基于ITR-Net多源域迁移学习的高铁轴箱轴承故障诊断OA

Objective Efficiently assessing the health status of axlebox bearings in high-speed trains is crucial for maintaining reliable train operation.Cur-rent deep learning-based bearing fault diagnosis faces two significant challenges:it requires many labeled actual fault samples,and the training and test sets need to satisfy independent and identically distributed conditions.Transfer learning relaxes the limitations of these issues for intelli-gent bearing fault diagnosis,and it utilizes transferable knowledge learned from existing labeled datasets to accomplish tasks within different but similar unlabeled datasets.However,the current transfer learning model based on a single source domain suffers from underutilization of labeled data,reduced transfer diagnosis accuracy,and potential negative transfer when the dataset distribution varies significantly.This study proposes ITR-Net(Inception Transformer and ResNet),a multi-source domain deep transfer learning method that integrates IFormer(Inception Trans-former)and ResNet for high-speed train axlebox bearing fault diagnosis research. Methods The method selected supervised labeled data under various operating conditions in the multi-source domain,and first obtained the time-frequency spectrograms of the one-dimensional vibration signals of the bearings as the model input by using the continuous wavelet transform based on the Morlet wavelet basis.The main structure of the proposed network framework consisted of three parts,namely the common feature extractor,the specific feature extractor,and the specific classifier.The common feature extractor adopted the IFormer network,which used the classical structure of the convolutional neural network(CNN)with depth-wise separable convolution(DWConv)and maximum pooling to cap-ture the local information of the input data.It employed the multi-head self-attention(MSA)mechanism in the Transformer network to capture the global information of the input data,so the IFormer network mined more comprehensive feature information.The common feature extractor was utilized to extract domain-invariant features in different source and target domains.The specific feature extractor adopted the classical convo-lutional neural network ResNet,which efficiently extracted the feature information of the input patch while effectively avoiding gradient disap-pearance or gradient explosion that can have occurred with the increase of network depth.The specific classifier was utilized to output the classification results for different source domains and the target domain,which facilitated subsequent metrics to measure the distance between the different predicted labels output.In applying the transfer strategy,the study optimized the multi-kernel maximum mean difference(MK-MMD)after the common feature extractor to align the overall distributions of the source and target domains;optimized the local maximum mean difference(LMMD)after the specific feature extractor to enable the model to extract fine-grained information from the input features;optimized the cross-entropy loss(CEloss)after the spe-cific classifier to improve the model classification accuracy on the source domains;and optimized the mean-squared error(MSE)loss after the specific classifier to reduce the differences between the predicted labels of the target domain output by different classifiers. Results and Discussions Six multi-source domain transfer tasks were set using the Integrated High-Speed Train Bearing Experiment Station and the Integrated Power Transmission Fault Diagnosis Experiment Station datasets to demonstrate the effectiveness of the proposed method.Analyz-ing the results of multi-source domain transfer and single-source domain transfer showed that the effect of multi-source domain transfer was sig-nificantly better than that of single-source domain transfer.Comparing the proposed method ITR-Net with other popular transfer learning meth-ods,namely deep adaptive networks(DAN),joint adaptation network(JAN),correlation alignment(CORAL),domain adversarial neural network(DANN),and Multi-feature spatial adaptation networks(MFSAN),the proposed method achieved an average transfer accuracy of 96.66%in six transfer tasks,while the comparative methods achieved 87.24%,88.30%,92.45%,94.11%,and 93.35%,respectively.This result demonstrated the supe-riority of the proposed method.The t-distribution stochastic neighbor embedding(t-SNE)visualized the feature clustering of the target domain features extracted from the six migration tasks.It was observed that the target domain features in the proposed method achieved more distinct clustering based on different bearing fault types,and the overall clustering of the unsupervised target domain features under the same fault types was improved,which proved the method's effectiveness.In the ablation experiments,the average transfer accuracies of using MK-MMD,LMMD,and MSE alone were 92.30%,93.19%and 93.18%,respectively;when MK-MMD and LMMD were utilized together,the average migration accuracy reached 95.26%;when the complete loss function was applied,the average accuracy reached the maximum of 98.63%.The ablation results proved that the adaptive migration strategy constructed using MK-MMD,LMMD,and MSE further enhanced the degree of domain feature alignment among multi-source domains,as well as between individual source domains and the target domain,resulting in the best migration learning effect. Conclusions The results showed that the proposed method can fully utilize the data information of multiple source domains,and the transfer us-ing multiple source domains can effectively improve the diagnosis performance of faults in the target domain.The distributions of the source do-mains and the target domains can be aligned,and the ablation experiments confirmed the effects of different loss functions on the transfer perfor-mance of the network models by applying the MK-MMD,LMMD,CELoss,and MSE loss functions to construct the transfer strategy at different network stage positions.The results provide a new approach for applying transfer learning to axlebox bearing fault diagnosis.