Due　to　the　fact　that　the　resource　is　prone　to　be　wrong　during　tasks　execution　in　cloud,　which　leads　to　failed　tasks,　in　view　of　the　recent　research,　the　Primary-Backup　model　(PB　model)　is　mostly　used　to　deal　with　fault-tolerant　tasks,　but　the　selection　of　passive　scheme　and　active　scheme　is　assumed　in　advance,　and　the　advantages　between　the　two　schemes　are　not　fully　utilized.　Based　on　the　deep　reinforcement　learning,　this　paper　proposes　an　adaptive　PB　model　selection　algorithm,　Active-Passive　Scheme　DQN　(APSDQN).　The　process　of　faulty　task　tolerance　is　regarded　as　a　Markov　decision　process,　taking　the　passive　scheme　and　active　scheme　as　the　action　spaces,　the　shortest　completion　time　of　the　task　and　the　highest　resource　utilization　as　the　reward　feedback,　combine　with　the　real　environment　state　information,　select　the　most　suitable　fault-tolerant　scheme　for　faulty　tasks　to　save　resources　and　improve　the　robustness　of　cloud　system.　The　experimental　results　show　that　APSDQN　has　certain　advantages　in　the　total　task　finish　time　of　task　allocation,　and　significantly　improves　the　resource　utilization　and　the　task　success　rate　in　the　cloud.　©　2022,　Springer　Nature　Singapore　Pte　Ltd.