Research On Fault Diagnosis Of Rolling Bearings Under Noise Environment And Unbalanced Samples

Rolling bearings,as the most common support parts in various large-scale machinery and equipment,often work under high-strength and high-speed working conditions.Therefore,they are prone to faults,which will cause the entire machinery and equipment to stop.In order to avoid such accidents,it is very necessary to conduct in-depth fault diagnosis research on rolling bearings.At present,the research methods of rolling bearings fault diagnosis are mainly divided into three categories: based on dynamic modeling,based on signal analysis,and based on data drive.Among them,the fault diagnosis method based on deep learning is an emerging branch of the data-driven method.Because of its excellent fault diagnosis ability in the face of big data,it has become the mainstream in current rolling bearing fault diagnosis methods.Unfortunately,complex noise environments,unbalanced training samples,and occasional new faults under actual working conditions will add a lot of difficulty to fault diagnosis based on deep learning,thereby affecting the final accuracy.At the same time,since most deep neural networks need to consume a lot of computing resources and time during training,if the deep neural network is retrained when the model is improved and upgraded,it will cause a great waste of resources.In view of the above-mentioned difficulties,this paper studies the fault diagnosis method of rolling bearing with low energy consumption optimized and upgraded in the case of complex noise environment and unbalance samples.The specific research content is as follows:(1)A fault detection framework for rolling bearings of large-scale mechanicals is proposed in this paper,which aims to solve the problem about fault diagnosis of rolling bearing in noise environment.This fault diagnosis framework combines a one dimensional de-noising convolutional auto-encoder(DCAE-1D)and a one-dimensional convolutional neural network(CNN-1D)to form a DCAE-CNN model,which is compared with the DAE-CNN model and the CNN-1D model under training sets and test sets of different signal noise ratios(SNR).Finally,the experimental results demonstrated the effectiveness of the DCAE-CNN.(2)Two fault diagnosis models(ACGAN-D and ACGAN-CNN)are proposed in this work,which both based on Auxiliary Classifier Generative Adversarial Networks(ACGAN),to solve the problem of sample imbalance and occurrence of new types of failures.This paper combines ACGAN-D model,ACGAN-CNN model,LSTM networks,BP networks,and CNN-1D model under the condition of imbalance samples with a 10:1 sample ratio in the training set.It is found that the fault diagnosis accuracy of ACGAN-D model and ACGAN-CNN model are both nearly 3% higher than CNN-1D model.Moreover,the main classifier of D network in ACGAN-D model and ACGAN-CNN model can effectively distinguish whether the input sample belongs to the existing sample type.Therefore,these two models based on ACGAN network have stronger robustness in the fault diagnosis of rolling bearing.(3)A novel fault fault framework for rolling bearings,which based on the pre-training network theory and DCAE-ACGAN model is proposed in this paper.This framework aimed at the problem of high energy consumption of the above-mentioned deep neural network model when training.Firstly,the trained DCAE-ACGAN model,which can classify the four failure states of rolling bearings and trained under the condition of 30 Hz rotation speed,is upgraded to a new model that can classify the six failure states of rolling bearings,under the conditions of low energy consumption and little time.And then,on this basis,the new model is migrated between different speed conditions.Finally,the effectiveness of migration models is verified through experiments.