Self-sensing capacity of strain-hardening fiber-reinforced cementitious composites: machine learning prediction and experimental validation

Abstract

This study focuses on the self-sensing capacity, which is indicated by gauge factor (GF) of strain-hardening fiber-reinforced cementitious composites (SH-FRCCs) for flexural specimen. At first, a machine learning model using a hybrid Random Forest–Particle Swarm Optimization (RF-PSO) technique was proposed to predict the GF for SH-FRCCs under direct tension. After that, an experimental program was conducted to validate the RF-PSO model in predicting GF of SH-FRCCs at the tensile zone of the flexural specimen. A dataset comprising 86 samples gathered from multiple previous studies was utilized to train and evaluate the proposed RF-PSO model. Eight potential input variables were considered: matrix strength (\(\sigma_{mu}\)), fiber type (FT), fiber geometry (\(L_{f} /d_{f}\)), fiber volume content (\(V_{f}\)), post-cracking strength (\(\sigma_{pc}\)), strain capacity (\(\varepsilon_{pc}\)), initial electrical resistivity (\(\rho_{i}\)), electrical resistivity at post cracking (\(\rho_{c}\)). The effectiveness of the hybrid RF-PSO model was assessed via four statistical metrics: R² (coefficient of determination), MSE (mean squared error), MAE (mean absolute error), and RMSE (root mean squared error). The analytical results showed that the proposed RF-PSO model showed excellent accuracy, with R² values of 0.935 in the training stage and 0.737 in the testing stage. The hybrid RF-PSO model demonstrated superior predictive performance compared to the pure RF model in predicting the GF of SH-FRCCs, improving the R² values by 1.05 and 1.14 times in the training and testing stages, respectively. Furthermore, one-dimensional partial dependence plot (PDP-1D) was used to investigate the effect of input variables on the GF of SH-FRCCs. It was found that the \(\sigma_{pc}\) and \(\rho_{c}\) extremely impacted to the GF predictions. The experimental results showed that the error between the experimental values and RF-PSO predictions is less than -13.63%, thus the proposed model in this study have high generalization capability in predicting the GF of SH-FRCCs.