Centrifugal pumps are widely used in chemical production. Cavitation happened in pumps is one of the major causes which leading to the reduction of efficiency. A diagnostic method of inception cavitation is put forward based on low frequency and high frequency characteristics of pump inlet pressure fluctuation signals. The marginal spectrum is getted through empirical mode decomposition (EMD) of experimental data and Hilbert-Huang transform (HHT). By qualitative analysis root-mean-square and marginal spectrum band energy of each intrinsic mode function can be used for cavitation recognition. But it takes too much time to recognize when the characteristic dimension is high

therefore it is necessary to quantitatively analyze to simplify characters. A four-dimensional feature vector was put into the back propagation neural (BP) networks for training and simulation

with the first and the second level root-mean-square energy value of intrinsic mode function get through EMD as high frequency feature and the 0 to 20 Hz and 20 to 40 Hz band energy values of marginal spectrum as low frequency feature. The method mentioned above has increased the recognition rate by 7.26% and 3.59% with simulation time decreasing by 77.72% when contrasting with wavelet analysis method and EMD energy entropy method. It has a strong influence on the training of networks that 3 to 9 level energy entropy of experimental data EMD varies little with cavitation conditions. So EMD energy entropy method takes much time of simulation with low recognition rate. Removing the redundant characteristics recognition rate increased by 3.59

simulation time decreased by 77.72% . During wavelet analysis characteristic vector selected varies much with flow rate

therefore recognition rate is high for the same flow rate and is low for different flow rate. For characteristic value in this paper varies not much and similarly with flow rate

recognition rate raises by 7.26% .