| Rotating machinery is the most critical part of industrial production,How to ensure the healthy and smooth operation of mechanical equipment is crucial for the industrial equipment to become intelligent and unmanned.Although traditional fault diagnosis methods have achieved good results in various scenarios,in the face of today’s intelligence and big data,traditional fault diagnosis methods can not meet the requirements of intelligent mining in the coal industry.In this thesisr,the hard in extracting signal characteristics and the low accuracy of fault diagnosis classification for bearing equipment in rotating machinery under the same and different working conditions are studied.The main research contents are as follows:(1)In view of the problem that the acquisition of statistical feature parameters in the traditional machine learning bearing fault diagnosis method needs to rely on empirical selection and has poor feature extraction ability.A bearing fault diagnosis method based on wavelet packet multiscale(WPT-MSCNN)is proposed to realize the adaptive extraction of fault depth features of vibration signals.Firstly,the vibration signal is analyzed by WPT,and the time-frequency characteristic information is extracted;Then use 3×1、5×1、7×1 three convolution kernels with different sizes and scales extract the depth features in the time-frequency feature information and complete the bearing fault classification.The excellent feature extraction ability and accuracy of this method are proved by experiments.(2)In view of the complex and changeable working conditions of field equipment and the distribution differences between the fault characteristics of bearing vibration signals under different working conditions,which affect the accuracy and scene adaptability of fault diagnosis and identification,based on the signal analysis and indepth learning methods in content(1),the feature migration method is studied,and a bearing fault diagnosis method combining WPT time-frequency analysis and feature migration(WMGRNMM)is proposed.Firstly,WPT is used to analyze the timefrequency of nonlinear and non-stationary vibration signals,and the wavelet packet time-frequency characteristic diagram(WPT-TFF)of vibration signals is constructed.Then,according to the characteristics of small size and scattered features of wpt-tff,a multi group parallel Res Net network structure(MGRN)is designed to extract the depth features of WPT-TFF.Finally,the multi-core maximum mean difference(MK-MMD)is used to analyze the distribution difference between the bearing source domain vibration signal depth features and the target domain signal depth features under variable working conditions,and the depth feature extraction network is optimized to make the extracted depth features have better cross domain invariance.Experimental results express that the WMGRNMM model has excellent fault classification accuracy and generalization ability under variable working conditions.(3)When the spatial distribution difference of multiple faults is too large,the constraint ability of feature migration measurement parameters on the distribution difference is weakened,which will lead to the decline of the diagnostic ability of WMGRNMM model.To solve this problem,a method of anti migration bearing fault diagnosis based on wide multi-scale domain(WMSDANN)is proposed.First,use layer32 × 1 large convolution kernel extracts the global features of vibration signals and then uses them 3×1、5×1、7×1 three different convolution kernels obtain the fault feature information of different scales in the global feature;Then,the characteristics of source domain and target domain are confused through domain confrontation network,so as to achieve Nash balance between them,align the distribution between source domain and target domain,and solve the problem of low accuracy of WMGRNMM migration diagnosis caused by too large difference in data distribution between the two domains.Experimental results show that WMSDANN has good off design multi-source fault diagnosis ability and strong model generalization ability.The thesisr has 44 pictures,25 tables,and 100 references. |
