Abstract: To solve the problem of crack detection of large turbine blades, the author proposed a non-contact online acoustic health monitoring system and studied in-depth on the adaptive crack feature extraction method. Firstly, a preprocessing algorithm is well designed to remove the complex background noise. Then 1/6 octave technique is used to reveal the spectrum change of acoustic signal roughly, and concluded that the octave energy ratios are extracted as input feature vector of the support vector machine classifier. Finally, the principal component analysis is introduced to optimize the high dimensional feature space adaptively. The measured data from wind field validates the effectiveness of proposed method.

Key words: turbine blades, fault diagnosis, feature extraction, 1/6 octave, principal component analysis