Investigating the Predictive Capabilities of ANN, RSM, and ANFIS in Assessing the Collapse Potential of RC Structures

Abstract

The contemporary study delves into the evaluation of the proficiency of artificial neural networks, response surface methodology (RSM), and adaptive neuro-fuzzy inference system (ANFIS) in estimating the collapse probability (CP) in various reinforced concrete (RC) structures. The research leverages the FEMA P695 methodology as a foundation for calculating structural collapse probabilities. The FEMA P695 process necessitates the execution of pushover and incremental dynamic analysis (IDA), involving 22 pairs of far-field records, rendering it a time-intensive procedure. To address this, a comprehensive analysis comprising approximately five thousand IDA simulations, as per FEMA P695 methodology, was performed on a diverse dataset of over a hundred RC frames, varying in height from 1 to 12 storeys. Five key parameters were chosen as the independent input variables for the models, while the CP served as the system's output variable. The ANN model, using a double hidden layer with Levenberg-Marquardt training, achieved a regression coefficient (R) of 0.977 and a mean absolute error (MAE) of 1.570. The RSM model, implemented with a cubic system, had an R of 0.975 and a higher MAE of 2.052. ANFIS, utilizing a Takagi-Sugeno fuzzy system, outperformed the other methods, with an R of 0.981 and the lowest MAE of 1.401. Comparative analysis revealed that ANFIS offers the highest efficiency and accuracy, demonstrating its superior capability in predicting CP with minimal error rates. This study provides a comprehensive evaluation of these techniques, highlighting ANFIS's robustness and precision, while also offering insights into the interrelationships among significant parameters.