Abstract: Aiming at the huge computational load of the structural damage identification method based on Bayesian model in the face of dealing with an intensive numerical model, this study proposes a high fidelity surrogate model construction method combining adaptive sparse Polynomial Chaos Expansions(PCE) and Kriging model. Based on this surrogate model, a new structural damage identification method is proposed. Firstly, the Least-Angle Regression algorithm was used to adaptively select the truncated PCE optimal expansion item as the trend item of Kriging model to improve the training and prediction efficiency of the surrogate model. The random item is modeled by Gauss’s random process. Secondly, the calculated value of the finite element model is replaced by the predicted value of the surrogate model to avoid repeated calls to finite element models in Bayesian inference. Then, the modified MCMC method was used to obtain the posterior probability distribution of the uncertain parameters to determine the location, degree and damage probability of the structure. Finally, two numerical examples show that the method proposed has a good robustness and can be applied to structural damage identification in complex environments with high identification accuracy. It provides an effective method for real-time health monitoring of large-scale civil engineering structures based on Bayesian finite element model modification.