DEVELOPMENT OF A DIAGNOSTIC SYSTEM FOR SLIDING SURFACES IN MACHINERY

该项目旨在开发一种系统，该系统可根据油样中所含的磨损碎屑预测机械滑动表面的状况。建立了磨粒显微图像的计算机分析技术和磨粒特征与滑动条件的人工神经网络关联技术。该系统已在钢的润滑磨损试验、齿轮试验和一些实际机床上得到应用和检验，并研制了一台计算机控制的图像分析仪，不仅能在显微镜上采集和分析图像，而且能自动获取图像。它还可以进行三维形状测量的大碎片。介绍了表征颗粒形状和表面特征的几种描述符，包括代表性直径、延伸率、圆度和修正圆度、反射率、厚度、傅立叶描述符和表面纹理参数。提出了平均三个参数的最佳样本容量。傅立叶描述符被认为是不太容易比其他形状参数所造成的图像digitalization.Wear碎片的错误被证明有不同的形态和表面特征，这取决于滑动的阶段，负载，滑动速度，油和添加剂，和滑动材料。利用人工神经网络分析和表征了岩屑特征与滑动条件的相关性。介绍了反向传播网络、Hopfield网络和Kohonen自组织映射，其中反向传播网络应用最为广泛。研究表明，神经网络只需输入碎片参数即可学习碎片特征和相关条件，并预测碎片来自的条件以及摩擦系数。该系统被证明是能够检测，从磨屑的特性，从初始磨损的变化，稳态磨损，并进一步剥落或齿轮试验中的得分。

This project was aimed at developing a system which predicts conditions of sliding surfaces in machinery from wear debris contained in sampled oils. Techniques for analyzing microscopic images of wear debris with computers, and association of debris features with sliding conditions using artificial neural networks were established. The system was applied and tested with lubricated wear experiments of steels, and also applied to gear tests and some practical machines.A computer controlled image analyzer was developed which not only captures and analyzes images on the microscope but is able to obtain images automatically. It can also conduct three-dimentional shape measurement of large debris. Several descriptors of shape and surface features of particles were introduced, including representative diameter, elongation, roundness and modified roundness, reflectivity, thickness, Fourier descriptors and sueface texture parameters. The optimum sample size for averaging there parameters was proposed. Fourier descriptors were found to be less susceptible than other shape parameters to errors caused by image digitization.Wear debris were shown to have different morphological and surface features depending on the stages of sliding, loads, sliding speeds, oils and additives, and sliding materials. Correlations of the debris features and the sliding conditions were analyzed and represented with artificial neural networks. The back-propagation net, Hopfield net, and Kohonen self-organizing map were introduced, among which the back-propagation nets were extensively used. It was shown that the neural nets learned debris features and associated conditions, and predicted conditions the debris came from as well as the coefficient of friction with only inputs of debris parameters. The system proved to be capable of detecting, from wear debris characteristics, changes from initial wear to steady-state wear, and further to flaking or scoring in gear tests.