Augmented Hybrid Learning for Visual Defect Inspection in Real-World Hydrogen Storage Manufacturing Scenarios

Ensuring product quality while reducing costs is critical in manufacturing scenarios. However, real-world operational factories, particularly in the hydrogen storage industries, pose several challenges, including strict quality control standards and limited but extremely biased data available for model development. To address these challenges, we propose an augmented hybrid learning method for visual defect inspection that leverages the strengths of deep learning and unsupervised learning. The proposed method is developed using only one-class OK samples and then validated on a real-world operational manufacturing line. The experiment results demonstrate that our method achieves a recall rate of nearly 90% with an overkill rate of only 0.6%. This method outperforms several benchmark methods that often struggle to balance high recall and low overkill rates. Experiments with industrial setups show that our method provides a promising solution for visual defect inspection in real-world manufacturing scenarios.