Fault Diagnosis Of Rolling Bearings Based On Improved Generative Adversarial Network

Rolling bearings are crucial components in rotating machinery,which often work under harsh conditions such as high loads and high-speed rotation,inevitably resulting in failures.If faults are not detected in a timely manner,they can seriously endanger personal and property safety.Therefore,fault diagnosis of rolling bearings is of great significance.Vibration signal processing-based detection technology is a common method for rolling bearing fault diagnosis,but often relies on the experience of experts to select the features to be extracted.With the advancement of computer technology,methods based on deep learning have begun to be applied in the field of rolling bearing fault diagnosis.Deep learning is highly influenced by data,and generally,the more samples obtained,the better the effect.However,in practice,there are problems such as label scarcity and class imbalance in rolling bearing vibration data.To solve these problems,this paper studied an improved method based on Generative Adversarial Networks(GAN)for rolling bearing fault diagnosis and developed a corresponding fault diagnosis system.The main research work is as follows:(1)To address the issue of class imbalance in the training samples of rolling bearings,a fault diagnosis method based on Constrained Auto Encoder-Generative Adversarial Network(CAE-GAN)is proposed.Firstly,an AE-GAN network model is constructed to effectively expand the number of samples.Secondly,an self-attention mechanism is added to the generator to enhance the network’s learning and generating capabilities for features.To further solve the problem of mode collapse,a sample improvement method based on distance constraints is proposed.The distance between generated samples of different categories is constrained to maintain the distribution of generated samples and prevent the generator from generating the same samples from different real samples.Finally,a residual network is used for classification.Experimental results on rolling bearing fault diagnosis show that the constrained AE-GAN model effectively improves the quality of generated samples and has higher fault diagnosis accuracy than other comparison methods.(2)To address the problem of difficult access to labeled data for actual fault diagnosis and inadequate utilization of a large amount of unlabeled data,a rolling bearing fault diagnosis model based on semi-supervised CAE-GAN is proposed on the basis of(1).This model combines the data generation ability of the CAE-GAN with the utilization ability of unlabeled data under the semi-supervised mechanism,fully utilizing unlabeled data for training and improving the fault diagnosis capability of the model in the case of insufficient labeled data.Using sample quality filters to select suitable generated samples for model training,further improving the model’s fault diagnosis ability.Experimental results show that this method has significantly improved diagnostic accuracy compared to other models in the case of few labeled samples.(3)Design and develop a fault diagnosis system for rolling bearings.Firstly,according to the overall requirements of the system,the experimental platform is built and tested.This includes formulate the data collection plan and the functional modules required for design completion.Next,Python is selected as the programming language to build the software backend,which includes functions such as realizing the functions of information input,data analysis,data processing,data generation,model training and model testing.Each functional module of the software is modularized to provide a foundation for further development of the diagnosis system.The effectiveness of the developed rolling bearing fault diagnosis system is preliminarily verified through experimental testing on the rolling bearing experimental platform.