一种零件综合质量评定方法研究

零件的质量评定是柔性智能制造中十分重要的环节。现有的自动化识别装置一般采用非人工接触的光学检测系统，但由于工况环境复杂，诸多干扰因素均会影响零件质量检测与评定的准确性。另外，工业现场的连续作业对工控机硬件的运行速度、光学检测系统的环境适应性以及质量评定算法的预测准确性都提出了更高的要求。基于此，提出一种基于机器视觉与机器学习的零件综合质量评定方法。首先，借助机器视觉技术完成被测零件图像的实时采集与处理，并利用灰度匹配算法与几何匹配算法对零件的图像与CAD（computer aided design，计算机辅助设计）机械加工图进行比较，求解得到灰度匹配分数与几何匹配分数这2个几何特征参数。然后，针对零件表面的缺陷（如划伤、磨损、边缘缺料及锈蚀等），在图像预处理（灰度化、图像增强、高斯降噪和二值化）的基础上，求解得到图像灰度的均值和标准差这2个表面缺陷特征参数。最后，借助主成分分析（principal component analysis, PCA）对零件的四维特征数据集进行降维处理，并利用K最近邻（K-nearest neighbor, KNN）算法对降维后的数据集进行训练和预测，完成零件综合质量评定；在此基础上，比较KNN算法与其他机器学习算法的准确率、召回率和特异度等指标，以验证其可行性。实验结果表明，所搭建的光学检测与处理系统在不同光源条件下的识别准确率达到96.15%以上；当相机的快门时间设定为100 μs时，该系统的图像处理速度达到45.2 帧/s。所提出的零件综合质量评定方法具有较高的准确率与处理速度，适用于复杂工况下零件的综合质量评定。

Research on comprehensive quality assessment method for parts

The quality assessment of parts is a very important link in flexible intelligent manufacturing. The existing automatic identification device generally adopts the non-artificial contact optical detection system, but due to the complex working environment, many interference factors affect the accuracy of quality detection and assessment for parts. In addition, the continuous operation of industrial site puts forward higher requirements on the running speed of industrial control machine hardware, the environmental adaptability of optical detection system and the prediction accuracy of part quality assessment algorithm. Based on this, a comprehensive quality assessment method for parts based on the machine vision and machine learning was proposed. Firstly, the real-time acquisition and processing of the measured part image was completed with the help of machine vision technology, and then the gray matching algorithm and geometric matching algorithm were used to compare the images and the CAD (computer aided design) machining drawings of parts, so as to solve the geometric feature parameters of gray matching score and the geometric matching score. Then, according to the surface defects of parts (such as scratch, wear, edge material shortage and rust), the surface defect feature parameters of mean and standard deviation of image gray were solved on the basis of image pretreatment (gray, image enhancement, Gaussian noise reduction and binary). Finally, the four-dimensional feature data set of parts was dimensionally reduced by the principal component analysis (PCA), and the K-nearest neighbor (KNN) algorithm was used to train and predict the data set after dimension reduction to complete the comprehensive quality assessment for parts. On this basis, the accuracy, recall rate, specificity and other indicators of the KNN algorithm and other machine learning algorithms were compared to verify its feasibility. The experimental results showed that, the recognition accuracy of optical detection and processing system was more than 96.15% under different lighting conditions; when the camera shutter time was set to 100 μs, the image processing speed of this system reached 45.2 frames/s. The proposed comprehensive quality assessment method for parts has high accuracy and processing speed, which is suitable for comprehensive quality assessment of parts under complex working conditions.