Abstract: Wavelet packet analysis overcomes the shortcoming of wavelet analysis in poor resolution of high frequency band, thus providing a more accurate decomposition of signals. Basis on this analysis, a method for signal feature extraction is described for fault diagnosis of hydropower units, which combines wavelet packet energy spectrum with power spectrum analysis. Extraction and analysis of the in-situ test data and vibration fault signal at a hydropower station shows that this extraction method is accurate and efficient. It is a method of high application value and would lie a basis for accurate unit fault diagnosis.