Research On Intelligent Cross-domain Fault Diagnosis Method Based On Deep Learning

To solve the problem of difficult fault pattern recognition for key rotating mechanical components under cross working conditions,eliminate the dependence on expert knowledge in fault feature extraction,automatically extract domain invariant and class separable features of samples in different working conditions,and improve the effectiveness of cross domain fault intelligent diagnosis,the paper conducts research on cross-domain multi-level fault diagnosis method,adaptive clustering cross-domain intelligent fault diagnosis method and weighted domain adaptation open set intelligent diagnosis model ground on deep learning algorithms.The main research work is as follows:(1)A cross-domain multi-level fault diagnosis model is constructed.To reduce the distribution discrepancy of samples in the feature space and improve the diagnostic performance of the method under a cross-domain scenario,a multi-level classification model based on CNN is constructed.The fault features in the sample are automatically extracted using the model,and the prediction results are output through the classifier,multiple kernel maximum mean discrepancy is introduced to map samples under different working conditions into the regenerative Hilbert space,evaluate the distance between sample distributions,optimize the composition of the model loss function,realize the gradual corresponding alignment of different types of samples in the feature space,and achieve the purpose of cross-working fault diagnosis.In addition,classifiers with different classification tasks are embedded in different positions of the model to realize multi-level fault diagnosis by identifying the fault mode and identifying the fault degree.The effect of the model is verified in multiple sets of experiments,and the results showed that it can accurately ascertain the fault location and size of bearings under cross operating conditions.(2)A cross-domain fault diagnosis method ground on adaptation clustering is developed.Aiming at the conundrum that the domain adaptation algorithm cannot make full use of the domain label and the domain adaptation model misjudges the samples on the decision boundary,an adaptive clustering domain adaptation model is constructed.The feature extractor projects the sample into a high-dimensional space;The classifier plays the role of correctly classifying tagged samples;The domain discriminator is responsible for distinguishing the source of samples.Through the GRL,various modules form confrontation,so that the extracted features can not be correctly recognized by the domain discriminator while being correctly classified by the classifier,to achieve the goal that the feature extractor can extract the domain invariant and class separable features of the samples under different working conditions,so that the samples can achieve domain alignment in the feature space.In addition,an adaptation clustering module is embedded in the model.By calculating the distance of features in the feature space,the target domain samples are pseudo-labelled,so that the unlabeled samples in the target domain continue to close to the direction of the corresponding source domain class center,aloof from the inter-class decision boundary,which contributes to decrease the domain offset.The results of validation in several datasets show that this method has a good effect in cross-domain fault diagnosis.(3)An open set fault diagnosis method based on weighted domain adaptation is proposed.To address the negative transfer problem caused by the model’s inability to recognize unknown categories due to different label spaces under cross working conditions,a weighted domain adaptation method based on double classifiers is proposed,which extracts high-dimensional features of samples.By comparing the probability distribution of the output sample of the classifier with the set threshold,the loss is calculated,which makes the model have the function of judging whether the sample is new;The weighting module is designed to evaluate the divergence between the samples and each class in the two domains.By constructing a weighted loss function,the weight and corresponding loss update the model,reduce the negative impact of fixed threshold,and promote the positive transfer between shared class in the two domains.In addition,the generalization of the model is improved by maximizing the difference between different classifiers.The results show that the model has good diagnostic performance in the cross-condition open set problem.