Recently,　intelligent　fault　diagnosis　has　made　great　achievements,　which　has　aroused　growing　interests　in　the　field　of　bearing　fault　diagnosis　due　to　its　strong　feature　learning　ability.　Sufficient　bearing　fault　samples　are　taken　for　granted　in　existing　intelligent　fault　diagnosis　methods　generally.　In　practice,　however,　the　lack　of　fault　samples　has　been　a　knotty　problem.　Therefore,　in　this　paper,　a　novel　method　called　data　reconstruction　hierarchical　recurrent　meta-learning　(DRHRML)　is　proposed　for　bearing　fault　diagnosis　with　small　samples　under　different　working　conditions.　This　approach　contains　data　reconstruction　and　meta-learning　stages.　In　the　data　reconstruction　stage,　noise　is　reduced　and　the　useful　information　hidden　in　the　raw　data　is　extracted.　In　the　meta-learning　stage,　the　proposed　method　is　trained　by　a　recurrent　meta　learning　strategy　with　one-shot　learning　way.　This　approach　is　demonstrated　on　the　bearing　fault　database　with　92　working　conditions　from　Case　Western　Reserve　University　and　with　56　working　conditions　from　laboratory.　Results　show　that　the　proposed　method　is　effective　for　bearing　intelligent　fault　diagnosis　with　small　samples　under　different　working　conditions.