A Complex Welding Defect Detection Method Based on Active Learning in Pipeline Transportation System

In the fields of industrial information processing and reliability security, accurate and intelligent detection of pipeline welding defects is a necessary yet challenging task. Due to the complex pipeline welding process, the existing automated defect inspection systems are inadequate to detect weld defects with high uncertainty. In this paper, an effective detection method based on active learning is proposed, which can improve the accuracy for complex weld defects in unseen scenarios. Firstly, a set of diversified regression paradigms based on mutual-assistance learning for welding defect detectors with different initialization hyperparameters are introduced to generate the initial model, reducing the prediction bias of traditional active learning baselines. Then, multiple reasonable sample mining strategies are presented based on unlabeled datasets, which can accomplish feature alignment with unseen defects and optimize the fuzzy classification hyperplane. Next, network fine-tuning mechanisms are introduced to continuously update model performance and save training costs. In the experiment, we conducted extensive effectiveness analysis and compared it with state-of-the-art methods. A large number of real-world experimental results and practical application cases have demonstrated the efficiency and accuracy of this method in pipeline system maintenance.