Design And Implementation Of Wheel Defect Recognition System Based On Deep Learning

With the change in people’s travel preferences,the number of vehicles has been increasing,and the automotive industry has become an important pillar of the national economy.The demand for automotive wheel hubs has also been growing.Wheel hub defect recognition is an important process in wheel production and is typically achieved through X-ray nondestructive testing.However,the traditional approach of defect recognition through manual visual inspection of X-ray scanned images suffers from issues such as low accuracy and efficiency.Therefore,in this study,a deep learning-based wheel defect recognition system is designed and implemented for a specific wheel production process in a company.This system assists human operators in identifying wheel defects,thereby improving the recognition accuracy and efficiency to a certain extent.The main research contents of this paper are as follows:(1)Construction of the WheelXray Dataset:The WheelXray wheel defect recognition dataset is constructed.Based on DICOM format wheel images obtained from X-ray scans,145 images are obtained by converting the format using dicom2jpg and annotating casting defects using labelme.To address the issues of sample imbalance and insufficient quantity in the wheel recognition scenario,data augmentation is performed.A total of 1,315 wheel samples are selected from 2,727 casting X-ray image samples in the publicly available GDXray dataset.The images are adjusted with padding to match the dimensions and annotation format of the existing data.After data augmentation,the sample count is increased to 1,460.Experimental results show that the defect recognition accuracy improves with the increase in sample quantity.The accuracy stabilizes after the sample count reaches 1,300,showing a 64.8%improvement compared to a sample count of 100,validating the effectiveness of dataset construction.(2)Design of Wheel Defect Recognition AlgorithmThis paper proposes an improved Faster R-CNN algorithm.The algorithm utilizes specific data augmentation and pre-training methods during training.Compared to the original algorithm,the improved algorithm achieves a 6.0%improvement in accuracy and a 2.37 FPS improvement in speed on the WheelXray dataset.Moreover,the accuracy is higher than that of current popular defect recognition models.Specifically,the ResNet101 is used as the backbone network,and the following design improvements are made:firstly,the conv5_x block used for image classification in the backbone network is streamlined to reduce overfitting and improve network convergence speed;secondly,the FPN structure is incorporated into the backbone network to achieve multi-scale feature extraction;thirdly,the anchor box configuration in the RPN network is adjusted to adapt to the scale of casting defects in the dataset;finally,a linear interpolation algorithm is used to improve the quantization operation in the RoI Pooling layer and reduce accuracy loss.Experimental results show that the improved model achieves a 1.9%improvement in accuracy compared to the Faster R-CNN model.Furthermore,data augmentation and pre-training methods are further applied to optimize the improved model.Experimental results show a 4.1%accuracy improvement when using data augmentation methods such as horizontal flip,vertical flip,and random cropping,combined with pre-training on the ImageNet and COCO datasets.(3)Design and implementation of a wheel defect recognition systemThis paper presents the design and implementation of a wheel defect recognition system,which features automatic defect recognition,wheel data management,and model updating based on manual annotation feedback.Following the principles of software engineering and combining them with the wheel defect recognition process,functional and nonfunctional requirements are proposed.To address these requirements,the system adopts a layered architecture design,abstracting the functional modules into user module,data module,system module,and recognition module.Database and interface design work is carried out based on the functional modules,followed by detailed design and coding implementation of the main functional modules.This study utilizes an improved Faster R-CNN model for automatic defect recognition.For defects that cannot be accurately identified,the system provides a manual annotation method to obtain feedback.By leveraging the experiential knowledge from this feedback,the model’s recognition accuracy can be improved,leading to updates of the existing model.Finally,the system passes functional testing and analyzes its limitations and future optimization directions.