To　acquire　an　impact　component　aroused　by　mechanical　fault,　a　novel　feature　extraction　method　based　on　nonlinear　manifold　learning　of　continuous　wavelet　coefficients　was　put　forward.　Firstly,　the　wavelet　entropy　method　was　adopted　to　optimize　the　Morlet　wavelet　shape　factor　in　order　to　match　with　the　impact　components　to　obtain　the　optimal　continuous　wavelet　coefficients.　Secondly,　the　nonlinear　manifold　learning　algorithm　named　local　tangent　space　alignment　was　used　to　reduce　the　optimal　wavelet　coefficients　matrix,　and　according　to　the　principle　of　the　maximum　kurtosis　index,　the　low-dimensional　embedded　vectors　introduced　to　reflect　impact　failures　were　extracted　from　the　global　coordinate　feature　matrix.　Finally,　simulations　and　industrial　applications　showed　that　compared　with　the　singular　value　decomposition,　this　approach　is　effective　to　extract　not　only　the　weak　impacts　with　the　greater　kurtosis　in　time　waveform,　but　also　the　fault　feature　frequencies　in　frequency　spectrum.