Fault　diagnosis　of　chiller　is　of　great　significance　for　saving　energy　and　maintaining　indoor　comfort.　Deep　neural　network　(DNN)　has　been　extensively　studied　for　fault　diagnosis　of　chiller　due　to　the　excellent　feature　extraction　ability　from　fault　data.　However,　the　neglect　of　redundancy　among　the　fault　features　of　chiller　degrades　performances　of　these　DNN　in　fault　diagnosis　of　chiller.　In　this　paper,　a　two-branch　sparsity　and　manifold　regularized　temporal　convolutional　neural　network　(TBSMRTCN)　is　used　for　fault　diagnosis　of　chiller.　Firstly,　a　two-branch　architecture　is　applied　to　TBSMRTCN　to　improve　the　result　of　fault　diagnosis.　Secondly,　a　sparsity　regularization　method　is　applied　to　the　TBSMRTCN　that　it　ensures　TBSMRTCN　to　filter　out　some　features　to　reduce　the　redundancy　of　extracted　fault　features.　Finally,　a　manifold　regularization　method　is　applied　to　the　TBSMRTCN,　which　enhances　the　generalization　performance　of　TBSMRTCN.　According　to　the　dataset　of　American　Society　of　Heating,　Refrigerating　and　Air-Conditioning　Engineers　(ASHRAE),　the　results　show　that　the　feature　extraction　method　proposed　in　this　paper　improve　the　fault　diagnosis　rate　(FDR)　of　chiller.　The　overall　fault　diagnosis　rate　of　TBSMRTCN　reaches　97%.　©　2022　IEEE.