Investigating uncertainties in shear resistance prediction of beams without stirrups.

Abstract

Optimizing a model’s performance should keep the functionality within the confines of safety and economy; deviation from these calls for a reliability investigation of such models. In a bit to optimize shear models for simplicity, safety functionality and economic performance have been an issue of a trade-off as the overestimation or underestimation of the model's intended purpose may occur. Overestimating the shear resistance of flexural members raises safety concerns since it might lead to unsafe design practices that ultimately cause the entire structure to collapse. In the same manner, underestimating the shear resistance may give rise to uneconomical designs. In this research, the predictions of various code-based & authorial shear resistance models in terms of their structural performance were assessed through the model uncertainties. According to Gino et al. (2017), identifying and quantifying the uncertainty related to a specific model is of high relevance to structural safety verification in the course of reliability assessment. Uncertainties related to models adequately capture the inconsistency of models’ performance across varied structural conditions. Hence, the extent of conservatism demonstrated by shear models of beams without stirrups is investigated towards structural reliability assessment and calibration. A database of 784 experimental beams without shear reinforcements compiled by Reineck et al. (2013), consisting of beams with varying geometrical properties was investigated in this study. Analyses conducted in this study include a mean values analysis (best-estimate prediction without any form of bias) and a deterministic design value analysis (inclusion of partial safety factor or reduction factor and characteristic material properties). Shear values derived from mean value analysis are used as the input parameter to determine the uncertainty of each model for the same structural condition. Supervised machine learning models based on the architecture of the Artificial Neural Networks, Support Vector Machine, Decision Tree Regressor and Random Forest were also used for shear resistance predictions. Model uncertainty was also derived for machine learning predictive shear models. A comparative analysis between the experimental shear resistance and all considered predictive model was done. Statistical characterization of each model factor in terms of the bias, standard deviation, coefficient of variation and skewness was carried out to evaluate the model’s performance in order to adopt a general probabilistic model for subsequent reliability evaluation. Sensitivity analysis of model uncertainty to parametric variation of input parameters was also carried out with the measured value of correlation. The provision of a carefully calibrated partial factor for model uncertainties that will take into account the uncertainty associated with shear methods is the most efficient management of the reliability performance for any resistance method. To this end, the calibration of a partial factor of safety according to EN 1992-4 was done for models with poor performance in terms of uncertainty performance indexes such as sensitivity to model parameters, a large degree of variability in shear prediction and significant bias in prediction.