Sensitivity and robustness analysis of adaptive neuro-fuzzy inference system (ANFIS) for shear strength prediction of stud connectors in concrete

The shear strength of stud connectors is essential for designing steel-concrete structures, which is assessed only through a push-out test or available design codes. An alternative technique that eliminates the need to conduct the push-out test is soft computing (SC). The performance of any machine learning (ML) based prediction model depends on the sensitive parameters used in the model development. This paper performs a sensitivity analysis on the shear strength prediction of stud connectors embedded in concrete. A system identification (SI) was conducted using an adaptive neuro-fuzzy inference system (ANFIS) to find the most sensitive combinations of input variables. Six different models were developed based on the SI results. Three machine learning algorithms, including ANFIS, extreme learning machine (ELM), and artificial neural network (ANN), were used to estimate the shear strength of stud connectors in each developed model. The results show that the number of studs (n) is the most sensitive parameter in predicting shear strength. Irrespective of concrete compressive strength (f_c), the combination of the stud diameter (ϕ), number of studs, and stud spacing (s) can predict the shear strength with the accuracy of ±8.67 kN. The robustness of the three AI algorithms was evaluated using the Monte Carlo Simulation method. The individual conditional expectation (ICE) was also presented to visualize the correlation between the target shear strength and the six predictors. The results of this study show that sensitivity analysis is an essential tool for any data-driven ML model for accurate prediction.