| Rolling bearings appear in many equipment as a key component,its fault has become a hot topic of research at home and abroad,divided into single fault and compound fault diagnosis.Intelligent diagnosis of rolling bearings has become a new trend in the future.The combination of time-frequency analysis and Convolution Neural Networks realizes the application of deep learning in signal processing,and the signal classification can be realized without relying on the experience of signal processing experts.Therefore,time-frequency analysis technology is used for fault identification of composite faults.For a single fault,the transfer learning of Convolutional Neural Network is used to solve the problem of fault type recognition under small samples,and it is possible to realize intelligent online diagnosis at the same time.The main research contents are as follows:(1)For the composite failure of the rolling bearing,the parameter k and alpha of the variational mode decomposition are optimized by using the surface-simplex swarm evolution optimization algorithm,and the appropriate modal functions are broken down to avoid the problem of over-decompositon and the under-decompositon.According to the mean square error and the permutation entropy,the method of the decomposition of the decomposition is reconstructed,and the denoising of the composite fault signal is realized.First,the signal is used to analyze the frequency of the internal and external loop failure by using the single sub-band to reconstruct the wavelet packet and the auto-correlation envelope spectrum.Secondly,the denoising signal is analyzed by the surface-simplex swarm evolution optimization algorithm to optimize the analysis of the signal of the independent component analysis,and the failure frequency of the internal and external loop is found through the spectral correlation diagram.Through the above two methods,the frequency recognition diagnosis of the compound failure of the internal and external loop is realized,and the error value of the theoretical calculation frequency is compared to the error value of5%.(2)This thesis mainly focuses on the diagnosis of single fault classification and recognition under small sample data.Firstly,the four one-dimensional data of the normal state of the rolling bearing,the inner ring fault,the outer ring fault and the rolling body fault are converted to the two-dimensional time-frequency samples.Secondly,classical Convolution Neural Network is used for transfer learning,and the network is retrained by modifying the network layer.Finally,the parameters of the network are optimized to select the higher values of network identification parameters.VGG(Visual Geometry Group)Convolutional Neural Network is used to transfer and learn,and convolutional processing of different sizes is designed for the same input in the full connection layer.The results show that the two-layer 3×3convolution processing structure has a high average recognition rate for the time-frequency and spectral kurtosis maps of the continuous wavelet transform,which is more convenient for classification.The classification of transfer learning solves the classification problem under the small sample data,avoids the limitation of expert experience and provides the possibility to realize the online intelligent monitoring.(3)The feature extraction of single fault data with three different fault diameters is done by using the wavelet packet reconstructed by single sub-band,sample entropy,and then it is made by using adaptive mutation particle swarm optimization support vector machine classification.Wavelet packet and permutation entropy are better than sample entropy and wavelet packet in fault feature extraction. |
