Statistical evaluation of a machine learning model as shear strength prediction on reinforced concrete beam

The shear strength (Vs) computation of reinforced concrete (RC) beams has been a major topic in structural engineering. Several methodologies have been introduced for the Vs prediction; however, the modeling accuracy is relatively low owing to the complex character of the resistance mechanism involving the dowel effect of longitudinal reinforcement, concrete in the compression zone, the contribution of the stirrups if existed, and the aggregate interlock. It is difficult, if not impossible, to shear design RC beams with and without stirrups utilizing laboratory trials. The span-todepth proportion, web width, and reinforcement proportion are only a few of the various factors that must be considered concurrently. Additionally, empirical techniques for shear design are developed within the confines of their testing regimes owing to the complicated shear failure process. As a result, these methodologies have limited generalizability and application. To overcome this problem, this work applies machine learning strategies for shear design. The current thesis is adopting the developing the Random Forest (RF) model as a robust machine learning (ML) predictive model for Vs prediction for reinforced concrete beams. The proposed ML model is developed based on collected experimental data 349, including the beam geometric and concrete properties parameters. Nine input combinations are constructed based on the associated input parameters for the proposed predictive model. The validation was conducted against the support vector machine (SVM) model, considered a well-established ML model introduced in the literature. In addition, several empirical formulations (EFs) are calculated for comparison. Research findings evidenced the potential of the proposed RF model for modeling the Vs reinforced concrete beams. Based on quantitative metric for the testing phase modeling, the RF model achieved the best results of the seventh input combination with root mean square error (RMSE = 89.68 KN), mean absolute error (MAE = 35.59 KN), mean absolute percentage error (MAPE = 0.16). The modeling accuracy performance comparison with the established ML models and the EFs confirmed the capacity of the proposed model. Results indicated that all the parameters utilized beam geometric and concrete properties are significant for the development of the predictive model. However, the model structure emphasizes the incorporation of seven predictors by excluding (beam flange thickness and coefficient). In general, the research provided a reliable a robust soft computing model for Vs of RC beams computation that contributes to the basic knowledge of structural engineering design and sustainability.

Text 113370.pdf Download (1MB)

Actions (login required)

View Item