Research On Rolling Bearing Fault Diagnosis Method Based On Multi-level And Multi-output Deep Learning Model

The failure of the rolling bearing directly affects the production efficiency of the factory and personal safety,so it is indispensable for long-term monitoring of the healthy state of rolling bearings to ensure its safe operation.The early rolling bearing fault diagnosis method has become stretched today,and the characteristics of artificial extraction faults are becoming more and more difficult under complex conditions.In view of this,this article uses the rolling bearing vibration signal as the input,and the multi-level and multi-output deep learning model is used to achieve the end-to-end to high-efficiency and accurate diagnosis of rolling bearings in complex scenarios such as high noise,changing conditions and unbalanced data.The main research content is as follows:(1)In response to the real-time problem of fault diagnosis,a multi-level fault diagnosis model based on convolutional neural networks and residual networks is proposed.First,the shallow convolutional neural network is used for abnormal detection.The classification of faults not only realizes the end-to-end diagnosis of rolling bearings,but also reduces the computing resource needs of long-term monitoring through multi-level diagnosis.The results show that the proposed method can greatly reduce parameters and calculate quantities when the accuracy is as high as 100%.(2)In response to the existence of actual industrial data,and the failure of rolling bearings in daily equipment maintenance,it is also extremely important.In order to realize the identification of the type and failure of the rolling bearings in the same situation and the strong noise scenario,the CWRU dataset after applying the noise ratio of different signal-to-noise ratios is used to verify it.The accuracy of the classification accuracy of faulty type and fault size can reach 97.60%and 96.87%,respectively.Finally,it explains the explanatory explanations of deep learning based on the model proposed.(3)In response to the problem of imbalance of industrial data,build a raw data set with a proportion of normal samples to fault samples of 8: 1,and then convert the vibration signals of various types of faulty samples into gray graphs,and then use deep convolution to form After the sample expansion of the network is expanded,then turn back to the vibration signal.Finally,the generated sample data and the original data are used to build a normal sample and fault sample ratio of 4: 1,2: 1,1: 1 as a training set.Use it.The 4 normal samples and faulty samples constructed into different data sets with different data sets input into the same model training and used the same test set for testing.The results show that the accuracy of the test accuracy after the data set training is 75.40%,and the confusion is confusing.The main reason is that the main reason is that a large number of faulty samples are identified as normal samples,and the test accuracy rate of test accuracy can reach 100%after data set training after data expansion.