Abstract: The attempt of using acoustic emission signal detection to carry out the turbine blades metal fatigue crack on-line monitoring has been made.Acoustic emission signal parameters are acquired by using SAMOS Acoustic Emission Testing System of the American PAC Corporation;In actual large turbine environment, four kinds of acoustic emission signals are selected. Combining BP neural network and pattern recognition,a feature extractor is designed to extract the metal fatigue characteristics of acoustic emission signals.Compared the sensitivity of input parameters to output results of neural network, several most effective parameters are chosen for identification and classification;and the separableness criterion is used further confirm its accuracy. Finally, in total 13 characteristic parameters of acoustic emission, five parameters, such as centroid frequency, counts, duration, rise time and average signal level(ASL) can be most notably used to identify acoustic emission signal in actual environment.