The　scarcity　of　historical　data　on　fault　states　is　a　common　phenomenon　in　the　health　management　of　chillers.　This　phenomenon　leads　to　poor　performance　of　fault　diagnosis　methods　based　on　deep　learning　for　chillers.　Generative　Adversarial　Network　(GAN),　a　common　data　augmentation　method,　is　an　effective　approach　to　conquer　this　challenge.　However,　dynamic　coupling　characteristics　exist　in　the　operational　data　of　chillers.　Dynamic　characteristics　lead　to　distribution　differences　in　operational　data　at　different　times　within　the　same　category.　The　coupling　characteristics　result　in　higher　inter-class　similarity　in　the　operational　data.　These　characteristics　pose　challenges　for　GAN-based　data　augmentation　tasks.　Therefore,　an　encoder-decoder-based　residual　feature　extraction　method　is　introduced　to　enhance　inter-class　separability　while　improving　intra-class　cohesion.　Furthermore,　a　residual　feature　generating　network　(DRFGN)　is　designed　to　augment　the　limited　historical　dataset.　In　the　DRFGN,　to　address　the　issue　of　insufficient　distribution　paradigm,　an　extended　class　characteristic　information　called　latent　attributes　is　introduced.　During　the　training　process　of　DRFGN,　latent　attributes　are　introduced　as　extended　discriminative　information,　allowing　the　generator　to　comprehensively　learn　the　class　characteristics　of　real　data.　Additionally,　smooth　L1　loss　is　employed　to　further　enhance　the　similarity　between　the　synthesized　and　real　data.　Finally,　a　dataset　collected　from　the　actual　equipment,　referred　to　as　HY-31C,　is　used　for　the　validation　of　the　proposed　method.　The　results　indicate　satisfactory　distribution　similarity　between　the　synthesized　and　real　data,　benefiting　the　health　management　and　maintenance　decision　of　the　equipment.　©　2024　Technical　Committee　on　Control　Theory,　Chinese　Association　of　Automation.