Hydrodynamic　feature　extraction　of　pressure　pulsation　signals　(PPSs)　and　intelligent　identification　of　flow　regimes　in　vaneless　space　(VAS)　of　a　pump　turbine　(PT)　are　crucial　to　the　safe　and　stable　operations　of　the　pumped　storage　power　station.　In　this　work,　the　scheme　based　on　an　improved　empirical　wavelet　transform　(IEWT),　energy　feature　vector　(EFV)　and　Bayesian　optimized　convolutional　neural　network　(BOCNN)　is　innovatively　proposed.　Firstly,　the　IEWT　is　proposed　by　introducing　the　least　square　method　and　mathematical　morphology　to　improve　the　decomposition　shortcomings　of　existing　methods.　The　phenomenon　of　mode　aliasing　is　eliminated　and　the　influence　of　background　noise　is　greatly　reduced,　as　verified　by　both　simulated　and　measured　PPSs.　Secondly,　based　on　the　IEWT,　several　significant　mode　components　are　obtained,　and　the　energy　feature　indexes　of　them　are　calculated　to　extract　the　hydrodynamic　feature　information　and　construct　the　EFVs　that　can　accurately　reflect　the　features　of　different　flow　regimes　in　the　VAS.　Then,　the　BO　algorithm　is　adopted　to　optimize　the　important　hyperparameters　of　CNN,　and　the　intelligent　identification　model　of　BOCNN　is　constructed　and　trained　to　identify　four　typical　types　of　flow　regimes　in　the　VAS.　Finally,　the　average　identification　accuracy　of　the　proposed　IEWT-EFV-BOCNN　scheme　can　reach　99.15%,　which　is　much　higher　than　traditional　schemes,　illustrating　that　the　proposed　scheme　has　significant　engineering　application　value.