Intelligent Signal Denoising Method and Experimental Verification Based on a Hybrid SACNN-BiLSTM Neural Network

In a structural health monitoring (SHM) system, the data collected by sensors are inevitably affected by noise disturbance of varying levels in a real engineering environment, which can degrade signal quality and impact subsequent data analysis. To address this issue, this paper proposes a denoising method for SHM signals based on a hybrid self-attention convolutional neural network/bidirectional long short-term memory (SACNN-BiLSTM) neural network. The method with novel hybrid architecture, self-attention mechanism, intelligently denoising and robustness can effectively improve the quality of noisy signals, which is validated by applying monitored vibration data from an actual large-span bridge in engineering. Five separate models were established for training data samples with different noise levels, and the results demonstrate that when the noise level is 20%, the signal-to-noise ratio (SNR), root mean square error (RMSE) and [Formula: see text] of denoised signals are, respectively, 15, 0.01 and 0.98; for the noisy signal with 80% noise level, the denoised signal has an approximately 65.5% increase in SNR and a 43.75% decrease in RMSE, along with a 50.67% increase in [Formula: see text]; the higher the signal noise level is, the more pronounced the denoising effect. In engineering applications, the average error of modal identification results based on the denoised signals is 0.78%, the modal assurance criterion is greater than 0.95, and the power spectral density (PSD) plots are clearer with significantly reduced spurious patterns. By handling Gaussian noise, blue noise and pink noise, the experimental results confirm the excellent noise reduction capability of the proposed method under various noise conditions.