Signal generation for bolt loosening detection with unbalanced datasets based on the CBAM-VAE

Bolt looseness has adverse influences on the stability and safety of engineering structures. Neural network algorithms can effectively monitor health conditions using impedance signals. However, impedance data of engineering structures in damaged conditions is challenging to obtain. The data would also exhibit an imbalanced distribution, yielding deterioration of the accuracy of health monitoring. In this study, we propose a data augmentation method based on a Variational Autoencoder model with a convolutional block attention module. This method addressed the issue of imbalanced data by generating new data. A Transformer model was adopted for training and fault classification. Without employing data augmentation methods, the max accuracy is 85.71%. However, experimental results demonstrate the remarkable effectiveness of this approach in enhancing and classifying imbalanced datasets, with an average accuracy of 89.35% and the highest accuracy of 94.81% after enhancement. The proposed method can be applied to health conditions identification of buildings, bridges, and trusses.