A damage-informed neural network framework for structural damage identification

In this work, an effective Damage-Informed Neural Network (DINN) is first developed to pinpoint the position and extent of structural damage. Instead of resolving the damage identification problem by conventional numerical methods, a Deep Neural Network (DNN) is employed to minimize the loss function which is designed by combining multiple damage location assurance criterion and flexibility matrices to guide the training process. In our computational framework, the parameters of the network, which include both weights and biases, are treated as new design variables instead of damage ratios. Therein, the training data consists only of a set of spatial coordinates of elements, whilst corresponding the damaged ratios of elements unknown to the network are factored into the output. To achieve the goal, the loss value is calculated relying on the predicted damage ratios with supporting Finite Element Analysis (FEA). Additionally, Bayesian Optimization (BO) algorithm is used to automatically tune hyperparameters of the network for enhancing reliability in damage identification. Several numerical examples for damage localization of truss and frame structures are investigated to evaluate the effectiveness and reliability of the suggested methodology. The obtained results point out that our model not only correctly locates the actual damage sites but also requires the least number of structural analyses and faster convergence rate compared with other algorithms.