Structural damage detection using convolutional neural networks combining strain energy and dynamic response

Based on the classification ability of a convolutional neural network (CNN), this paper proposes a structural damage detection method in which a CNN is used to classify the location and level of damage in a structure. The dynamic responses are combined with the modal parameters of the structure as the inputs to the CNN to detect the damage. As structure damage can cause changes in multiple damage indicators, an individual indicator may not be enough to detect all damage scenarios. The combination of multiple damage indicators will provide more comprehensive information for damage situation. It is expected that this combination will overcome disadvantages of the damage index based on a single modal parameter. The finite element method was used to provide the training samples for the network. Damage in an element was introduced by reducing its Young's modulus. Two cases were considered for the input of the CNN: the first used the modal strain energy only, and the second used the combination of modal strain energy and dynamic response (acceleration). The comparison results show that the inclusion of dynamic responses in the damage index significantly improves the correctness rate of structural damage detection and enhance the convergence of the network.