Field-Validated deep learning model for Piezoelectric-Based In-Situ concrete strength sensing

Abstract

In the realm of concrete strength monitoring, the fusion of piezoelectric sensor data with machine learning algorithms holds significant promise. However, previous studies have been limited by datasets that lack diversity, often confined to homogeneous mix designs and sensor types, which restrict the applicability of developed models in variable real-world conditions. This study addresses this gap by constructing a comprehensive database that encompasses an unprecedented variety of mix designs and sensor deployments, capturing a wide spectrum of concrete behaviors. Our research introduces a novel 1D Convolutional Neural Network (1DCNN) architecture paired with a baseline mechanism, specifically tailored for the analysis of electro-mechanical impedance (EMI) signals. Rigorous validation through train-test splitting has demonstrated the high accuracy of our model within laboratory settings, achieving an R² value of 0.96 and a mean prediction error of 2.68 MPa for concrete strength. We’ve also conducted field validation tests on actual highway concrete pavement projects, marking a pioneering application of lab-trained models for strength prediction under field conditions. The successful application of our model in these field tests confirms its robustness and reliability, marking a pivotal step toward practical deployment. Our study not only bridges the gap between laboratory research and field application but also sets a new benchmark for dataset informativeness and model versatility in the domain of concrete strength prediction.