Fault Diagnosis Of Rolling Bearing Based On Time-frequency Analysis And Deep Neural Networks

Rolling bearings are key motion transmission components in rotating machinery and their performance can have a direct impact on the operational safety and reliability of the entire system.The failure of a rolling bearing can lead to the complete failure of the entire machinery and equipment,resulting in huge economic losses and even casualties.Therefore,in order to ensure the stable operation of mechanical equipment,timely detection and diagnosis the rolling bearing fault,this thesis takes rolling bearing as the research object,focusing on the bearing vibration signal fault feature extraction technology,bearing fault diagnosis technology under small sample conditions and noise environment.The main research content is divided into the following parts.First of all,this thesis starts from the rolling bearing failure mechanism and analyses the typical failure causes and failure types of bearings.On this basis,a typical fault simulation experiment for rolling bearings was designed to collect fault vibration data and build up data sets.The modal decomposition of the vibration signal is first carried out using the local mean decomposition method,and then the component signals containing fault characteristics are selected for envelope analysis based on the kurtosis criterion.The method effectively extracts the characteristic frequencies of typical bearing faults and achieves a preliminary screening of bearing faults based on the characteristic frequencies.To address the limitations of frequency domain features,this chapter investigates the time-frequency domain analysis method of bearing vibration signals and establishes a time-frequency image dataset of bearing fault vibration signals for subsequent fault diagnosis work.Based on the time-frequency image data,this thesis goes on to investigate the application of Convolutional Neural Network（CNN）in bearing fault diagnosis.Based on the principle of CNN algorithm implementation,this chapter proposes the use of timefrequency images as the input of CNN in response to the problems of poor utilization efficiency for vibration signals and large demand for network training samples.Next,this chapter builds a joint short-time Fourier transform-convolutional neural network（STFTCNN）fault diagnosis model applicable to bearing fault vibration data.The model is also optimized according to the characteristics of the bearing fault data by comparing the influence of the model structure and algorithm on the fault diagnosis results.The experiment proves that the model can achieve higher diagnostic accuracy with significantly reduced sample size,and the fault diagnosis accuracy of the experimental data reaches more than 99.7%.Finally,this thesis focuses on the bearing fault diagnosis algorithm in noisy environment and proposes a joint CAE-DRN fault diagnosis model based on Convolutional Auto Encoder（CAE）and Deep Residual Network（DRN）.Firstly,this chapter builds and optimizes the STFT-DRN model for bearing fault diagnosis in noisy environment with the strong feature learning ability of deep residual network and the characteristics of bearing vibration signal in noisy environment.The experiments verify that this model has better noise immunity performance compared with STFT-CNN under moderate noise intensity,and the fault diagnosis accuracy is above 90% in both cases,but still underperforms under large noise environment.To address this problem,this section proposes to use CAE to perform noise reduction on the time-frequency data in a noisy environment,and then feed the noise reduced data into the DRN for fault diagnosis.It is proved that the CAE-DRN model can achieve more than 88% fault diagnosis accuracy under the-9.5d B signal-to-noise ratio condition.