Bearing fault diagnosis with extensive labeled fault data has been achieved. In engineering, most machines are in a healthy state. When a fault does occur, the machine is shut down as soon as possible, thus it is difficult and uneconomical to collect enough fault data with labels to carry out a fault diagnosis. To solve the problem, a hybrid fault diagnosis method for rotating machinery, based on variational mode decomposition energy entropy (VMD-EE) and transfer learning (TL), is proposed. First, we decompose the original signal using VMD, calculate the EE value of the modal components, then build a dataset using fault data from these components. Second, a deep residual neural network is proposed to extract high-dimensional features from the dataset and divide the data into source and target domains according to the working conditions. Finally, W-distances are introduced to dynamically evaluate the importance of the conditional and marginal distribution probabilities to minimize the loss, with a feature-based TL method being used to dynamically balance the distribution adaption and reduce the difference in the probability distributions of the two domains. The proposed method is validated using the datasets of Case Western Reserve University and a machinery fault simulator platform. The VMD-based intelligent health detection and statistical analysis solve the problem of mode mixing very well and accurately detect signal faults, or not, by E E θ , the threshold of VMD-EE. Meanwhile, the accuracy of TL-based neural network fault diagnosis is up to 99.4%, and the losses are kept at around 0.02. These results show the accuracy and robustness of the proposed method in the absence of datasets and under varying operating conditions.

Data-driven methods have been applied in fault diagnosis. However, in practical engineering, workers are more concerned with the real-time health status of bearings. And it is difficult to complete the effective training of diagnostic models with insufficient labeled fault data. Therefore, this paper proposes a modular method based on a mask kernel and multi-head self-attention mechanism for rolling bearing fault diagnosis. First, the proposed method divides the diagnosis into two modules of status detection and fault recognition. The approach of sharing one backbone for both modules simplifies the optimization process. The method combines the translation invariance of the convolution kernel and the mask attention mechanism of the transformer by computing the local self-attention and superimposing the partial local attention by the mask to ensure the integrity of the information. Finally, a zero-shot training method is proposed to embed the query into the model to achieve cross-distribution fault diagnosis of bearings. The experiments on the data sets of Case Western Reserve University and machinery fault simulator are implemented to diagnose the bearings. The results show that the proposed method can obtain higher diagnostic accuracy and computational efficiency than the existing methods and can be valid for scenarios with cross-condition diagnosis or imbalanced samples.

Transfer learning can realize the cross-domain fault diagnosis of rotating machinery, where the model trained on plenty of labeled samples collected in one working condition can be transferred to insufficient samples collected in target working condition. Currently, the data features cannot be completely extracted by existing methods when the data distribution gap of the samples collected in different working conditions is quite large. In order to fully extract the data features of rotating machinery to achieve cross-domain fault diagnosis, this paper investigated a cross-domain fault diagnosis model of rotating machinery based on graph feature extraction. The proposed method can realize unsupervised fault diagnosis on rotating machinery running under different working conditions by extracting the numerical and structural features of source and target domains. First of all, data features with large data distribution gaps need to be fully extracted, so a convolutional network based on deformable convolutional network (De-conv) is designed to extract the features with large differences in data distribution under various working conditions. Secondly, features are extracted based on convolutional neural network for data values in existing domain adaptation methods while the structure features of machine monitoring data are ignored. Therefore, a composite spectral-based graph convolutional network (CS-GCN) is designed to extract structural features of data. Thirdly, fully extracted features are input into universal domain adaptation network to achieve cross-domain fault diagnosis with private faults of rotating machinery under changing working conditions. Finally, a benchmarking dataset and a dataset collected from a practical experimental platform are used to verify the effectiveness of the proposed model, and the results show that it is more suitable for cross-domain fault diagnosis of rotating machinery than other comparison models. Keywords: Graph feature extraction; Cross-domain fault diagnosis; Deformable convolutional network; Domain adaptation; Rotating machinery.