Ceramic Bearing Ball Surface Defects Inspection Based On Computer Vision

Homogeneous and fully densified silicon nitride has combinational properties that are suitable for rolling contact applications and has great future in aerospace applications. But as a type of brittle material, surface defects may be introduced in each processing step. These surface defects reduce the quality and reliability of ceramic bearing ball and thus limit wider applications of hybrid bearings. To ensure the reliability of hybrid bearings in important applications, this paper analyzes the formation and propagation of these defects and develops an experimental surface inspection system based on computer vision for ceramic bearing ball.In this paper, three kinds of silicon nitride ball blanks with different diameters are lapped and polished, and statistical data are established. Surface defects are classified into five columns according to their morphologies under optical microscope. All defects are examined using SEM and the formation mechanism of flaws and cracks were interpreted with fracture mechanics of ceramics and fractography. The results show that irregular lapping abrasives produce flaws. Various cracks are mainly produced by upper lapping plate. Wear and scratch defects are caused by false lapping load and hard abrasive particles in final process. Better blank ball roundness and lower lapping load and speed at early lapping stages may reduce various cracks. Ceramic bearing balls are generally classified into high quality for extreme condition applications, acceptable quality for common condition applications and inacceptable for any applications according to the type and size of defects.Fractographic and metallographic analyses is performed to study the potential relationships between material microstructures and surface defects. The results show that the volume fraction of pores, the size of inclusions and grain diameter are key parameters to ceramic bearing ball surface lapping quality which can be improved by optimizing the microstructure characters.According to the geography properties of surface defects and present human inspection method, an automatic ceramic bearing ball surface inspection system based on computer vision is developed. An preprocessing algorithm for obtained surface images based on median filtering, top-hat transformation and logarithmic transformations is set up to eliminate uneven lighting disturbance and thus reduces the difficulties of defects segmentation. An automatic seeded region growing algorithm based on histogram concavity analyses and minimum inner-class cross entropy theory is developed for defects segmentation. The results show that this segmentation algorithm fully utilizes the gray level continuous properties of defects and overcomes the weak anti-jamming capabilities of traditional methods. This new segmentation method is robust for ceramic ball surface defects inspection.Region shape features of defects are computed after segmentation. Comparative experiments show that the first seven normalized Fourier descriptors, defects areas, ratio of length to width of the biggest object and number of objects are the best descriptors to distinguish flaw, crack, snowlike and scratch defects. Because of the discontinuous character of wear after segmentation, textile features of original image are directly computed. Comparative tests suggest that intensity histogram descriptors are the best features to present wear defects among textile features. Two BP Neural Network classifiers were established and trained according to these two groups of features. In experiments, 97.5% of the samples is classified correctly by the intensity histogram descriptor classifier and 90.3% is classified correctly by the region shape descriptor classifier.This paper integrates the software and hardware into the ceramic bearing ball inspection system. The theoretical resolution of this system reaches 4.85μm. Mathematic and control models are constructed for various ball surface unfolding methods. Sample images were captured and processed by this real system. In this way, 85.56% of the defects is classified correctly. Up to now, ceramic bearing balls fromΦ10 toΦ14 can be inspected by present system. Additionally, system reliability is analyzed and it is concluded that image quality and illumination constancy, as well as mechanism precision greatly influence the inspection accuracy. Improving the performance of classifiers is effective method to reduce the error rate of the inspection system.