Vibration-based structural damage detection using 1-D convolutional neural network and transfer learning

Abstract

This paper presents a novel vibration-based structural damage detection approach by using a one-dimensional convolutional neural network (1-D CNN) and transfer learning (TL). The CNN can effectively extract structural damage information from the vibration signals. However, the CNN training needs enough samples, while some damage samples (scenarios) obtained from real structures are limited, which will compromise the CNN ability to detect structural damage. As a solution, the numerical models have potential to provide sufficient CNN training samples; meanwhile, the state-of-the-art TL technique can significantly shorten the network training time and improve the accuracy. Therefore, this paper proposes a new method to detect the damage of a bridge model. The 1-D CNN is firstly trained with the samples of the single damage scenarios of the numerical bridge model. And then it is transferred to the complex scenarios of multi-damage (double or triple simultaneously), random size structures, and experimental model. The results demonstrate that: with the TL, the accuracy of damage detection is increased by about 47% at most, and the convergence speed is increased by at least 50%; in particular, the TL can inhibit over-fitting, and for the real bridge case, the accuracy also increased by 44.4%. It is demonstrated that: the TL can effectively improve the damage detection accuracy and convergence effect, and the application of this method to the random size structures also proves its generalization.