Structural damage detection based on convolutional neural networks and population of bridges

The CNN-based detection methods have been widely used in the field of structural health monitoring (SHM), however, they can only be used for individual structures under certain conditions; for the structures in-service, damage detection will be affected by a variety of external factors (unknown/uncertain load and geometric dimension, etc.). Therefore, in order to improve the applicability of the CNNs, their compatibility and robustness need to be thoroughly investigated. In this paper, a large number of random models were produced to establish a population of bridge structures, the damage features of the population were extracted by the CNN; subsequently, the CNN was applied to damage detection of the new randomly-created models. The results show that the best detection results (99.4% accuracy) can be obtained by using the acceleration signals as the CNN input. This demonstrates that the proposed method will expand the detection ability of the CNN beyond an individual structure.