To　extract　fault　feature　of　mechanical　signal　corrupted　by　noise,　a　new　method　to　construct　redundant　second　generation　wavelet　(RSGW)　is　presented　to　extract　fault　feature　in　time　domain.　The　initial　prediction　operator　and　initial　update　operator　are　interpolated　with　zero,　and　then　the　prediction　operator　and　update　operator　at　different　decomposition　scale　are　gained.　Splitting　operation　is　unnecessary　for　RSGW,　and　the　approximation　signal　is　predicted　and　updated　directly,　the　length　of　approximation　signal　and　detail　signal　for　all　scales　remains　identical,　and　the　thresholds　at　different　scales　are　selected　according　to　the　noise　characteristics.　Experimental　and　engineering　results　confirm　the　advantages　of　RSGW　over　the　other　wavelet　methods　for　signal　de-noising,　and　the　fault　features　of　bearing　inner　raceway　damage　and　vibration　excited　by　steam　in　high-pressure　turbine　of　turbo-generator　in　time　domain　are　desirably　extracted　by　RSGW.