In　severe　few-shot　scenarios,　fault　samples　are　extremely　limited,　challenging　the　health　management　of　the　chiller　significantly.　Generative　adversarial　network　(GAN),　a　common　data　augmentation　method,　suffers　from　a　feature　shift　issue　in　severe　few-shot　scenarios.　This　is　because　the　extremely　limited　fault　samples　cannot　provide　sufficient　distribution　paradigms　for　GANs.　Therefore,　a　severe　few-shot　fault　diagnosis　method　for　chillers　based　on　an　extended　conditional　mapping　feature　generation　network　(ECMFGN)　is　proposed　to　tackle　this　issue.　In　ECMFGN,　the　ECM　module　is　built　to　provide　an　extended　distribution　paradigm　for　the　generator.　This　extended　distribution　paradigm　can　adequately　reflect　class-specific　characteristics.　Then,　through　adversarial　supervised　learning,　the　feature　similarity　between　the　samples　synthesized　by　the　generator　and　the　real　samples　is　improved.　Experiments　demonstrate　that　the　feature　shift　issue　in　severe　few-shot　scenarios　is　solved　effectively.　In　addition,　the　characteristics　of　different　fault　modes　for　chillers　are　close.　Thus,　insufficient　intraclass　cohesion　and　interclass　separation　can　lead　to　interclass　confusion　of　generated　samples,　which　directly　affects　diagnostic　accuracy.　To　conquer　this　challenge,　a　smooth　$L_{1}$　triplet　loss　(TL)　is　designed　to　promote　intraclass　cohesion　while　improving　the　separability　between　classes.　The　experimental　results　indicate　that　the　performance　of　the　fault　diagnosis　model　is　significantly　improved　in　the　chiller　dataset　with　similar　characteristics　among　the　fault　modes,　benefiting　the　health　management　and　maintenance　decisions　of　the　equipment.