GTAW Penetration Prediction Model Based on Convolution Neural Network Algorithm

Abstract

：The penetration state can be reflected by the information of the molten pool, but it is difficult to establish a function between the molten pool and the penetration state. To solve this problem, a penetration prediction model based on convolution neural network (CNN) is proposed. Based on the introduction of CNN principle, a molten pool sensing system based on passive vision is designed to collect 2D images of the molten pool. The acquired images are preprocessed to generate the training set and test set for CNN training and testing. Then the prediction mode is built, and the network parameters such as learning rate, batch-size and iterations are optimized. It is found that the model can achieve the best comprehensive performance in prediction accuracy and training time when the size of the first layer convolutional kernel is 99  and the last layer contained 64 convolutional kernels. After training the model with the training set, the trained model is used to predict the penetration state on the test set, and the prediction accuracy is higher than 96%. By visualizing the feature mapping of the prediction model, it is found that the model predicts the penetration state through the features of edge, the position of reflective point and the molten pool tail, thus explaining how the model the judges the penetration state.

Defects detection of GMAW process based on convolutional neural network algorithm

“…However, it lacks robustness across various welding scenarios. These models [18][19][20] achieve high accuracy in predicting defects within a specific scenario based on CNN and its improved algorithms. However, there are two significant issues persist: (1) The mechanisms of welding defect prediction remain unclear.…”