Study On Fault Diagnosis Method Of Rotor System Based On Modified Generative Adversarial Network

Rotating machinery is a common equipment in many fields,such as aerospace,chemical industry,energy,metallurgy and others,which plays an important role in modern industrial production.As the core component of rotating machinery,fault diagnosis research of rotor system is of great significance for maintaining smooth operation of equipment system,avoiding economic property loss and casualties.Most mechanical equipment operates under normal conditions,so the absence of fault samples causes imbalance of data categories,which results in that existing fault diagnosis models extract their characteristics ineffectively.Therefore,imbalance of data samples is an urgent problem to be solved in the process of fault diagnosis.Generative adversarial network(GAN)by countermeasure mechanism to train in-depth learning model and generate new data samples,which has achieved great success in image processing field.However,traditional GAN has problems of mode collapse and gradient disappearance,and there is little research on vibration signal generation of GAN.In order to solve existing problems,this paper uses the data-driven deep learning method to extract the hidden features of vibration signals through deep network structure and carry out the research on rotor system fault diagnosis.Through a lot of experiments,the feasibility of generating vibration signal by improving generative adversarial network’s method is studied.At the same time,the fault diagnosis under unbalanced data is studied by combining two-dimensional convolution neural network,a method of fault diagnosis based on improving generative adversarial network’s method is proposed.The main research contents are as follows:(1)In order to solve the problems of mode collapse and gradient disappearance in traditional GAN,supervised learning and Wasserstein distance are introduced and a new network structure and training method are designed,which constrains GAN training to generate specified samples,and an improved GAN method for generating vibration signal is proposed.(2)Aiming at the problem that fault characteristics are difficult to extract,data preprocessing is carried out by time-frequency analysis method,and one dimensional time series vibration signal is transformed into time-frequency picture and input into convolution neural network.Sample characteristics are extracted and classified by convolution neural network,and a model of two-dimensional convolution neural network fault diagnosis is proposed.(3)Aiming at the problem of data imbalance caused by few fault samples in fault diagnosis,combining with improved GAN method,a method of two-dimensional convolution neural network fault diagnosis based on generated data is proposed,the fault diagnosis effect of this method under different unbalanced proportion data sets is analyzed,and the feasibility of solving the problem of data imbalance is verified by experiments.(4)Based on the integration of Python and Py Qt5 development environment and the model proposed by this paper,a set of rotor fault diagnosis system software is designed and developed.