Perpustakaan Digital ITB

Boiler tubes are critical heat-transfer components in industrial boiler systems, and their failures cause costly, prolonged shutdowns and safety hazards. Conventional failure diagnosis relies on manual visual inspection, which is slow, subjective, and exposes personnel to hazardous conditions. This study develops a machine learning (ML) model to automatically detect boiler tube surface defects from visual inspection data, intended as a foundation for correlating these defects with their underlying damage mechanisms. A surface defect-to-damage mechanism mapping matrix was constructed, correlating six surface defect classes with corresponding damage mechanisms: 3 for Crazing, 15 for Patches, 12 for Pitted Surface, 8 for Scratches, 1 for Rolled-in Scale, and 2 for Inclusion. A ResNet-18 model, a deep learning architecture for image recognition, was trained using transfer learning on combined NEU-DET (a public steel surface defect image dataset), field, and synthetic data. The baseline model achieved 95.24% validation and 66.67% testing accuracy, while an OpenCV (image-processing tool)-optimized version reached 100% validation and 80.00% testing accuracy, though the Patches class remains a limitation.