Design And Implementation Of Edge End Rim Weld Detection System In Industrial Internet

The application of industrial internet,computer vision,and the Internet of Things technologies has injected new power into industrial production lines.As a load-bearing component of automobiles,the wheel undergoes processes such as steel plate forging,welding,and inspection in actual production,among which the detection of wheel rim weld plays a significant role.The accuracy of detection affects the efficiency of subsequent procedures such as airtightness and quality inspection.Computer vision can achieve high-precision and highefficiency detection of rim weld,avoiding the misjudgment caused by subjective factors such as human factors and environmental conditions.Meanwhile,adopting edge computing can reduce data transmission delays and improve data security.Therefore,using computer vision and edge computing to achieve wheel rim weld detection tasks has significant practical significance and application value.Therefore,this thesis focuses on algorithm research and edge system application development for the accuracy and inference speed of the rim weld detection model.The main work of this thesis is as follows:(1)Starting from the learning of model parameters and aiming at the problem of low detection accuracy of small object in the three target scales of small,medium,and large after using data augmentation methods,this thesis proposes a loss function based on dynamic weighting of target scales.The loss function implements exponential weighting based on the target scale,and maps the different weighted penalty values smoothly for different scales of small,medium,and large object.By increasing the weight of locating loss for small object during model training,the feature learning of the model parameters is biased towards small object.Experimental results on a public dataset show that the proposed loss function based on dynamic weighting of target scales can improve the detection accuracy of the model without increasing the model parameters or the computational complexity of inference.(2)This thesis proposes an optimization scheme for optimizing the redundant feature receptive field based on the idea of optimizing the objective detection model.By using a context fusion module to optimize the receptive field of the main output and adopting a multi-input single-output and effective pruning detection head,the method achieves fast inference for wheel rim weld detection.Through the use of the two proposed improvement methods,and the lowprecision quantization of model parameters for inference testing on embedded edge computing platforms,the experiments show that the model parameters are only 0.28 M,the average precision(AP)is improved by 2.2% compared to the baseline,and the inference speed is increased by 13%.The inference time for a single image on the embedded edge platform is only74 ms,achieving a balance between speed and accuracy,and meeting the real-time requirements of production lines.(3)To gain a comprehensive understanding of the application requirements for wheel rim weld detection on wheel rim production line,this thesis integrates the proposed model optimization methods and designs and implements a development of a wheel rim weld detection system.The system combines functions such as dataset acquisition,model training,model testing,model conversion,and online operation.It supports multiple platforms and possesses portability,stability,and robustness,which greatly enhances the efficiency of wheel rim production on the production line.