Cross-domain Fault Diagnosis Of Rotating Machinery Based On Deep Learning And Transfer Learning

Modern industrial equipment has become large-scale,complex,and intelligent.Ensuring that industrial equipment operates in a safe and stable state has become a hot spot for researchers in recent years.Key components of industrial equipment that rotate,including gears and bearings,are necessary.Intelligent monitoring,fault diagnosis,and life prediction for industrial machinery are of the utmost importance to ensure the safe and reliable operation of the machinery and prevent accidents.Data-driven fault diagnosis technology can adaptively extract fault features and identify fault modes from big data and can realize comprehensive intelligent fault diagnostics in dynamic environments.Although data-driven fault diagnosis has recently seen some success,there are still a few issues to be resolved: First,the hyperparameters of the current data-driven fault diagnosis network are mostly set by experts based on experience.Second,the fault signal is limited and readily contaminated by outside noise due to the lack of defects at the real diagnosis site.Overfitting is a common problem with the classic diagnostic network since it cannot fully extract useful information from the noise.Finally,the real industrial scenarios have changeable equipment operating conditions and severely inadequate labels for the defect data that have been gathered.The challenges of multi-domain mismatch and domain negative transfer are still insurmountable for the fault diagnosis approach of single-source domain adaptive transfer.The problem of challenging deep neural network hyperparameter configuration is a common challenge of fault diagnosis models based on deep learning,while the other two problems are industrial problems.To conduct out research,this topic chooses the time-frequency analysis approach,deep neural networks,and transfer learning strategies.The following are the primary research findings:(1)The adaptive optimization of fault diagnosis network hyperparameters is researched under the guidance of the synergistic enhancement of network classification accuracy and training effectiveness.The Deep belief network’s learning rate,batch size,and other hyperparameters are combined and optimized using the Sparrow search algorithm(SSA),which uses the Deep belief network as the branch network and the network’s root mean square error as the optimization index.The goal of the comparative experiments is to demonstrate the sparrow search algorithm’s superiority to alternative optimization methods,such as particle swarm optimization.The classification diagnosis using DBN after SSA hyperparameter search provides higher diagnostic accuracy and faster diagnosis speed than standard DBN due to the verification of many data sets.(2)In the background of few samples and high noise,useful information is often submerged in noise and network training is often overfitting.Based on the autoencoder(AE),from the perspective of reconstructing the network error,zeros are randomly set at the output layer nodes of the AE,and interference items are added to the network loss function to prevent network overfitting.To enhance the network’s capacity to extract useful features,a stacked denoising autoencoder(SDAE)is built,and its hyperparameters are integrated and tuned globally.SDAE is used in the experiment to extract features,categories,and diagnose the samples after adding Gaussian white noise to the original vibration signal and a small number of samples.The findings demonstrate the superior resilience and noise resistance of the fault diagnosis model based on optimized SDAE.(3)A cross-domain fault diagnosis methodology based on multi-source domain feature adaption and selection is created for the cross-domain transfer task of few-sample labels.To quickly obtain multiple feature adaptive extractors(FAEs),which are used to fully mine the inter-domain transferable features in the data,a multi-branch two-dimensional convolutional neural network(2DCNN)is used as the backbone network.To solve the domain mismatch problem,the metric value of the Multiple Kernel Maximum Mean Discrepancy(MK-MMD)between the source and target domains is integrated in the loss function of FAEs.As the network loss converges,the inter-domain distance is shortened,and the inter-domain offset is reduced.To solve the domain negative transfer problem,ensemble learning,and information entropy gain are introduced to perform ensemble selection on the adaptive feature matrix,intercept effective features,and align feature distributions.Finally,the beneficial features are categorized and identified using the fully connected network.Using multiple data sets to construct data sets with different rotating components and working conditions,from the perspective of feature visualization,compared with single-source domain transfer methods and other domain adaptive networks,the algorithm suggested in this research can improve the degree of domain matching and prevent negative transfer.