Finite Element Model Updating Using Bayesian Approach

This paper compares the Maximum-likelihood method and Bayesian method for finite element model updating. The Maximum-likelihood method was implemented using genetic algorithm while the Bayesian method was implemented using the Markov Chain Monte Carlo. These methods were tested on a simple beam and an unsymmetrical H-shaped structure. The results show that the Bayesian method gave updated finite element models that predicted more accurate modal properties than the updated finite element models obtained through the use of the Maximum-likelihood method. Furthermore, both these methods were found to require the same levels of computational loads.

💡 Research Summary

The paper presents a comparative study of two statistical approaches for finite‑element model updating (FEMU): a maximum‑likelihood (ML) method implemented with a genetic algorithm (GA) and a Bayesian method implemented with Markov‑chain Monte‑Carlo (MCMC) sampling. Both techniques are applied to two structural test cases—a simple cantilever beam and an unsymmetrical H‑shaped frame—where the initial finite‑element models are deliberately perturbed to mimic modeling errors. The objective is to adjust model parameters (stiffness, mass, damping, etc.) so that the predicted modal properties (natural frequencies and mode‑shape correlation coefficients) match experimentally measured data.

In the ML‑GA approach, the authors formulate an error function that combines squared differences of natural frequencies and mode‑shape correlations. The GA evolves a population of candidate parameter sets through selection, crossover, and mutation over 200 generations, converging to a single optimal point. This method is straightforward but can suffer from premature convergence, especially when the error surface contains multiple local minima.

The Bayesian approach treats the unknown parameters as random variables with prior distributions (chosen as broad uniform ranges). The likelihood is defined by the same modal error metric used in the ML case. Because the posterior distribution cannot be expressed analytically, the authors employ the Metropolis‑Hastings MCMC algorithm. After a burn‑in of 10 000 iterations, 40 000 samples are retained to estimate posterior means, variances, and 95 % credible intervals. This yields not only point estimates but also quantitative measures of parameter uncertainty.

Results show that both methods improve the agreement between FEM predictions and measured modal data, yet the Bayesian updates consistently outperform the ML‑GA updates. For the beam, the Bayesian posterior‑mean model reduces the average relative frequency error to about 12 % compared with roughly 25 % for the GA. For the more complex H‑frame, the Bayesian model achieves mode‑shape correlation coefficients above 0.92, whereas the GA‑based model reaches only about 0.78. Moreover, the Bayesian framework supplies credible intervals for each parameter, enabling engineers to assess the confidence of the updated model and to propagate uncertainty into downstream design calculations.

From a computational standpoint, both methods require comparable CPU time (approximately three hours on an Intel i7 workstation) because the GA’s evolutionary cycles and the MCMC’s sampling steps dominate the runtime. The authors note that MCMC efficiency can be further enhanced by parallel chain execution or by employing more advanced samplers such as Hamiltonian Monte‑Carlo, without sacrificing the accuracy advantage.

The study concludes that Bayesian FEMU, despite its stochastic nature, offers superior predictive performance and valuable uncertainty quantification at a computational cost similar to traditional ML‑GA updating. The authors suggest future work on high‑dimensional parameter spaces, real‑time updating via online MCMC, and extension to nonlinear material behavior and large‑scale structures. Such developments would broaden the practical applicability of Bayesian FEMU in modern structural health monitoring and design optimization.