Deep Transfer Learning For Online Welding Quality Monitoring

With the continuous development of welding industry,the realization of automation,flexibility and intelligence in welding process has become an inevitable trend of industrial development.It is of great significance to improve the quality of welding process and related welding workpieces to study the on-line inspection of welding seam quality.In this paper,the online monitoring method of welding quality based on deep migration learning is studied.It is difficult to detect and identify the weld surface defects generated in the current welding process.By studying welding defect data enhancement,welding defect image classification and welding defect target detection method,Efficient and accurate online welding quality monitoring.The main findings are as follows:(1)Based on the generation of anti-network,the data enhancement of welding defect learning samples is realized.By introducing condition variables,Gaussian mixture model and self-attention mechanism optimization,the anti-network loss function is generated,A generator and discriminator model based on deep convolutional neural and self-attention mechanisms is proposed;The difference evaluation method between the generated sample distribution and the real sample distribution was studied.The experimental results show that the proposed data enhancement algorithm can provide an effective way for deep learning algorithm to train small welding defect samples.(2)Introducing DropBlock convolution kernel optimization and global average pooling,combined with lightweight convolutional neural network model MobileNet,proposed a welding defect classification algorithm based on optimized lightweight convolutional neural network,and using Fashion-MNIST small-scale image public data set and welding X-ray internal defect enhancement data set to test the algorithm,The experimental results show that compared with the commonly used convolutional neural network classification algorithm,the proposed welding defect classification algorithm takes into account the recognition accuracy and training efficiency,and has the advantages of low computational complexity and good deployment ease of use welding defect target identification provides means of implementation;(3)Study the internal and external feature mapping relationship and transfer learning strategy of welding defects,and use the network-based deep transfer learning strategy to realize the weight migration of internal and external feature models of welding defects.Improve the convergence speed and generalization ability of the welding external feature defect classification model,and effectively use the limited welding external feature data set for training;The object detection algorithm is combined with the welding external feature defect classification model and One-Stage object detection algorithm based on deep transfer learning.The welding defect object detection algorithm based on deep transfer learning is proposed and deployed to the embedded AI platform to perform real-time welding surface image detection,and welding quality real-time online monitoring.