Reliable strain monitoring in asphalt pavement using piezoelectric sensors in bending mode and machine-learning-based calibration

Abstract

Ensuring the durability of asphalt pavements is essential for maintaining transportation infrastructure. Structural health monitoring (SHM) supports this by enabling early detection of deterioration. Among SHM techniques, piezoelectric sensors offer real-time strain measurement capabilities, but accurate calibration is crucial for reliable use in applications such as fatigue life monitoring. This study calibrates lead zirconate titanate (PZT) sensors for strain measurement in asphalt pavements by employing machine learning (ML) models to convert voltage signals into accurate strain data under bending loads. In the experimental phase, PZT sensors were tested during four-point bending (4 PB) experiments on Teflon and asphalt beams over a strain range of 15–450 μm/m and loading frequencies from 1–35 Hz. Teflon, with homogeneous mechanical properties, was used to establish baseline strain–voltage relationships in the elastic regime. The resulting dataset was used to analyse sensor behaviour and develop ML-based calibration models. A variety of ML algorithms were evaluated for predictive accuracy, robustness, and consistency. Results showed an approximately linear strain–voltage relationship across both materials, with nonlinearity at higher frequencies attributed to the capacitive nature of the sensors. Ensemble ML models, particularly Extra Trees Regression (R2 = 0.978, RMSE = 24 μm/m) and CatBoost Regression (R2 = 0.970, RMSE = 28 μm/m), achieved the highest accuracy, demonstrating controlled-condition strain calibration. These findings demonstrate that integrating PZT sensors with ML-based calibration can enhance strain measurement reliability in asphalt pavements, providing a foundation for advanced SHM systems aimed at improving pavement performance and service life.