Smart prediction of rock crack opening displacement from noisy data recorded by distributed fiber optic sensing

Abstract

The commonly used method for estimating crack opening displacement (COD) is based on analytical models derived from strain transferring. However, when large background noise exists in distributed fiber optic sensing (DFOS) data, estimating COD through an analytical model is very difficult even if the DFOS data have been denoised. To address this challenge, this study proposes a machine learning (ML)-based methodology to complete rock's COD estimation from establishment of a dataset with one-to-one correspondence between strain sequence and COD to the optimization of ML models. The Bayesian optimization is used via the Hyperopt Python library to determine the appropriate hyper-parameters of four ML models. To ensure that the best hyper-parameters will not be missing, the configuration space in Hyperopt is specified by probability distribution. The four models are trained using DFOS data with minimal noise while being examined on datasets with different noise levels to test their anti-noise robustness. The proposed models are compared each other in terms of goodness of fit and mean squared error. The results show that the Bayesian optimization-based random forest is promising to estimate the COD of rock using noisy DFOS data.