235 / 2021-04-13 13:27:59

Vibration-based structural damage detection using one-dimensional convolutional neural network and transfer learning

Transfer learning,Structural damage detection,1-D CNN,Vibration signals,Bridge model

This paper presents a novel and efficient approach to automatically classify the structural damage from acceleration signals via a one-dimensional convolutional neural network (CNN) and transfer learning (TL). As structural damage usually induces changes of the structural dynamic responses, the CNN can effectively extract structural damage information from vibration signals and classify them into the corresponding damage states. However, for actual structures it is difficult to obtain enough samples for the CNN, which will sacrifice the CNN ability to obtain enough damage information and affect the damage detection results. Numerical simulations have the potential to provide sufficient CNN training samples. The most important implementation strategy is to establish the relationship between the numerical model and experimental model, so as to transfer the damage detector from a numerical model to the corresponding experimental model. Therefore, this paper employs a bridge model and the state-of-the-art TL technology to establish the communication between the numerical model and the experimental model, that is, using the numerical model to train a 1-D CNN, and then using the experimental data to fine-tune the 1-D CNN model, so as to transfer the1-D CNN detector from the numerical model to the experimental model. The results confirm that: through TL, the accuracy of the 1-D CNN reaches 96%, which was 21% higher than the control experiment of non-TL model (75%), and TL makes the 1-D CNN more stable. Interestingly, by visualizing the change process of acceleration signal in the different CNN layer, it was found that the 1-D CNN amplifies the distinctness of different damage scenarios through the convolution layers, summarizes and maps them into corresponding categories in the fully connected layer. It is demonstrated that: the TL can effectively improve the damage detection accuracy of the experimental model and the damage information obtained from the numerical model can be used to make up for the incomplete damage information of the experimental model.