Abstract:This paper deals with the analysis of the bearing fault mechanism and the extraction method of fault characteristic parameters.A wavelet analysis method was introduced and an improved wavelet packet algorithm was put forward to reduce the frequency aliasing in wavelet analysis after a full consideration of the nonlinear system and the uneven surface vibration signals of rolling bearings.The improved wavelet analysis method largely avoided the frequency aliasing phenomenon and overcame the problem of indistinguishable high and low frequency overlapping in the traditional wavelet packet algorithm. The new analysis method can also separate the noise signal containing the fault signal by using wavelet frequency band.A structural model of the improved wavelet neural network was built by the combination of the advantages of wavelet and neural network.To solve the problem that the convergence speed of traditional BP algorithm is slow and easy to fall into local minima,the algorithm of wavelet neural networks was studied so as to improve it from two aspects: the learning rate and the connection weights.A simulation was then carried out and in the process, N205 type rolling bearing was tested on the test bench and the test data were used in the training network.The results of the network training were obtained by using the vibration signal as input vector for the network.Through the simulation example,it can be found that the improved wavelet neural network can well classify faults,and its convergence speed is obviously faster than of the wavelet neural network under the same condition and the improved BP network, which proves that it can effectively diagnose the faults of rolling bearings.