Reserch In Fault Recognition Of Rolling Bearing Based On Deep Learning Network

Rolling bearing is an important part of rotating machinery,which plays a very important role in industrial equipment,and its running state has great influence on the health state of mechanical equipment.In this paper,a fault identification and condition monitoring method based on deep learning network is studied for the fault signal of rolling bearing.In order to extract the characteristic of signal and reduce the dimension of signal feature space,the theory and method of wavelet packet decomposition of signal are studied.The influence of wavelet function on the result of signal characteristic decomposition and the method of determining the number of decomposed layers are emphatically discussed.The effect of wavelet packet decomposition based on Daubechies wavelet(dbN,N =4,10,20)is analyzed.When the N is larger,the smaller the overlapping region of the filter band between the scaling function and the wavelet function,the better the separation characteristics of the small wave packet frequency domain.With the increase of N,the entropy value of wavelet packet decreases gradually.When N reaches a certain value,the entropy value of wavelet packet is basically unchanged.After comprehensive analysis,the wavelet function is determined.In addition,according to the characteristic frequency of bearing damage and Nyquist frequency,combining theory and engineering practice,the number of layers of wavelet packet decomposition is determined,which provides the basis for engineering.In order to realize the automatic fault state recognition,the fault recognition method based on BP network is studied.With the input of the original signal as the BP network,the input dimension is large,the feature is not obvious,the state recognition effect is poor and the efficiency is low.Wavelet packet transform is used to decompose the original signal,and wavelet packet energy is used as the feature,which can greatly reduce the input dimension of BP network.In addition,the Wavelet packet energy contains statistical information of time domain and frequency domain,which improves the accuracy and efficiency of recognition.However,the problem of local extremum exists in BP network.The Restricted Boltzmann machine(RBM)is a kind of deep learning network,which has high efficiency in the classification of large dimension signal.To overcome the local extremum problem of BP network,this paper presents a method based on the combination of wavelet packet decomposition and RBM network.The experimental results show that the RBM network is better than the BP network for classifying the signal with large dimension,the results of wavelet packet decomposition belong to statistic information in time-frequency domain,which contain more information than time domain signal,and the method of combining wavelet packet decomposition and RBM network improves the accuracy of fault recognition.In order to monitor the working state of mechanical equipment,a monitoring method of bearing running state based on RBM network is studied.The wavelet packet energy value of real-time monitoring signal is used as the input of RBM monitoring network to monitor the running state of mechanical equipment.When the energy value changes,it means that the condition of equipment changes,and it can be determined in time whether its running state is normal or not.