Finite Element Model Updating Based on Neural Network Ensemble

Over the last few decades, structural health monitoring (SHM) has been gaining more and more attention, especially in civil engineering. Due to the assumptions and uncertainties in finite element (FE) modelling, there are inevitably various errors between the dynamic characteristics predicted by the FE model and the measured data. So it is necessary to calibrate the initial structural model, which is a typical inverse problem, and generally ill-posed. As a powerful artificial intelligence technology, artificial neural network (ANN) has been widely used in model updating due to its excellent pattern recognition ability. Compared with the traditional ANN approach, the Bayesian Neural Network (BNN) method is more robust to noise. However, with the increase in the number of dimensions and hidden neurons, the amount of samples required for training neural networks and the corresponding time consumption shows catastrophic growth, especially for training a single neural network to update large-scale FE models of civil engineering structures. To make progress, the ensemble of multiple neural networks fed with divided training sample sets is a feasible strategy. It is expected to improve the generalization performance compared to a single network for handling large-scale FE models, which is seldom emphasized in the current literature related to the FE model updating. This paper proposes a FE model updating method that utilizes the strategy of neural network ensemble by utilizing the modal flexibility matrix as the training input. The entire set of training samples is further divided into a series of smaller sample sets and used to train multiple BNNs, the final identification result is obtained by summing the outputs weighted by the evidence of each individual model. A truss model is employed in this paper to validate the feasibility and effectiveness of the method.