Assessment of machine learning models for predicting the deflection of concrete beams

The deflection of a concrete beam was measured using various sensors, and the resulting data were used in this article to assess the machine learning models. Support Vector Regression (SVR), Random Forest (RF), Genetic Programming (GP), and the Back Propagation Neural Network (BPNN) were used to predict the deflection of concrete beams at L/3, L/2, and 2 L/3. Various types of sensors, including a force-resisting sensor (FSR), a flex sensor (FLS), a piezoelectric sensor (PZS), and a microelectromechanical systems (MEMS) accelerometer, were employed in structural health monitoring (SHM). A total of 400 experimental load–deflection data points, obtained from four sensors under gradually applied static loads ranging from 0 kN to 100 kN, were used to train and validate SVR, RF, GP, and BPNN models. Loads and sensor types were considered inputs, and the corresponding deflection, as computed by the sensor, was the output. The adopted ML models have performed well in predicting the deflection; however, the BPNN and GP models predict more accurately with fewer errors. To justify the potential of the adopted models, various statistical analyses have been carried out. Statistical evaluations show that the BPNN model achieved the lowest MAPE (10.68%) and highest correlation (R = 0.92), outperforming other models. Moreover, the Taylor diagram was used for further assessments, and plots were used to determine outliers and leverage.