System identification of concrete gravity dams using artificial neural networks based on a hybrid finite element–boundary element approach

Abstract

System identification is an emerging field of structural engineering which plays a key role in structures of great importance such as concrete gravity dams. In this study, an artificial neural network (ANN) procedure is proposed for the identification of concrete gravity dams, in conjunction with a hybrid finite element–boundary element (FE–BE) analysis for the prediction of dynamic characteristics of an existing concrete gravity dam with an empty reservoir. First, a dam–reservoir interaction analysis is carried out by the hybrid FE–BE approach in the frequency domain. A two-dimensional (2D) FE model (FEM) is used for linear-elastic analysis of the gravity dam on a rigid foundation, while the unbounded reservoir with inviscid, compressible, and frictionless fluid is discretized by BEs. Various analyses are performed for different height to base width ratios of dams in terms of different wave reflection coefficient of the reservoir bottom. The use of ANNs is motivated by the approximate concepts inherent in system identification approaches, and the time-consuming repeated analyses required for dam–reservoir interacting systems. The conjugate gradient algorithm (CGA) and the Levenberg–Marquardt algorithm (LMA) are implemented for training the ANNs, using available data generated from the results of coupled dam–reservoir system analyses. The trained ANNs are then employed to compute the dynamic amplification of dam crest displacement and natural frequencies of existing concrete gravity dams through forced vibration tests. The results obtained by solving the present inverse problem are compared with existing FEM solutions to demonstrate the accuracy and efficiency of the proposed method.