Deep learning-based adaptive denoising method for prediction of crack opening displacement of rock from noisy strain data

Precise prediction of crack opening displacement (COD) is necessary to control it within a certain threshold to ensure safety of critical buried rock structures (e.g. geological disposal facilities), and rock COD nowadays tends to be estimated using distributed fiber optic sensing (DFOS) data. The frequently used methods for computing COD of engineering structures from DFOS data are based on analytical models derived from strain transferring. However, these analytical models face significant challenges in computing COD when DFOS data exhibits high levels of noise. To address this limitation, this research aims to develop a deep attention threshold processing network (DATPN) to improve feature extraction ability from highly noisy DFOS data and accurately predict the COD of rock. A channel attention module is firstly designed and used as the main module by the DATPN to automatically learn thresholds adapting to different noises, so that the professional expertise is not required when determining noise thresholds. The learned noise thresholds are then provided to a built-in improved thresholding functions of the DATPN to eliminate data noise. Therefore, the DATPN can easily learn more useful target features to accurately predict the COD of rock owing to the elimination of data noise. The results indicate that the proposed DATPN achieves a maximal goodness of fit (R²) of 0.9924, a minimal mean squared error of 5.22, and a percent error for COD prediction ranging within ±10 % when evaluated using the original measured dataset. The R², correlation coefficient, root-mean-square error, and standard deviation for the DATPN (0.8407, 0.9193, 10.31, and 24.37) are superior to the other four machine learning models on the dataset with a signal-to-noise ratio of 0. This demonstrates that the DATPN is satisfactory in predicting the COD of rock from highly noisy DFOS data.