1 Introduction
Si3N4 ceramic materials have high mechanical properties such as low coefficient of thermal expansion, heat resistance, corrosion resistance and wear resistance[1]. Ceramic materials also have thermal, electrical and chemical properties [2]. As a result, ceramic bearings are widely used in aerospace engines, biomedical devices and transportation. Since the process of ceramic products is susceptible to factors such as mixing and sintering, they are prone to external defects [3]. The main types of defects are pit, crack, wear and snowflake [4], which affect the performance of ceramic materials. The defects of ceramic products are relatively small and difficult to be recognized by the human eye. At present, the inspection methods for surface defects of ceramic materials at home and abroad are mainly divided into two methods: manual inspection and nondestructive testing. Manual inspection method test results have the disadvantages of low detection efficiency, high false detection rate, randomness, and poor stability. Non-destructive testing (NDT) [5-7] is a method for detecting the type, number, shape, location and size of surface defects in ceramic materials without changing the surface or internal structure of the ceramic.
With the rapid development of artificial intelligence (AI) and machine learning, convolutional neural networks(CNNs) have been widely used in computer vision. Such as object detection, student teaching[8], health analysis, etc. Aiming at the vehicle navigation light guide plate (LGP) image characteristics, Li et al[9] proposed proposed a visual detection method based on improved RetinaNet. They proposed and used an improved feature pyramid network module to improve the feature fusion network in the retinal network. Experimental results show that the method is effective. In the vehicle LGP data set, the average detection rate is 98.6%. Mingming Zhu [10] et al. proposed a RetinaNet-based method for detecting arbitrarily oriented ships. By a rotated RetinaNet, a refined network, a feature alignment module, and an improved loss function, rotated detection, achieving better detection progress and solving boundary discontinuity are achieved to locate ships with high accuracy. Yantong Chen [11] et al. proposed a fly species recognition method based on the refined RetinaNet and the convolutional block attention module (CBAM). A multi-scale feature pyramid was constructed. Kullback-Leibler (KL) loss superseded smoothed L1 loss for simultaneously learning bounding box regression and localization uncertainty. The method achieves an average accuracy (mAP) of 90.38%, which is better than existing methods. The average time to identify a single image is 0.131s.
Combining NDT techniques with deep learning methods to process Si3N4 ceramic bearing inner ring surface defect images, online data augmentation is used to expand the images and form the dataset. Attention mechanisms SE, CBAM and NAM are added to the tail of the Resnet-50 feature extraction network, respectively. It is found that the NAM attention mechanism improves the model recognition and classification accuracy most significantly by comparison. Improved loss function replaces Smooth L1 loss for learning bounding box regression and uncertain localization. The Si3N4 ceramic bearing inner ring surface defect recognition and classification model is obtained by training the network for testing surface defects. The main contributions of this paper are as follows: ⑴the platform for surface defects of Si3N4 ceramic bearing inner ring is built independently to collect surface defects images, ⑵the online data augmentation is used to expand the surface defect dataset, ⑶adding NAM to obtain a more complete attention map and better global information of the image. The results show that the proposed method has better performance compared with the Faster RCNN and RetinaNet.
The rest of this paper is organized as follows. Chapter 2 introduces the platform for surface defects of Si3N4ceramic bearing inner ring. Chapter 3 presents the Si3N4 ceramic bearing inner ring surface defect analysis and dataset fabrication. Chapter 4 introduces the Si3N4 ceramic bearing inner ring surface defect identification and classification model. Chapter 5 outlines the experimental results and analyzes the feasibility of the proposed method. Chapter 6 presents the conclusion.