Dynamic material parameter inversion of high arch dam under discharge excitation based on the modal parameters and Bayesian optimised deep learning

Vibration-based dam safety monitoring methods have increasingly become a research hotspot for assessing dam safety. The accurate material parameters of a high arch dam and its foundation are critical for monitoring vi-bration safety. Accordingly, a dynamic material parameter inversion framework for an arch dam is developed based on modal parameters and deep learning. Firstly, an determined-order stochastic subspace identification method with adaptive variational modal decomposition is proposed based on the prototype vibration signal obtained under the discharge excitation, which can effectively identify the modal parameters. Secondly, the sensitivity of the dynamic elastic modulus (DEM) in different regions to the modal parameters of the arch dam was analysed using the orthogonal test method and variance analysis method, and it was utilised to ascertain the DEM to be inverted. Finally, a Bayesian optimised multi-output long short-term memory neural network is used to establish a nonlinear mapping relationship between the DEM and the modal parameters as an alternative to finite element calculations, and the identified modal parameters are employed as network inputs to inverse the actual DEM of each zoning. An engineering example shows that the proposed DEM inversion method for high arch dams is effective and accurate, providing a good basis for the vibration safety analysis of arch dams. This study overcomes the limitations of difficult to effectively extract modal parameters of arch dams under discharge excitation, and advances the application of deep learning technology in hydraulic engineering by combining modal parameters.

Automated summary

Vibration-based dam safety monitoring methods have become crucial for assessing dam safety. This study developed a dynamic material parameter inversion framework for arch dams using modal parameters and deep learning. A determined-order stochastic subspace identification method was proposed to effectively identify the modal parameters. The sensitivity of the dynamic elastic modulus (DEM) in different regions to the modal parameters was analyzed, and a Bayesian optimized multi-output long short-term memory neural network was used to establish a nonlinear mapping relationship between the DEM and modal parameters. The proposed method was proven effective and accurate for high arch dams, advancing the application of deep learning technology in hydraulic engineering.