| The fault diagnosis of rotating machinery is critical to the safety and reliable operation of huge mechanical equipment.As rotating machinery is serving in a complex environment for a long time,it is prone to breakdowns,and even causes huge accidents and financial losses.In recent years,the issue of fault diagnosis of rotating machinery has received an increasing attention from domestic and foreign scholars.Data-driven technologies continue to be developed and many results have been achieved in the application of deep learning models in fault diagnosis.However,the traditional deep learning models still reveal many problems during the application in real industrial scenarios.On the one hand,it is difficult to collect sufficient marked fault data in complex mechanical equipment,and the limited data constraint the application of intelligent fault diagnosis algorithms.On the other hand,mechanical equipment often works under variable speeds or variable load conditions due to carrying different tasks.The target domain data and the source domain data have different distributions,thus deep learning-based diagnostic methods would show the limitations of poor performance and generalization ability.This paper is based on the deep learning approach to carry out research on fault diagnosis in the case of imbalanced monitoring data and transfer diagnosis under variable working conditions.The main research includes the following three parts:To address the problem of data imbalance in real industrial scenarios,this paper uses generative adversarial networks(GAN)to solve the problem.Concretely,a generative model based on spectral normalization(SN)and Wasserstein generative adversarial network(WGAN)is constructed,which is mainly built by one dimensional convolutional neural network(1-D CNN)and fully connected layers.At first,the fault data are fed into the generator network,and the generator learns the distribution of the original data.The generator and discriminator continuously optimize the quality of the generated data by dynamic confrontation until the network reaches Nash equilibrium.Secondly,the generated data are mixed with the original data to reconstruct the training set,thus alleviating the impact of data imbalance.After that,we use a feature extractor built by a deep 1-D CNN for feature extraction on the training set data.And besides,we set Softmax as the activation function to perform fault diagnosis at the decision level with respect to the feature information.Finally,the performance of the proposed method is validated by conducting tests on the dataset CWRU.Aiming at the problems of instability of network training and poor learning ability of generators exposed during the application of adversarial generative networks in the field of fault diagnosis,this paper proposes to improve the generative model.Specifically,we introduce Gradient Normalization(GN)Mechanism and Full Attention Mechanism(FA)to improve the WGAN model.Gradient normalization is a model-level constraint,which not only makes the setting of discriminator layers and parameters more flexible and simplified,but also ensures the stability of the network,which in turn allows the network to generate high-quality data.The full attention mechanism can compensate for the small perceptual area of convolutional neural networks by making them pay more attention to the deep discriminable features of the input data,which could enhance the generation capability.The experimental part is tested using CWRU bearing fault diagnosis data set and real bearing fault diagnosis platform,and the final fault diagnosis accuracy of the proposed method tested on the real bearing fault diagnosis platform is 95.0%.Compared with the comparison algorithms,the proposed method has an average improvement of 3.92%in diagnostic accuracy,and the results indicate that the improved generative model can address the fault diagnosis problem under data imbalance more effectively.To deal with the problem of low diagnosis accuracy under variable working conditions such as variable speeds and variable loads that often occur in real industrial scenarios,this paper presents a domain adaptation method that incorporates multi-kernel maximum mean discrepancy and conditional domain adversarial network.To begin with,the target domain data and the source domain data are passed through a network of feature extractors to obtain the transferable features.Furthermore,a multi-kernel maximum mean difference mechanism is introduced to enable features to be matched in the training phase of the network,reducing the distance between the distribution of features of the same class of data in the source and target domains.After that,we use conditional domain adversarial networks to reduce the distribution differences between source and target domain data by simultaneously considering both feature information and label information during the training process,in order to improve the diagnostic accuracy of the model on target domain data.Finally,the proposed method is tested on the publicly available Jiangnan University bearing fault diagnosis data set and the 2009 PHM gearbox fault diagnosis platform,and the experiments illustrate that the proposed method can effectively solve the fault diagnosis problem under variable working conditions. |
PDF Full Text Request