Automatic Recognition of Multiple Weld Types Based on Structured Light Vision Sensor Using Deep Transfer Learning

Abstract

Real-time and high-precision extraction of groove types and key features is an important factor to achieve high performance of weld quality in the process of automatic welding. Based on the complexity and diversity of weld groove types, a method for identifying weld groove types (TL-Alexnet-ELM) is presented, which combines deep transfer learning with extreme learning machine. Firstly, to avoid over-fitting of the model, the weld data set is expanded using data enhancement technology. Then, to improve the generalization ability of the model, transfer learning is used to fine-tune the structure of Alexnet (TL-Alexnet) to improve the feature extraction accuracy of the source weld image. Finally, the extracted image eigenvectors are input into the extreme learning machine classifier to get the classification results of the groove types. To validate the effectiveness of the algorithm, model self-comparison study, comparison study of different deep learning network models, and comparison study of different classifiers are carried out. The experimental results show that the recognition accuracy of this method is 99.9%.