A　translation-invariant　based　on　adaptive　threshold　denoising　method　for　mechanical　impact　signal　is　proposed.　Compared　with　traditional　wavelet　denoising　methods,　it　suppresses　pseudo-Gibbs　phenomena　in　the　neighborhood　of　signal　discontinuities.　To　remedy　the　drawbacks　of　conventional　threshold　functions,　a　improved　threshold　function　is　introduced.　It　possesses　more　advantages　than　others.　Moreover,　by　utilizing　characteristics　of　signal,　a　adaptive　threshold　selection　procedure　for　impact　signal　is　proposed.　As　it　is　data-driven　and　level-dependent,　it　is　more　rational　than　other　threshold　estimation　methods.　The　method　is　compared　to　alternative　existing　methods,　and　its　superiority　is　revealed　by　simulation　and　real　data　examples.