Precursory information identification and integrated multi-parameter warning model to fracture in thermally treated sandstone

Studying the early warning information of thermally treated rocks is of paramount significance for understanding the permeability of fractured geothermal reservoirs. We conducted uniaxial compressive tests on thermally treated sandstone, synchronized with acoustic emission (AE) monitoring. The precursory information of rock failure is analyzed by using the precursor indices based on AE b-value, AE energy concentration, AE parameters variance, and AE hit rate. The results show that when the sample approaches instability failure, both the AE b-value and AE energy concentration significantly decrease, whereas the variances of AE parameters and AE hit rate dramatically increase. The collected 18 groups of data were divided into 13 groups for the test set and five groups for the validation set, and the early warning model based on the random forest (RF), support vector machine (SVM), neural networks (NNs), logistic regression (LR), and LightGBM algorithms are primarily trained. Compared with SVM, NN, LR, and LightGBM algorithms, the RF-based early warning model can not only reduce the error effect due to environmental noise, but also effectively improve the warning accuracy. As the number of input features increases, the performance of the early warning model is remarkably enhanced. The RF-based early warning model integrates four input features and achieves the highest early warning ability with 97% accuracy and 93% precision. In addition, the indicators selected in the feature combinations also have a great impact on the performance of the early warning model. Compared with other input feature combinations, the performance of the early warning model that includes AE b-value is significantly improved.