基于多重分形的水泥回转窑工况识别研究

鉴于多重分形理论在定量描述复杂系统非线性运行规律方面具有的独特优势,将多重分形理论引入到工况特征识别研究中来,确认了水泥回转窑窑电流信号的多重分形特性.在此基础上,研究了窑电流多重分形谱参数随工况变化的情况,发现多重分形谱参数的变化趋势与回转窑内工况状态的变化趋势之间具有较强的关联性,进而提出了基于多重分形谱参数进行水泥回转窑异常工况特征提取的新方法,为水泥生产过程中工况状态的在线监控和预报提供了有力支持.     

基于形态谱的磨粒图像特征参数提取方法

为提高磨粒识别的精度,提出一种基于形态谱磨粒图像特征参数提取新方法,给出磨粒图像的归一化形态谱的计算方法,并将磨粒的形态谱作为其特征向量,采用径向基函数神经网络对磨粒进行自动识别。结果表明:利用磨粒的形态谱实现了对球形磨粒、切削磨粒、严重滑动磨粒、疲劳剥块4种典型磨粒的分类识别,磨粒的形态谱可以作为磨粒的有效特征参数。     

基于多重分形的摩擦振动信号研究

为提取摩擦振动的特征信号和实现摩擦振动特征信号的表征,在摩擦磨损试验机上进行船用柴油机缸套-活塞环摩擦副摩擦磨损模拟试验,应用总体经验模式分解对缸套-活塞环摩擦副的非线性、非平稳的摩擦振动信号进行降噪,并应用多重分形对摩擦振动信号进行分析,得到多重分形谱图及其参数。研究结果表明,随着摩擦磨损试验的进行,缸套-活塞环摩擦副摩擦振动信号振幅分布多重分形谱和频率分布多重分形谱呈现一定的变化规律;多重分形谱图及其参数能够实现摩擦振动信号特征的表征,可以用于摩擦副摩擦振动状态的识别。     

滚动轴承振动信号的非线性分形特性研究北大核心

为研究球轴承振动信号的非线性分形特性,对Hurst指数及多重分形计算分形特征的方法进行改进。采用2次分割充分利用数据,并对不同状态振动的Hurst指数和多重分形谱的变化规律进行了仿真及试验研究。结果表明:正常状态振动呈现正的持续性,故障状态振动具有反持续性,故障状态振动的波动影响范围比正常状态振动的小;振动信号具有不均匀的自相似性(即多重分形特征),振动状态与多重分形谱的形状之间存在着对应关系。     

磨粒分形识别及发展

相互作用表面间必然会产生磨粒,磨粒含有大量的有关材料摩擦磨损的信息。磨粒形态分析是确定磨损方式和磨损程度的有益手段。磨粒并非是欧氏几何体,而是展示出了分形性质。基于分形几何理论,可获得尺度不变的分形参数,用这类参数可对磨粒形态进行客观、全面的表征。本文综合评述了磨粒分形表征以及磨粒形态与磨损方式、磨损程度间的定量耦合关系等的研究进展,对将来磨粒分形研究的趋势和注意的问题进行了探讨。     

缸套-活塞环磨合过程摩擦力矩多重分形研究

针对缸套-活塞环磨合模拟试验过程中摩擦力矩的变化特点,提出一种用多重分形谱描述摩擦力矩信号复杂性的新方法。给出了摩擦力矩信号多重分形谱的计算方法,并对活塞环磨合过程中摩擦力矩信号多重分形谱的变化规律进行了试验研究。结果表明:随着磨合过程的进行,多重分形谱的宽度Δα值呈上升趋势,与活塞环磨损失重的变化规律是一致的;多重分形谱描述了摩擦副表面的动态变化过程,运用多重分形谱可有效地评价磨合磨损过程。     

基于改进MF-DFA的液压泵退化特征提取方法

针对液压泵振动信号通常具有非线性强和信噪比低的特点,提出了一种基于改进多重分形去趋势波动分析(multi-fractal detrended fluctuation analysis,简称MF-DFA)的液压泵性能退化特征提取方法。首先,引入滑动窗口技术改进传统MF-DFA方法在时间序列数据分割过程中存在的不足,提高了MF-DFA方法的计算精度;然后,利用改进的MF-DFA方法计算液压泵多重分形谱参数,分析了不同分形谱参数对液压泵退化状态的反映能力,选取奇异指数α0和多重分形谱宽度Δα作为退化特征量;最后,以液压泵不同退化状态下的实测数据为例验证了该算法的有效性。试验结果表明,该方法能够准确提取液压泵退化特征,提高了退化状态识别的准确率。     

滚动轴承振动信号的非线性分形特性研究

为研究球轴承振动信号的非线性分形特性,对Hurst指数及多重分形计算分形特征的方法进行改进。采用2次分割充分利用数据,并对不同状态振动的Hurst指数和多重分形谱的变化规律进行了仿真及试验研究。结果表明:正常状态振动呈现正的持续性,故障状态振动具有反持续性,故障状态振动的波动影响范围比正常状态振动的小;振动信号具有不均匀的自相似性(即多重分形特征),振动状态与多重分形谱的形状之间存在着对应关系。     

局部Walsh谱描述图像纹理特征的旋转不变性

为了有效地进行纹理分析,提出一种基于局部Walsh谱的纹理特征旋转不变性描述方法。首先,比较每个像素点与其邻近点的灰度值生成局部二值序列,并计算序列离散Walsh变换的功率谱;然后,采用功率谱的各谱点值构成特征直方图描述纹理特征;最后,从序列的列率特性出发,构造了新的两族局部Walsh谱,揭示了局部Walsh谱与局部二值模式之间的联系。因为离散Walsh变换功率谱具有循环移位不变性,所以局部Walsh谱具有先天的旋转不变性。实验结果显示,与灰度共生矩阵和Gabor滤波器组相比,局部Walsh谱的纹理分类准确率较高;与局部二值模式相比,在相同尺度下局部Walsh谱的分类准确率比其高出3%以上,对两幅旋转纹理图像分割的错误率比其低11%和3%,表明提出的方法具有较好的纹理鉴别能力和旋转不变性。     

曲波域铁谱磨粒图像特征提取方法研究

曲波变换具有多尺度分析能力,与小波变换相比可更好地表达图像的曲线特征.为有效描述铁谱磨粒的形貌特征,提出一种曲波域图像特征提取方法.利用曲波变换将磨粒图像进行分解,得到不同尺度的曲波系数;根据曲波系数统计分布特点,采用广义高斯分布模型对细尺度和精细尺度曲波系数分布进行建模;提取粗尺度曲波系数的均值、标准差、能量和熵等统计特征,以及细尺度和精细尺度曲波系数的广义高斯分布模型参数描述磨粒特征.将提取的特征用于发动机典型磨粒识别,识别成功率达到了88.9%,表明该方法所提特征能很好地表达铁谱磨粒的形貌特征.     

多层次信息融合在铁谱图像磨粒识别中的应用

针对铁谱图像磨粒识别中异类信息综合利用率较低的问题,提出多层次信息融合的铁谱图像磨粒识别方法。首先,在铁谱图像二值化分割的基础上进行二值滤波,结合彩色铁谱图的R、G、B三分量,实现铁谱图像的彩色滤波。其次,以实际采集的磨粒图像样本为例,提取滤波后二值图像的形态特征,以及滤波后彩色图像的颜色特征;在特征层利用PCA对异类特征进行维数约简,并结合SVM和k-fold交叉验证,实现形态特征和颜色特征的特征层融合;在决策层将异类特征的SVM概率输出结果作为D-S证据理论的基本概率分配函数,实现形态特征和颜色特征的决策层融合。通过与形态学滤波结果对比,验证了本文提出滤波方法的优越性;其次,不同层次的信息融合结果表明,与单独使用颜色特征和形态特征相比,异类信息融合后可实现优势互补,有效提高故障磨粒的识别准确率。     

Two-dimensional fast Fourier transform and power spectrum for wear particle analysis

Statistical parameters, such as R_a and R_q, have been widely used to investigate the roughness of wear particle surfaces in the literature. It has been reported that wear particle analysis based only on numerical characterization is often insufficient to distinguish certain types of wear debris. In this study, two-dimensional fast Fourier transform, power spectrum and angular spectrum analyses are applied to describe wear particle surface textures in three dimensions. Laminar, fatigue chunk and severe sliding wear particles, which have previously proven difficult to identify by statistical characterization, have been studied. The results show that spectral analysis effectively identifies the surface texture pattern (e.g. isotropy or anisotropy) and can be applied to classify these three types of wear particles.     

Automated classification of wear particles based on their surface texture and shape features

In this study, the automated classification system, developed previously by the authors, was used to classify wear particles. Three kinds of wear particles, fatigue, abrasive and adhesive, were classified. The fatigue wear particles were generated using an FZG back-to-back gear test rig. A pin-on-disk tribometer was used to generate the abrasive and adhesive wear particles. Scanning electron microscope (SEM) images of wear particles were acquired, forming a database for further analysis. The particle images were divided into three groups or classes, each class representing a different wear mechanism. Each particle class was first examined visually. Next, area, perimeter, convexity and elongation parameters were determined for each class using image analysis software and the parameters were statistically analysed. Each particle class was then assessed using the automated classification system, based on particle surface texture. The results of the automated particle classification were compared to both the visual assessment of particle morphology and the numerical parameter values. The results showed that the texture-based classification system was a more efficient and accurate way of distinguishing between various wear particles than classification based on size and shape of wear particles. It seems that the texture-based classification method developed has great potential to become a very useful tool in the machine condition monitoring industry.     

基于yolov5在线可视铁谱图像磨粒多目标识别方法研究

鉴于在线图像可视铁谱获取的磨粒谱片图像分辨率低,磨粒种类复杂多变,磨粒图像背景复杂等问题,使得磨粒在线智能识别面临挑战。为了实现在线可视铁谱图像磨粒多目标实时检测与识别,提出基于yolov5在线可视铁谱图像磨粒多目标识别方法。以正常磨损磨粒、疲劳磨损磨粒、滑动磨损磨粒、球形磨粒、氧化磨损磨粒、切削磨损磨粒6种磨粒作为研究对象,基于yolov5深度神经网络模型对复杂油液环境下的异常磨损磨粒进行分割与识别。结果表明：基于yolov5算法的磨粒智能识别模型能够实现复杂环境下多目标、多类型磨粒在线实时识别,其识别速度和准确率基本满足油液在线监测需求,为装备在线图像可视铁谱技术工业化应用提供了技术支撑。     

Shape and texture features in the automated classification of adhesive and abrasive wear particles

In this study the automated classification system, developed previously by the authors, was used to classify wear particles. Two kinds of wear particles, adhesive and abrasive, were classified. The wear particles were generated using a pin-on-disk tribometer. Various operating conditions of load, sliding time and abrasive grit size were applied to simulate adhesive and abrasive wear of different severity. SEM images of wear particles were acquired, forming a database for further analysis. The particle images were divided into eight groups or classes, each class representing different wear test conditions. All eight particle classes were first examined visually. Next, area, perimeter and elongation parameters were determined for each class and the parameters were statistically analysed. The automated classification system, based on particle surface texture, was then applied to all particle classes. The results of the automated particle classification were compared to those based on either the visual assessment of particle morphology or numerical parameter values. It was shown that the texture-based classification system was a more efficient and accurate way of distinguishing between various wear particles than classification based on size and shape of wear particles. It seems that the texture-based classification method developed has great potential to become a very useful tool in the machine condition monitoring industry.     

Automatic wear‐particle classification using neural networks

Although the study of wear debris can yield much information on the wear processes operating in machinery, the method has not been widely applied in industry. The main reason is that the technique is currently time consuming and costly due to the lack of automatic wear particle analysis and identification techniques. In this paper, six common types of metallic wear particles have been investigated by studying three‐dimensional images obtained from laser scanning confocal microscopy. Using selected numerical parameters, which can characterise boundary morphology and surface topology of the wear particles, two neural network systems, i.e., a fuzzy Kohonen neural network and a multi‐layer perceptron with backpropagation learning rule, have been trained to classify the wear particles. The study has shown that neural networks have the potential for dealing with classification tasks and can perform wear‐particle classification satisfactorily. This revised version was published online in June 2006 with corrections to the Cover Date.     

基于改进支持向量机的磨粒铁谱识别的研究

提出了一个基于改进支持向量机的磨粒模式识别系统。该系统首先对磨粒的铁谱分析图像进行预处理，然后提取其特征参数，最后利用支持向量机对磨粒所属的类型进行分类。