Image Analysis And Welding Quality Research Of Robot Welding Pool Based On Deep Learning

The weld pool is the result of the comprehensive action of the welding process,and contains a wealth of welding quality information.However,due to the complex welding environment,the molten pool is mixed with a lot of noise information,which affects the expression of image features;in addition,the learning model of molten pool image features has the problems of low accuracy and poor real-time performance.Therefore,in order to effectively solve the above problems,a morphological dehazing algorithm is proposed to improve the image quality;in order to obtain a model with better performance,the Alexnet-LSTM welding quality prediction optimization model is proposed.The main research contents are as follows:Based on the arc welding robot workstation,a visual monitoring platform is built,including the installation and debugging of hardware equipment and image acquisition software.Through a large number of exploratory tests,a reasonable test plan was determined,and the task of acquiring molten pool images of different quality levels was completed.In order to filter out smoke and splash interference in molten pool images,a morphological dehazing algorithm is proposed to preprocess the images.The algorithm uses dark channel prior theory and guided filter to filter out smoke and dust in the image,and uses15×15 morphological operator to eliminate splash.The grayscale features and texture features of the dehazed image are calculated.The grayscale feature results show that the algorithm can filter out the smoke in the image and improve the richness of image information.The texture feature results show that the image texture changes processed by the algorithm are more regular.In order to determine the mapping relationship between molten pool and weld bead,and divide reasonable welding quality grades,a molten-pool weld bead evaluation model was proposed,and finally six different welding quality standards from good to bad were obtained.In order to ensure the rationality of the melt pool image dataset,some images are taken for expansion operations to form a dataset with a capacity of 12000.The Alexnet-LSTM welding quality prediction optimization model is proposed.First,the Alexnet convolutional neural network model is used to train the data set,and the optimal learning rate is determined to be 0.0001,and the optimal convolution kernel size is 9 ×9;Then the long short-term memory network is used to train separately,and the optimal learning rate is determined to be 0.0001 and the optimal number of hidden layers is 200.The accuracy of the above two models trained separately cannot achieve the purpose of high accuracy and low loss value at the same time,so the Alexnet-LSTM model is used for training.After adding the LSTM model to the fifth convolutional layer in the Alexnet network,the model realizes the screening of convolutional features.The experimental results show that the average accuracy of the proposed optimization model stably converges is91.46%,and the average loss value is 0.32.The accuracy of the optimized model is 17.46%higher than that of Alexnet,and the loss value is reduced by 0.87 for a longer short-term memory network.In conclusion,the proposed morphological dehazing algorithm effectively solves the problem of smoke and splash interference in molten pool images.The design of the molten pool-weld evaluation model lays the foundation for the formulation of optimization strategies for welding process specifications.The research on the optimization model of welding quality prediction provides a new idea for solving the problem of real-time monitoring of welding,and ensures the stability and good rate of welding bead formation.