Enhanced prediction of early-age concrete compressive strength utilizing Non-bonded piezo sensors and deep learning methods

This study uses a non-bonded piezoelectric sensor (NBPS) on an aluminum-foil-affixed concrete cube specimen to assess the early mechanical properties of concrete using deep learning (DL). NBPS and aluminum foil accurately measure electromechanical impedance (EMI) signals during hydration, revealing concrete microstructural change. A computational model in COMSOL Multiphysics generated a large data set for DL analysis from high-resolution EMI signatures. Experimental observations and numerical simulation results verified the model's accuracy, proving its reliability and robustness. Using EMI signatures, one-dimensional (1D CNNs) were trained to predict concrete compressive strength. Lasso regression chose features from the original dataset to reduce computational cost and increase model efficiency. This strategy eliminated less significant features that improved model performance. The predictive accuracy rose from 0.940 (original dataset) to 0.959 (Lasso-modified dataset) as evidenced by the R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value. MAE dropped from 5.75 to 4.85, and RMSE dropped from 7.73 to 6.33. These improvements demonstrate feature selection's ability to improve model predictions. Smoother training curve convergence showed training stability, and correlation coefficient and scatter plot analysis showed better value alignment. Lasso regression reduced dataset overfitting and redundancy, boosting CNN heterogeneous data generalization. Due to the indirect connection in NBPS signal strength prediction accuracy is somewhat lower than embedded (EPS) and surface-bonded (SBPS) piezo sensor sets. While noninvasive, easy to use, and real-time monitoring without changing specimen structure, the NBPS has considerable implementation advantages.