Improving Classification of Imbalanced Classes in Industrial Data: Enhancing Defect Detection for Type IV High-Pressure Hydrogen Vessels

This study focuses on enhancing defect detection in Type IV high-pressure hydrogen vessels (HPV) using advanced machine learning techniques. Type IV vessels, crucial for hydrogen storage and transportation in industrial and automotive sectors, feature composite materials with polymer liners to ensure mechanical strength and hydrogen tightness at pressures up to 700 bar. Detecting defects in these vessels is critical for maintaining structural integrity and safety. The primary challenge addressed is the significant class imbalance among defect types, where critical defects are infrequent compared to minor defects or defect-free instances. Standard classification models often fail to effectively detect these critical defects due to their bias towards majority classes. To overcome this, we evaluated and compared methods including Weighted Logistic Regression, Weighted Decision Trees, and the Discrete Minimax Classifier (DMC), which adapt their strategies to improve minority class detection. Our findings demonstrate that these adapted algorithms enhance sensitivity to minority classes, particularly in reducing false negatives for critical defect detection. This study emphasizes the importance of tailored machine learning approaches in industrial defect detection, paving the way for safer and more reliable hydrogen storage technologies through optimized predictive maintenance and quality control processes. In particular, we include in our study the recently studied minimax approaches, which have strong theoretical foundations and good empirical performances, as show our experiments.

Keywords: Classification, Imbalanced classes, High-pressure type IV hydrogen vessels, DMC, Weighted algorithms, Defect detection.