Abstract: The paper regards clad defect, channel defect and lack of penetration(LOP) in the FSW joints as object, makes research on application of wavelet analysis theory in feature extraction, and uses the three feature extraction methods based on wavelet packet(WP) signal component node energy, WP node coefficients, wavelet decomposition of the power spectral density(PSD) of the defects echo signal to extract the features of the three types of defects. To assess the classification performance of the feature extraction methods above by classification criteria based on Euclidean's distance, then the features can be loaded to the artificial neural network(ANN) that is used for recognition of the defects. The result shows that the feature extraction method based on wavelet decomposition of the PSD of the defects echo signal has the best classification performance, and the ANN that use the feature gets the rate of defects recognition 85.71%.