Coal　energy　plays　a　pillar　role　in　the　development　of　national　economy.　The　safe　mining　of　coal　energy　has　always　been　an　important　research　topic　of　domestic　and　foreign　scholars.　In　view　of　the　problems　existing　in　the　current　mine　fire　detection　methods,　such　as　the　number　of　measurement　points,　the　difficulty　of　maintenance,　the　complexity　of　installation　and　the　high　rate　of　false　alarm　and　missing　alarm,　an　intelligent　mine　fire　detection　method　based　on　convolution　neural　network　is　proposed.　In　view　of　the　problem　that　the　learning　rate　parameter　selection　is　not　suitable　and　easy　to　interfere　with　the　convergence　of　the　model,　the　selection　of　subjective　factors　is　strong　and　it　is　not　easy　to　find　the　best　learning　rate,　this　paper　proposes　a　method　of　the　whole　process　adaptive　learning　rate.　This　method　takes　the　mine　temperature,　humidity,　smoke　concentration,　CO　concentration　and　O-2　concentration　as　input,　through　the　self-learning　of　the　whole　process　adaptive　learning　rate　convolution　neural　network,　and　outputs　the　prediction　results　respectively,　namely,　the　probability　values　of　open　fire,　smoldering　fire　and　no　fire.　By　using　Anaconda　environment　to　build　model　simulation　results　show　that　the　recognition　error　of　open　fire,　smoldering　fire　and　no　fire　probability　is　less　than　3%,　which　can　greatly　reduce　the　rate　of　missing　and　false　alarm.