Identification of coal structures by semi-supervised learning based on limited labeled logging data

Abstract

Coal structure is a critical parameter in coalbed methane (CBM) development due to its significant impacts on methane enrichment, fluid flow and hydraulic fracturing. Traditional statistical analysis and data-driven machine learning methods for coal structure identification are highly dependent on the labeled logging data and have potential limitations when labeled logging data is limited. To address this issue, this paper proposed a semi-supervised learning method based on Laplacian support vector machine (LapSVM) to identify coal structure by using few labeled logging data. By mining the structure information from abundant unlabeled data, LapSVM can improve the model performance and alleviate the over-reliance on labeled data. To evaluate and verify the effectiveness and reliability of the proposed LapSVM method in coal structure identification, datasets collected from 32 CBM wells in the southern Qinshui Basin, China, are utilized in this study. The particle swarm optimization (PSO) is adopted for parameter optimization of LapSVM models. For the LapSVM model, the addition of unlabeled data is conducive to enhance model accuracy, and unavoidably increases the computational cost at the same time. The comparison of training, testing and blind-well test results between the LapSVM and standard support vector machine (SVM) models indicates that the LapSVM outperforms traditional SVM and possesses higher accuracy and generalization in coal structure identification. It has been demonstrated that the LapSVM can be a reliable tool for coal structure identification when limited labeled logging data is available.