and the outer "8" to only use the modified chain code histogram to extract the characteristics. It is difficult for classification and identification the state of the machinery in the case. Therefore, the thesis describes an improved extraction method based on boundary described and shape number. Meanwhile, the thesis uses probabilistic neural network to learn the mapping between the geometric characteristics and the type of shaft orbit. In addition, the fusion feature extraction methods of modified chain code histogram and other geometric information are described. The experimental results show that modified chain code histogram and shape number fusion feature extract method is better than the other fusion form, and have a high recognition rate.(3) Rotating machinery vibration signals are often associated with noise, which cause the shaft orbit curve not smooth. Therefore, feature extraction must have noise immunity. By pulse coupled neural network model derived from the study of the mammalian visual cortex neurons, the thesis improved pulse coupled neural network time signature and the roundness of shaft orbit fusion feature extraction method without training. In addition, the pulse coupled neural network information entropy time signature for the shaft orbit feature extraction is further used. Then, the feature vectors of two methods are all send to probability neural network and radial basis function neural network for classification. The experimental results show that the pulse coupled neural network time signature and shaft orbit roundness fusion feature extraction method can express original shaft orbit more accurate. And the probability neural network is superior to radial basis function neural network in multi-class recognition.(4) Contourlet transform can realize any direction decomposition at any scale, and work well at description of the image contour information. It can effective express the image contour by only using a small amount of the coefficient. The study improves wavelet transform and contourlet transform fusion shaft orbit feature extraction method. The method combines the good sparsity for singularity of wavelet transform and good detect ability for linear feature of contourlet transform to accurate description of the shape of the shaft orbit. It provides a new feature vector for shaft orbit automatic classification, also provides a new approach for rotating machinery fault diagnosis automation. The introduction of support vector machines to classify the feature vector, the shaft orbit automatic classification process becomes more rapid and effective.更多还原