Welding Defects Detection Improve By Data Augmentation: Fusing Negative Image’s Edge Maps To Positive Image

With the continuous advancement of the "China Intelligent Manufacturing 2025",China’s manufacturing industry is developing in the direction of intelligence and information technology.Welding is a kind of technology to join metal by high temperature heating,which is widely used in China’s automobile manufacturing,aerospace equipment production.Therefore,the quality of welding will directly affect the reliability of the final product,and then affect people’s production and living safety.In this paper,the following problems were found in the welding inspection of actual automobile parts production: first,the number of defect samples produced in the production process was insufficient;second,the number of different types of defects was greatly different;finally,the defect features of different forms existed in the same defect.The recalibration of welding equipment in the production line will be adjusted according to the feedback results of quality inspection,so the detection algorithm should be able to feedback the specific defect problems in the defective parts in order to provide support for the recalibration of production equipment.In view of the above problems,The main content of this paper consists of the following parts:1)In this paper,a new data augmentation strategy is proposed in combination with the conditional generation antagonism network to alleviate the unbalanced problem of welding defect data in three dimensions.In this paper,in the positive image into negative image after Sobel operator to extract the edge of the defect area,using GAN to fusion for filling out the edges of the graph to generate new defects after samples,the experiment shows that the generation strategy to generate samples of specific form has better controllability,can more effective use of improved the characteristics of the new training sample model generalization performance.2)This paper designs an adversation network generated by edge fusion filling conditions,which can generate new defect samples after filling the images with defect edge features in the positive samples.On the network structure,the generator uses the spatial adaptive normalized structure to integrate the generating conditions and guarantees the stability of the training high-resolution image and the better detail performance of the generated image through the gradually increasing structure.3)In the detection model,this paper uses Focal Loss Loss function to balance the impact of a large number of simple samples in training on the learning of difficult samples.In terms of network structure,this paper redesigned the model for the multiscale and multi-objective overlapping structure in view of the relatively fixed welding detection scene and the similar size of welding pads without overlapping,and proved the effectiveness of the improved structure and expansion method through practice.