Graded identification of air leakage in coal-mine gas drainage boreholes using a physics-augmented transformer–LSTM framework

Air leakage in gas drainage boreholes in coal mines reduces methane concentration and extraction efficiency, yet reliable grading remains difficult because weak leakage signatures can be obscured by noise and changes in operating conditions. This study proposes a physics-augmented Global–Local Transformer–LSTM (GL-TransLSTM) framework for classifying normal, slight, moderate, and severe leakage. The four grades were defined in controlled underground tests by setting the valve opening of a drainage branch exposed to ambient mine air to 0%, 10%, 50%, and 100%, respectively. Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) was used to decompose four synchronized signals—drainage negative pressure, methane concentration, gas flow rate, and acoustic intensity—into multiscale components. Residuals based on effective methane flow and pressure–flow–concentration coupling were incorporated as auxiliary features with physical meaning. Parallel Transformer and LSTM branches captured global dependencies and local transitions through adaptive gated fusion, while a coarse-to-fine classifier and state continuity regularization used only during training improved discrimination between adjacent grades. The framework was evaluated on 3,900 windows generated from 492 uninterrupted monitoring segments, with the data partitioned by monitoring segment before window generation to prevent temporal leakage. Across five independent runs, the mean accuracy and macro-F1 were 90.26 ± 0.41% and 90.38 ± 0.37%, respectively. In Run 3, which was used for detailed analysis of individual classes, recalls for slight, moderate, and severe leakage were 89.00%, 90.00%, and 92.22%, respectively. The measured processing time from input to classification was approximately 6.7 ms per window under the tested configuration, including feature generation and classification but excluding external data acquisition, communication, storage, and visualization delays. These results support the feasibility of GL-TransLSTM for graded leakage monitoring in near real time under the tested configuration.