Objective CO₂ displacing coalbed methane (CBM) provides substantial environmental and energy benefits. However, differences in coal structures represent a key geological factor influencing the effectiveness of this technique. Constrained by the complex geological conditions for CBM occurrence and the impacts of multiple factors on the log responses of coal structures, traditional log-based methods for coal structure identification face challenges such as much human intervention, complex feature engineering, and poor adaptability. This study aims to achieve effective identification of complex coal structures within a single coal seam.

Methods An asymmetric convolution kernel-based convolutional neural network (CNN) model for coal structure identification was proposed in this study. When constructing feature matrices for the model, a feature weighting mechanism based on mutual information values was incorporated to underline sensitive log curves. Meanwhile, a Gaussian weighting strategy was integrated along the depth direction to highlight the log information from central points. Using two asymmetric convolution kernels of sizes 3×1 and 3×2, the CNN model enabled differential feature extraction from the log data in the vertical and lateral directions. Vertically, the weights of the contributions of various log values to feature extraction were adaptively adjusted based on the differences in the coverage times of the convolution kernels across different lines of depth. Laterally, the synergistic variation characteristics among multiple log curves were fully preserved.

Results The design of asymmetric convolution kernels allows for the effective utilization of the spatial structural characteristics of log data. Accordingly, complex nonlinear mapping relationships between log responses and coal structures are established. Experiments were conducted on log data from primary coal seams in the CBM exploration area at the periphery of the Panji coal mine, Huainan mining area, Anhui Province. The experimental results reveal that the CNN model yielded an accuracy of up to 82.64% on the test set, significantly outperforming the multilayer perceptron (MLP, 78.13%), support vector machine (SVM, 73.67%), and K-means clustering (64.42%) models.

Conclusion The integration of priori geological knowledge with deep learning can effectively enhance the accuracy of log-based coal structure identification. This finding provides a novel technical approach to the high-accuracy identification of coal structures in structurally complex areas, holding great practical value for efficient CBM production.