Machine learning-assisted ultrasonic assessment of tendon duct grouting in post-tensioned concrete structures

Voids left by incomplete grouting make post-tensioning strands vulnerable to corrosion, yet the grouting condition is still judged largely by visual interpretation of ultrasonic tomograms, which is operator-dependent. This paper presents a machine learning-assisted framework for grouting assessment based on ultrasonic pulse-echo data from the MIRA A1040 tomograph. Seventeen physics-based features are extracted, nine from raw in-phase/quadrature waveforms and eight from images reconstructed with the Synthetic Aperture Focusing Technique, each linked to a specific mechanism of shear-wave interaction with grouted and ungrouted ducts. The dataset comprises 331 B-scans: 15 from a laboratory calibration specimen and 316 from the Gisund Bridge, an in-service post-tensioned bridge affected by alkali–silica reaction, labelled from endoscopy-confirmed grouting conditions. Over 100 repeated stratified train–test splits, six classifiers reached balanced accuracies of 0.66–0.70 against a 52.3% majority-class baseline; XGBoost, selected for the most stable generalisation, achieved a balanced accuracy of 0.70 and a Matthews correlation coefficient of 0.40. Grouped leave-one-section-out testing showed weaker cross-section generalisation, and the model agreed with five of six core-drilled verification openings. Physics-based features therefore carry transparent discriminative information, and the framework is best used as decision support for inspectors.