Research On Diagnosis Scheme Of Small Sample Data Of Rolling Bearing Faults Based On Deep Learning

As a kind of commonly used support and rotating parts,the health state of rolling bearings has a very important impact on the safety and maintainability of the whole machine.If this component has local wear,fracture,corrosion,fatigue peeling and other faults during operation,the staff will face serious safety hazards,and the manufacturer will also face huge economic losses.In order to accurately detect the dangerous conditions of rolling bearings,reduce the risk of mechanical operation,and improve the efficiency of industrial production,it is necessary to monitor the health state of rolling bearings and identify their failure modes.The traditional fault diagnosis methods mainly adopt the manual fault feature extraction mode,whose diagnosis accuracy is closely related to the experts’ practical engineering experience and signal knowledge reserve.In comparison,deep learning technology can use self-learning method to represent fault characteristics,realize intelligent fault diagnosis and reduce the degree of human participation,which has obvious advantages.However,in industrial production,it is often difficult to collect fault information,which leads to poor diagnostic performance based on deep learning.Aiming at the small sample condition,this paper proposes a new diagnosis method based on FFTGAN-SSAE model to realize the effective identification of rolling bearing fault modes.The main research work is as follows:Firstly,the FFT method is used to convert the CWRU bearing monitoring signals into frequency domain to form frequency domain samples,and compare and analyze the characteristics of time and frequency domain rolling bearing samples.Based on the signal frequency domain information,GAN is used to generate new samples similar to the frequency domain samples to provide sufficient and high-quality training sample data for subsequent diagnosis models.Secondly,the structure of the SSAE diagnostic model is designed,and the optimal parameter selection is analyzed through repeated experiments.For different stages of rolling bearing signal samples,T-SNE technology is used to conduct two-dimensional visualization of their features,and the separability presented by the features is analyzed and compared.The optimal parameters are used to determine the SSAE fault diagnosis model,and the diagnosis results are compared with those of FFT-GAN-PCA-SVM model and FFT-GAN-2D-CNN model.Finally,the bearing signal measured by the wireless sensor device is used to verify the fault diagnosis method proposed in this paper.By adding different numbers of generated samples to the training samples,the failure pattern recognition experiments were carried out to compare the influence of the number of samples on the accuracy of model diagnosis.At the same time,GAN is used to generate time-domain samples,and the recognition accuracy of the SSAE diagnostic model is compared when the corresponding generated samples are added to the time and frequency domains as a whole.The experimental results based on the above two different data sets show that in the case of insufficient samples and small numbers,the method proposed in this paper can identify different failure modes with high accuracy and has better stability.