| Rolling bearing is an indispensable part of mechanical equipment,its normal operation and health status has a significant impact on the performance,life and reliability of the machine,so it is of great significance to carry out intelligent fault diagnosis and condition monitoring.Most rolling bearing signals are one-dimensional signals,so data preprocessing will increase the steps of fault diagnosis,and feature extraction of original data using a single structure will lead to incomplete feature extraction.Most intelligent diagnosis methods are based on the same distribution of training data and test data.However,when the machine is working in practice,the data distribution changes due to the change of working conditions,which limits the generalization ability of fault diagnosis network model.The introduction of multi-scale method,domain adaptive method and domain generalization method in transfer learning provides a new way to solve bearing fault diagnosis problems with inadequate features and different data distribution.The main research content of this paper is as follows:(1)Aiming at the problem of how to comprehensively extract the timing characteristics of bearing vibration signals and realize end-to-end fault diagnosis,a lightweight 1D MSCNN network model was studied,which takes the original data as input without the prior knowledge of experts and can extract bearing fault characteristics more comprehensively.The multi-scale method can integrate rich features extracted from convolution kerns of different sizes,and the CBR network layer can reduce training parameters to accelerate model convergence.1D MSCNN can comprehensively extract the depth features of rolling bearing faults,and finally complete the classification by Softmax classifier.The experimental results show that the model is an effective intelligent fault diagnosis network for rolling bearings.(2)Aiming at the problem of no label in the target domain under different working conditions,a multi-representation dynamic adaptive(MRDA)fault diagnosis method for rolling bearings was studied under the theoretical framework of domain adaptive theory.Firstly,MMD and WCMMD were used to measure the marginal and conditional distribution distances of source and target domains.The weighting factor?calculated by A-distance was used to evaluate the proportion of marginal and conditional distribution distances,thus forming the dynamic distribution distance(DDD).Then,a multi-representation adaptive structure(MRDAM)is proposed to replace the fully connected layer of 1D MSCNN,forming a 1D hollow multi-representation dynamic adaptive migration network(1D DMRDATN).MRDAM uses multiple substructures to extract different representation distributions,and uses?to align each representation distribution dynamically.It helps to learn more invariant features;Finally,Adam optimizer was used to optimize the sum of distance loss composed of multiple DDDS and the target loss composed of label loss.High precision identification of rolling bearing faults under different working conditions was achieved.The experimental results show that 1D DMRDATN not only has high fault classification accuracy in different working conditions,but also has strong generalization ability.(3)In order to solve the problem that the target data cannot be obtained under unknown working conditions,a one-dimensional cavity multi-representation domain generalized migration network(1D DMRDGTN)is studied under the theoretical framework of domain generalization.Firstly,the network uses the knowledge distillation framework to learn the invariant features of Fourier phase transform in the domain and capture the intrinsic semantic information of the data.MRDA algorithm was used to obtain the interdomain invariant features,and the common feature knowledge was obtained.Then,the distance function is used to maximize the regularization of intra-domain invariant features and inter-domain invariant features to form regular loss,reduce the possible repetition and redundancy between intra-domain invariant features and inter-domain invariant features,and obtain more invariant features.Finally,the model is optimized by the loss function composed of classification,invariant feature learning within the domain,invariant feature learning between the domains and regular loss,and the classification of unknown fault samples is realized.The experimental results show that 1D DMRDGTN has the highest fault identification accuracy,which proves the effectiveness of the proposed method.The proposed method has been verified on the Case Western Reserve bearing data set and the Paderborn University bearing data set,and has been compared with the traditional mainstream methods.Experimental results show that this method has the highest prediction accuracy and excellent overall performance. |
