As　it　is　always　submerged　by　the　supply　frequency　component　in　current　signal,　the　rotor　fault　feature　component　is　difficultly　detected.　A　novel　feature　extraction　method　of　rotor　faults　based　on　continuous　wavelet　transform　(CWT)　and　singular　value　decomposition　(SVD)　is　proposed.　The　current　signal　which　consists　of　the　supply　frequency　component　and　the　rotor　fault　feature　components　is　transformed　into　the　time-frequency　space　by　CWT,　and　the　time-frequency　information　is　decomposed　to　a　series　of　orthogonal　sub-spaces　by　SVD,　which　include　different　feature　components.　With　the　reconstruction　of　the　sub-spaces　selected,　the　rotor　fault　feature　components　are　extracted　effectively.　Experimental　results　and　industrial　measurement　analysis　show　that　the　approach　provides　a　more　effective　means　to　diagnose　rotor　faults.