Bridge Vibration Signal Denoising Based on Convolutional Denoising Autoencoder Neural Network

To mitigate the influence of noise on vibration signals employed in bridge monitoring, a denoising approach based on a Convolutional Denoising Autoencoder Neural Network (CDANN) is proposed. In this approach, the reconstruction mechanism of DAE is systematically integrated with the feature extraction capability of CNN. Skip connections and dropout regularization are incorporated to enhance the stability and generalization performance of the model; wherein noisy signals are used as inputs and the corresponding denoised signals are generated as outputs. The effectiveness and robustness of the proposed approach are validated through a numerical simply supported beam model and an experimental bridge. The results indicate that, for the numerical simply supported beam model subjected to Gaussian white noise and pink noise interference, CDANN improved the average output SNR by 11.6 dB and 6.4 dB compared with CNN and WTD, respectively, while reducing the average RMSE by factors of 1.26 and 1.18. Moreover, the RMSE was reduced by more than 64% relative to the original noisy signals. For the experimental bridge, the denoising ratio of CDANN exceeds 70%, and the R is consistently maintained above 0.98. These findings confirm the effectiveness of CDANN in denoising vibration signals and demonstrate that the proposed method can serve as a reliable data-driven solution for structural health monitoring systems.