Intelligent Fault Diagnosis Of Rolling Bearing Based On Wavelet Threshold Filtering And Neural Networks

Rolling bearing is one of the most widely used accessories in the mechanical equipments, the running state of rolling bearing will have direct impact on the performance of the whole mechanical equipments and the occurrence of fault. Therefore, how to diagnose the state of rolling bearing,whether it fails or not and the sorts of fault within a short time and without disturbing the running of the whole system has become a urgent problem to be solved, which is also a hot spot of the field of condition monitoring and fault diagnosis.Now, vibration signal analysis is a major mean to realize condition monitoring and fault diagnosis of rolling bearing. There are many different theories and methods to process vibration signal, such as FFT,Cepstrum,Wavelet analysis,Hilbert-Huang etc, among which Wavelet analysis is very widely used compared with other methods. At present, combining the methods of wavelet transform and envelope demodulation to get wavelet high-frequency coefficients of envelope spectrum of different levels, thus the character of fault can be got out, but the fault character is not very good. Wavelet transform can realize the function of a group band filters and get the signals of different bands, to the signal of every single band, the noise out the signal of this band can be filtered, however, the noise in the signal of this band is still high, which is bad for extracting good fault characters in the later stage.To overcome these shortcomings, this paper proposes the method of threshold value filtering to filter the noise in the early stage after careful study on the theory of wavelet ,and aiming at promoting the defects of hard-threshold function and. soft-threshold function, this paper propose a new threshold function, which can overcome the defects of burrs bringing by hard-threshold function to some extent and the defects of approaching problems between the signal being filtered and the real signal.This function is adopted to realize filter stimulation imposed on the leleccum with white gaussian noise before the practical application, the SNR and SME of the method of hard-threshold filtering,the method of soft-threshold filtering and the new threshold function are presented. It proves that the new threshold function can successfully overcome the defects of hard-threshold function and soft-threshold function, thus promote SNR of whole signal.During the practical application, this paper firstly adopt the new threshold function to filter the noises of vibration signals of various working conditions, then envelope demodulation is applied to get envelope spectrum of the vibration signal being filtered, finally, a comparison is made between envelope spectrum of the signal being filtered by the new threshold function and envelope spectrum of wavelet coefficients with high frequency level. The results proves that noises can be filtered by the method of threshold filtering, also provide a good condition for extracting good fault characters for recognition.In this paper, fault characters are dealt by the method of energy normalizing, the results of which are used as the input of BP neural networks. The fault recognition is carried out by BP neural networks , the fault recongnition rate of which is very high.