Reliability enhancement of Non-destructive testing methods for IN-SITU concrete compressive strength using Convolutional Neural Networks with destructive testing data.

Abstract

Reliable evaluation of in-situ compressive strength of concrete is vital for ensuring safety and longevity of ageing infrastructure particularly bridges that recently are subjected to increasing traffic loads and deterioration due to climatic conditions. While traditional core extraction methods have reported high accuracy, these are intrusive, time-consuming and costly, thus limiting their practical application in structural assessment of concrete. In response to the need for less intrusive, quicker and cost-effective alternatives, this study investigated the use of the two most common Non-Destructive Testing techniques (NDT)—Schmidt Rebound Hammer (REB) and Ultrasonic Pulse Velocity (UPV)—coupled with Convolutional Neural Networks (CNN) to improve predictive accuracy. The study developed and trained three CNN model configurations using MATLAB® R2025a with data collected from five existing concrete bridge structures. The architecture consisted of a Conv1D with an input layer for two-channel sequential data, a convolutional layer with five filters, BatchNorm, ReLU activations, two fully connected layers with dropout regularization and a final dense output neuron, optimized for regression tasks. Each test method (REB, UPV and Core) generated 30 measurements which were subsequently cleaned and organized into triplicate datasets. REB-only, UPV-only and Combined REB–UPV were evaluated against core test results as ground truth. Model results revealed that the REB-only model performed reasonably well (R² = 0.75, RMSE = 1.78 MPa, MAPE = 10.8%), UPV-only model exhibited lower predictive capacity (R² = 0.31, RMSE = 2.09 MPa, MAPE = 29.6%) whereas the Combined REB–UPV model outperformed both, achieving an R² of 0.90, RMSE of 1.39 MPa and MAPE of 7.8% during training. However, signs of overfitting were observed during testing, primarily attributed to the limited dataset size. It was demonstrated that combining NDT data with CNN-based deep learning networks can significantly enhance compressive strength prediction over single-method approaches. This study offers a novel application of CNNs—typically used in image recognition—for numeric prediction-based concrete assessment using NDT and Destructive Testing (DT) data. It presents a promising, scalable yet non-invasive practical tool for structural health monitoring.