Uncertainty-aware fuzzy knowledge embedding method for generalized structural performance prediction

Abstract

Structural performance prediction for structures and their components is a critical issue for ensuring the safety of civil engineering structures. Thus, enhancing the reliability of the prediction models with better generalization capability and quantifying the uncertainties of their predictions is significant. However, existing mechanism-driven and data-driven prediction models for reliable engineering applications incorporate complicated modifications on models and are sensitive to the precision of relevant prior knowledge. Focusing on these issues, a novel and concise data-driven approach, named “R2CU” (stands for transforming regression to classification with uncertainty-aware), is proposed to introduce the relative fuzzy prior knowledge into the data-driven prediction models. To enhance generalization capacity, the conventional regression task is transformed into a classification task based on the fuzzy prior knowledge and the experimental data. Then the aleatoric and epistemic uncertainty of the prediction is estimated to provide the confidence interval, which reflects the prediction's trustworthiness. A validation case study based on shear capacity prediction of reinforced concrete (RC) deep beams is carried out. The result proved that the model's generalization capability for extrapolating has been effectively enhanced with the proposed approach (the prediction precision was raised 80%). Meanwhile, the uncertainties within the model's prediction are rationally estimated, which made the proposed approach a practical alternative for structural performance prediction.