open access

https://hdl.handle.net/2115/100426

Neural Network with Variational Monte Carlo for Bayesian Inference and Response Estimation of Nonlinear Systems

英語

Variational inference

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

In the current environment, where health diagnosis, seismic response assessment, and control design of bridges, highrise buildings, and smart structures are becoming critical issues, methods for parameter identification and dynamic response estimation are needed for systems where highly nonlinear structural systems are subjected to highly nonstationary disturbances, such as earthquakes and wind. This study proposes an innovative computational framework for estimating structural parameters of seismically isolated bridges using neural network variational Monte Carlo (NNVMC). Conventional parameter identification methods, such as transition Markov chain Monte Carlo, often suffer from low computational efficiency and convergence problems in high-dimensional parameter spaces. To address these issues, we developed a new NN-VMC approach that combines neural network approximation with the probabilistic inference of variational Monte Carlo. We applied the proposed framework to the estimation problem of key structural parameters of a two-degree-of-freedom seismically isolated bridge model, using seismic response data as observational input. Using numerical example, we validated the parameter-estimation accuracy and computational efficiency of the NN-VMC method. The proposed NN-VMC method is expected to contribute to real-time structural integrity monitoring, damage identification, and seismic risk assessment of critical infrastructure.

会議発表論文

Version of Record