Rotating Machinery Fault Diagnosis Based On Improved Generative Adversarial Networks

Condition monitoring and fault diagnosis of machinery and equipment is an important means to realize the safe operation of modern industry.With the development of computer,sensor and communication technology,the condition monitoring of machinery and equipment has entered the era of big data,which brings new challenges to mechanical fault diagnosis due to the massive amount of data,various types and extremely fast generation speed.Therefore,the use of new theories and methods to implement monitoring and diagnosis of mechanical faults under big data to ensure its accuracy and efficiency has become a hot research direction at present.Deep learning-based artificial intelligence technology is a powerful way to achieve intelligent defect detection by mining deep aspects of data and gaining knowledge autonomously.However,in the actual working environment,the limited fault data collected restricts the accuracy and stability of diagnosis.Generative adversarial networks(GAN),as a branch in the field of deep learning research,is currently the most effective way for dealing with data scarcity.In this paper,with the goal of improving the diagnostic accuracy and reliability of diagnostic results,we investigate the technique of generating fault samples and diagnostic methods for rotating mechanical equipment based on generative adversarial networks.To address the issue of insufficient fault data,a data enhancement method(CWT-GANSP)based on two-dimensional time-frequency maps and quadratic potential generation adversarial network is proposed.Firstly,the original one-dimensional time-series signal is converted into a two-dimensional time-frequency map using continuous wavelet transform(CWT)to extract the fault features more accurately.Based on this,the data is enhanced by generating an adversarial network based on the quadratic potential.Through the training of the generative adversarial network,new data with realism and diversity can be generated more accurately with better signal-image conversion adaptability,which can improve the fault diagnosis accuracy more effectively.To address the problem that existing deep neural networks are susceptible to interference from fault-independent features,resulting in poor model generalization and thus reducing the accuracy of fault diagnosis,we propose incorporating residual self-attentiveness into the auxiliary generative adversarial network(ACGAN)and developing a fault diagnosis model based on the residual self-attentive generative adversarial network(RSA-ACGAN).Fault features are identified on different time and frequency scales for signals,and attention is adaptively focused on different moments to generate high-quality samples to mitigate crosstalk between different fault types.Semi-supervised learning is introduced to address the problem of low diagnostic accuracy of fault diagnosis models in scenarios with limited number of markers.Specifically,semisupervised generative adversarial network(Semi-RAGAN)uses RSA-ACGAN to generate high-quality samples to expand the training set.The auxiliary classifier is used as a pre-trained model for the fault diagnosis model and fine-tuned using generated and unlabeled samples.The main advantage is that high quality sample generation is guaranteed and unlabeled data is used to improve the performance of the fault diagnosis model.Compared with traditional GANbased fault diagnosis methods,the Semi-RAGAN methodology dramatically reduces the amount of data that must be labeled,increasing the model’s efficiency and scalability..