Adaptability of identification algorithms while MWD application in various geological conditions

Machine Learning (ML) has been extensively utilized in the Measurement While Drilling (MWD) technology in the coal industry. However, there existed uncertainty about whether mainstream algorithms can effectively and accurately identify rock strength across diverse geological conditions. Therefore, this study compared the effects of the three mainstream algorithms, K-means clustering, BP neural network, and SVM classification, to clarify the geological conditions suitable for each algorithm, and finally proposed the optimal selection and application methods of various algorithms under different geological conditions. The study revealed that the accuracy for K-means clustering, BP neural networks, and SVM multi-label classification algorithms were 85%~88%, 92%~96%, and 89%~92% respectively. Additionally, BP neural networks, SVM multi-label classification, and SVM single-label classification are suitable under conditions where strength differences are less than 6.3 MPa, between 6.3 MPa and 22.7 MPa, and greater than 22.7 MPa, and their generalization rates are 98.12%, 98.48%, and 94.71%, respectively. K-means clustering is appropriate for the preliminary identification of strata information in scenarios with few rock types and simple existing conditions. At the same time, BP neural networks and SVM multi-label classification are well-suited for strata characterized by “multiple rock types and low strength differences.” In contrast, SVM single-label classification is more suitable for strata featuring “few rock types and high strength differences”, which would hold significant practical value in enhancing the applicability of ML for MWD technology in coal mines.