Design And Implementation Of Roller Bearing Surface Defect Detection Algorithm Based On Improved Convolutional Neural Network

Roller bearings are key parts in the operation of railway freight cars.Their daily working conditions combine three strict requirements of high-speed rotation,high bearing capacity and high stability,which directly affect the safe and stable running of railway freight cars.It is one of the daily tasks of railway support stations in various places to regularly inspect the surface of roller bearings for defects.Through participating in the field investigation of the operation of the railway freight car support station,it was found that the existing detection method for bearing defects is mainly manual inspection.After analysis,there are mainly the following reasons:(1)Due to the high technical requirements of domestic railway freight car bearing standards,the detection rate of actual defects is less than 0.1%,so the number of samples of various defects is low,resulting in insufficient training sets in traditional convolutional neural network learning.Therefore,data augmentation technology is one of the key issues to be solved in this paper.(2)The guarantee station involved in this article mainly needs to detect the353130 B bearing(hereinafter referred to as the bearing).The structure of the bearing is complex,and the shapes of each part are different,and the shape is quite different.It is difficult for traditional computer vision technology to use the same solution to deal with multiple surfaces.defect detection.(3)The railway freight car support station uses industry standards to judge the appearance defects of bearings.The standard records 21 types of defects.The data returned from the support station shows that under normal circumstances,defects account for 0.03% of the bearing surface area.Therefore,the detection of multiple nonobvious objects and recognition bring challenges to the existing computer vision processing technology.In order to solve the above three problems,this paper designs and implements an improved convolutional neural based target detection algorithm FCAM-YOLOv5(Four Detection Layer Coordinate Attention Mobil Netv3 YOLOv5).The main design ideas and implementation contents of FCAM-YOLOv5 are as follows:(1)Use methods such as flipping,contrast enhancement,and mixing public data sets to perform data augmentation operations on the original data set to solve the problem of difficult convergence and low accuracy of traditional convolutional neural networks under small sample data sets.Use the Gaussian fuzzy data augmentation method for some data to increase the amount of data while preventing the model from deteriorating robustness due to the large gap in the amount of various types of data.(2)FCAM-YOLOv5 adds a CA(Coordinate Attention)attention mechanism to the traditional target model,and adds a new detection layer to strengthen the detection of small targets.By replacing the feature extraction network in the target detection algorithm with the lightweight network Mobile Netv3,the detection speed of the model is increased,and the industry standard detection index is initially reached,and the automatic detection of bearing surface defects can be completed on the assembly line.(3)Using neural network technology,it abandons the shackles of traditional computer vision technology on defect surface requirements and defect ratios,so that normal detection can also be performed on curved surfaces.Field detection results and experimental data show that FCAM-YOLOv5 is better than traditional computer vision detection methods and has a wider range of application scenarios.The detection speed and detection accuracy of FCAM-YOLOv5 are better than the original YOLOv5 model,and the field test results also show that the missed detection rate of FCAM-YOLOv5 is significantly lower than the original YOLOv5.This paper forms a proprietary data set SKF-KS2022,which includes a total of4900 bearing defect images of 4 types of defects.The ablation experiment shows that the m AP(mean Average Precision)of the improved convolutional neural network on the basis of the data set reaches 0.8587,which is 0.06 higher than the original model,and the detection speed of each image reaches 54 ms,which is 17 ms higher than the original model,the experimental data show that the detection rate of the FCAMYOLOv5 model has initially reached the industrial detection standard.