Abdul Azziz Bin Abd Talib, Lim Chun Yee and Liew Chin Kian, Singapore Institute of Technology, Singapore

Anchor bolts are critical connectors that maintain structural alignment and load-bearing capacity in civil infrastructures. Defects in these bolts can compromise safety and operational reliability, yet traditional visual inspections are often insufficient for detecting internal damage. This study presents a feature-based machine learning framework integrated with ultrasonic testing to enhance defect detection in anchor bolts. A fabricated bolt inspection system with a 10 MHz ultrasonic transducer was used to acquire 477 ultrasonic signals from pristine, straight thinning, and tapered thinning bolts, with additional validation signals collected from field-installed bolts. Three dimensionless features were engineered to capture signal clarity and defect-related scattering. Six machine learning classifiers were evaluated using stratified crossvalidation, with Gradient Boosting achieving the highest accuracy of 93%, outperforming other classifiers, including ensemble methods. The model demonstrated strong robustness in distinguishing between nondefective (Green), monitoring required (Yellow), and defective (Red) bolts even under practical variability. Field deployment in train tunnels further validated the model’s reliability with no false Red classifications. The results confirm the viability of integrating ultrasonic testing with machine learning for automated anchor bolt inspection, enabling accurate, data-driven infrastructure maintenance and predictive safety strategies.

Anchor bolts, Ultrasonic testing, Machine learning, Structural health monitoring, Non-destructive testing, Infrastructure safety