Deep Learning-based data-driven technique for early-age concrete strength monitoring using the non-bonded piezoelectric sensor system

This study introduces a novel framework combining electro-mechanical Impedance (EMI) technology with machine learning (ML) and deep learning (DL) for real-time assessment of early-age concrete compressive strength. A non-bonded piezoelectric sensor (NBPS) was mounted on an OPC-43 grade cement concrete cube using aluminum foil to monitor EMI signatures during hydration. Conductance and susceptance data were collected via an LCR meter across 3 to 150 days of curing intervals. Experimental data were validated using a physics-based, simulation-driven approach in COMSOL Multiphysics, which replaces labor-intensive experiments by generating an extensive dataset for effective training and validation of ML and DL models. Mechanical property variations, including stiffness (k) and damping (c), were analyzed using an equivalent system approach. Prior ML studies achieved R² values of 0.94 and 0.95 for embedded piezo sensors (EPS) and surface-bonded piezo sensors (SBPS). In the present study, ensemble ML methods like bagged and boosted trees attained an R² value of 0.9 for NBPS, demonstrating its predictive efficacy. Despite a slightly lower R², NBPS exhibited stability, ensuring its reliability. While previous studies have applied 2D CNN and 1D CNN models for damage prediction with image-based inputs, this study introduces a novel 1D CNN-based DL framework for early-age concrete strength assessment, leveraging EMI data in dataset form rather than image-based input. Developing hybrid architectures, such as 1D CNN-LSTM and 1D CNN-BiLSTM, also represents a new approach in the field. Feature selection via Lasso regression enhanced model performance, with the 1D CNN achieving an R² of 0.934 and the 1D CNN-Bi-LSTM reaching 0.988, setting a new benchmark in EMI-based strength prediction, an area yet to be extensively explored. Once trained, the DL models can instantly predict compressive strength at any curing stage and can be easily adapted to different materials, curing conditions, or sensor setups, providing a scalable solution for structural health monitoring. While the NBPS configuration showed slightly lower predictive accuracy than EPS and SBPS, its non-invasive setup and real-time monitoring capability provide practical advantages.