Narrow gap GTAW defect detection and classification based on transfer learning of generative adversarial networks

Gas tungsten arc welding (GTAW) is the primary process employed for critical applications such as pressure vessels and power pipelines where the weld quality and integrity are essential. Traditionally, its quality and integrity heavily rely on engineers' experience to assess because the phenomena involved are extremely complex to interpret. Deep learning can significantly enhance the efficiency. However, acquiring a sufficient number of correct defect features during actual production presents a challenge. To address this issue, this study first analyzed the narrow GTAW process to identify suitable phenomenon that fundamentally correlates to the relevant defects. To mathematically explain these complex relationships, we explored two classification networks, ResNet and Vision Transformer (VIT), as potentially most effective network structures to filter image features from a public aluminum alloy GTAW database for transfer learning. Additionally, a Generative Adversarial Network (GAN) was utilized to generate defect feature images. Transfer learning from the public database and GAN-generated image features substantially improved the model's prediction accuracy. Validation on the test set revealed that the pre-trained GAN-generated images support the model achieved the best F1 score and accuracy, at 92.78 % and 92.8 %, respectively.