Research On Fault Diagnosis Of Rolling Bearings Based On CNN And GAN

Rolling bearings are the core components of rotating mechanical equipment,and fault diagnosis can ensure the safe operation of them.In recent years,rolling bearings fault diagnosis technology based on convolutional neural network has attracted extensive attention.Among them,in the fault diagnosis research based on short-time Fourier transform and convolutional neural network,the type of window function,window width and translation overlap width have an important impact on the signal transformation results,and further research is needed on them.On the other hand,the training of convolutional neural network models require a large amount of labeled data,but the generative adversarial network used for data enhancement has the problems of poor image data generation and weak feature extraction ability.In view of the above problems,this paper carries out research on the regulation of short-time Fourier transform parameters and the structural improvement of deep convolutional neural network models and their integrated applications,and the main research contents and results are as follows:(1)The STFT-CNN fault diagnosis model is constructed,and the influence of the relevant parameters of the window function in the short-time Fourier transform on the diagnosis result is analyzed.First,the one-dimensional vibration signal generates time-frequency images by short-time Fourier transform through five different window functions.Then,the generated time-frequency images are input to the STFT-CNN model for fault feature learning and classification.Finally,the optimal window function type,window width and translation overlap width are determined.In this study,the powerful image processing ability of convolutional neural network is used to realize the classification and recognition of bearing fault vibration signals.On both bearing datasets,the STFT-CNN model achieves a high recognition rate.(2)Aiming at the problem that the feature extraction ability of the proposed STFT-CNN model is weak,the network structure and model parameters of the STFT-CNN model are improved.The first convolutional layer adopts a 5×5convolution kernel,and the subsequent convolutional layers adopt 3×3 convolution kernels.The nonlinear expression ability of the model is improved by double-layer stacking method.By comparing and analyzing the recognition accuracy of the original network and the improved network,the effectiveness of the improved model is verified.At the same time,the combination scheme of the optimal short-time Fourier transform and the improved model is determined to establish a STFT-ICNN model suitable for rolling bearing fault diagnosis.Finally,the noise immunity performance of the STFT-ICNN model is tested under noise conditions.(3)Aiming at the problem that traditional generative adversarial networks generate poor fault images and weak feature extraction capabilities.In this study,a STFT-GAN fault diagnosis model is proposed,which adopts a deep convolutional generative adversarial network.In this paper,the generator and classifier network structures suitable for fault diagnosis are designed to improve the classification ability of deep convolutional generative adversarial networks.The model learns fault features through unlabeled data,gets rid of dependence on labeled data,and realizes the classification and recognition of fault data.Finally,experimental verification is carried out on two rolling bearing datasets.The experimental results show that the STFT-GAN model can generate fault image data and extract fault information well,and it has good fault classification ability.