Damage identification analyses are fundamental to guarantee the safety of civil structures. They are often formalised as inverse problems whose solution ignores any source of uncertainty that could be accounted for by using appropriate statistical models. Unfortunately, these models often exhibit an intractable likelihood function. We propose quantifying uncertainty through a fully Bayesian approach based on Approximate Bayesian Computation (ABC), a class of methods that overcome the evaluation of the likelihood and only require the ability to simulate from the model. Furthermore, we suggest a strategy to reduce ABC computational burden using Neural Networks. Finally, we test the method at work on a damaged beam to discuss its strengths and weaknesses.
Approximate Bayesian Computation for Probabilistic Damage Identification
Cecilia Viscardi
;
2023-01-01
Abstract
Damage identification analyses are fundamental to guarantee the safety of civil structures. They are often formalised as inverse problems whose solution ignores any source of uncertainty that could be accounted for by using appropriate statistical models. Unfortunately, these models often exhibit an intractable likelihood function. We propose quantifying uncertainty through a fully Bayesian approach based on Approximate Bayesian Computation (ABC), a class of methods that overcome the evaluation of the likelihood and only require the ability to simulate from the model. Furthermore, we suggest a strategy to reduce ABC computational burden using Neural Networks. Finally, we test the method at work on a damaged beam to discuss its strengths and weaknesses.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
