Research On Fault Diagnosis Of Rolling Bearing Based On Signal Processing And Deeping Learning

Rolling bearings are key components in normal operation of mechanical systems,but due to the complex working environment,they are prone to damage and fault.During the operation,bearing fault seriously affects related components and even the whole mechanical system.Bearing and other mechanical components produce regular vibration and impact in the running process.The analysis and processing based on vibration signals is of great significance to the state monitoring and fault diagnosis of mechanical equipment.With the trend of intelligent fault diagnosis technology,intelligent fault diagnosis technology that introduces deep learning has become a current research hotspot This paper mainly takes rolling bearings as the research object,takes signal processing technology and deep learning methods as the technical support,and proposes a series of methods for the intelligent and accurate bearing fault diagnosis under complex conditions and the high accuracy fault diagnosis under a small number of samples.The main contents of this paper are as follows:(1)Aiming at the bearing fault diagnosis problem under complex conditions such as noise,and the limitations of traditional envelope analysis and deep autoencode r,an intelligent bearing fault diagnosis method based on Teager energy operator demodulation and multiscale compressed sensing deep autoencoder is proposed.The method uses the Teager energy operator demodulation to obtain the demodulation envelope spectra with enhanced fault feature information.An energy Jarque-Bera statistic is designed to evaluate the quality of the demodulation envelope spectrum,and combines the particle swarm algorithm to filter multiple demodulation spectrum rich in diagnostic information.Meanwhile,a compressed sensing deep autoencoder based on unsupervised learning is proposed.The compressed coded data is obtained by compression measurement.The network weights in each stack layer are adjusted by the error back propagation,and the proposed network has unique data information fusion and feature representation capabilities.In addition,the multiscale coarse-grained procedure is integrated into the deep network by a parallel to constructe the multiscale compressed sensing depth autoencoder,and it can automatically learn comprehensive and rich multi-scale features from spectral data and identify bearing fault patterns.The experimental results and comparative analysis demonstrate that the proposed method is effective and robust in bearing fault diagnosis.(2)Aiming at the training problem of large-scale deep network model with a few samples and the limitation of traditional deep network structure,a high-accuracy fault diagnosis method based on Transformer convolutional network and transfer learning is proposed.This paper introduces the standard Transformer framework to design a Transformer convolutional network model for fault diagnosis,and makes reasonable modifications for Transformer to overcome its application limitations.The one-dimensional signal is creatively interpreted as the sequence of fixed-size patches,and learnable linear embeddings and position embeddings with position information are added to the sequence.In addition,one-dimensional convolutional neural network and classifier layer are integrated into the network model to effectively process the Transformer output sequence rich in fault feature information and achieve high-accuracy pattern classification.Transfer learning is used to improve the training efficiency of the model,and reasonable model parameters are initialized through source domain pre-training,which can quickly adapt to the target task.According to the experimental data,multiple transfer experiments are used to verify the effectiveness of the proposed method comprehensively.The results show that the Transformer convolution network has excellent model expression and fault pattern recognition ability.