Gas Indices and Deep Learning for Early Warning of Coal Spontaneous Combustion: A False-Alarm-Controlled Evaluation with Simulation and Field Evidence

Coal spontaneous combustion is both a mine-safety hazard and a source of uncontrolled greenhouse gas emissions, since seam fires burn underground for years and force premature panel sealing and resource sterilisation. Early detection is therefore a sustainability problem as well as a safety one, yet the gas ratios used to detect it, above all Graham’s index, are rarely assessed against periods in which no heating develops. Reported detection rates and lead times are consequently uninterpretable, because any alarm rule can look prescient by alarming continuously. We propose a false-alarm-controlled protocol in which every method is reduced to a continuous alarm score, its threshold fixed on validation episodes, and the test and field data reserved for evaluation. On 18 field episodes from three coal mines, a carbon-monoxide threshold of 11 ppm detected all three confirmed fires (95% CI 0.29 to 1.00) with no false alarms among fifteen event-free records and a median lead time of 276 h, while Attention–LSTM models also detected all three but raised false alarms in six and two of the fifteen event-free records, respectively, while Graham’s index detected none of the three. This ordering among the single-channel methods is unchanged across four persistence durations and three false-alarm targets. Achieved false-alarm rates nonetheless departed from the target when thresholds were transferred, indicating that alarm calibration does not carry across monitoring settings.