Perpustakaan Digital ITB

In-line inspection (ILI) data from multiple periods is used to track pipeline corrosion over time, but corrosion features often shift in position between inspections, which makes it hard to compare the same defect and lowers prediction reliability. This research develops an automated feature matching system to align corrosion features across multi-temporal ILI data and analyzes its effect on machine learning-based corrosion depth prediction. The matching used an iterative closest point transformation and an overlap-based correspondence optimization on three inspection periods of an offshore pipeline, evaluated through positional error reduction and a simulation-based sensitivity analysis. An XGBoost model was replicated from a reference study, then applied to the case study data with and without calibration from the matched dataset. The results show that feature matching reduces the median positional error by up to 99.87% and reaches a matching accuracy above 80% for both period pairs, while the replication matched the reference within 0.2%. The calibration lowered the prediction error from 7.909 %WT to 3.379 %WT, below the naive benchmark of 3.505 %WT. These findings show that feature matching is an important preprocessing step for improving machine learning-based corrosion prediction .