CN116304820B - Bearing fault type prediction method and system based on multi-source domain transfer learning - Google Patents

Bearing fault type prediction method and system based on multi-source domain transfer learning Download PDF

Info

Publication number
CN116304820B
CN116304820B CN202310251823.0A CN202310251823A CN116304820B CN 116304820 B CN116304820 B CN 116304820B CN 202310251823 A CN202310251823 A CN 202310251823A CN 116304820 B CN116304820 B CN 116304820B
Authority
CN
China
Prior art keywords
domain
comprehensive
source
source domain
fault classification
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202310251823.0A
Other languages
Chinese (zh)
Other versions
CN116304820A (en
Inventor
吴松松
刘毅
郑诗源
黄木盛
荆晓远
张清华
陈俊均
姚永芳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan Changfei Intelligent Network Technology Co ltd
Guangdong University of Petrochemical Technology
Original Assignee
Wuhan Changfei Intelligent Network Technology Co ltd
Guangdong University of Petrochemical Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan Changfei Intelligent Network Technology Co ltd, Guangdong University of Petrochemical Technology filed Critical Wuhan Changfei Intelligent Network Technology Co ltd
Priority to CN202310251823.0A priority Critical patent/CN116304820B/en
Publication of CN116304820A publication Critical patent/CN116304820A/en
Application granted granted Critical
Publication of CN116304820B publication Critical patent/CN116304820B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/04—Bearings
    • G01M13/045—Acoustic or vibration analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Acoustics & Sound (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention belongs to the technical field of vibration data identification in deep learning, and discloses a bearing fault type prediction method and a system based on multi-source domain transfer learning, wherein N source domain samples and prediction labels of comprehensive source domain samples are output by using a fault classification model and a fault classification comprehensive model; the countermeasure learning network learns the domain invariant features of the comprehensive source domain and the N source domains by using the fault classification comprehensive model through countermeasure training; outputting a prediction label of the target domain sample by using the fault classification comprehensive model; the countermeasure learning network learns the domain invariant features of the comprehensive source domain and the target domain through countermeasure training, and outputs the prediction labels of the target domain samples through the fault classification comprehensive model. According to the method, the label of the target domain vibration signal sample can be directly obtained through iteration, and the influence of domain migration problem caused by domain differences existing among different source domains in the multi-source domain is effectively relieved.

Description

Bearing fault type prediction method and system based on multi-source domain transfer learning
Technical Field
The invention belongs to the technical field of equipment fault diagnosis of industrial automation, and particularly relates to a bearing fault type prediction method and system based on multi-source domain transfer learning, which are a guaranteed technical means of normal operation of industrial mobile equipment.
Background
The rolling bearing is a basic key component of rotary mechanical equipment, the running state of the rolling bearing has great influence on the performance of the whole equipment, and the bearing fault diagnosis is a guaranteed supporting technology for the normal running of industrial dynamic equipment. The rolling bearing belongs to a consumable part, and the bearing monitoring data and the bearing state show complex corresponding relation along with the time and the dynamic change of the working condition environment. Therefore, it is a challenging real requirement to accurately determine whether a bearing fails and identify the type of failure based on a sensing signal, such as a vibration signal.
Vibration signals, which contain running state information of rolling bearings, are widely used for diagnosing the health state of the bearings. Bearing fault diagnosis based on machine learning extracts discrimination features reflecting bearing faults from vibration signal data in a data-driven manner, and a fault type label of the bearing is predicted through a classifier. Bearing fault diagnosis can be realized by a Support Vector Machine (SVM), a K Nearest Neighbor (KNN), a Logistic Regression (LR) and other shallow learning algorithms, and can also be realized by means of deep learning algorithms such as a deep confidence network, a convolution network, a recursive network and the like. The diagnosis thought based on machine learning adopts a supervision chemistry model, a fault diagnosis model is trained by using vibration signal data with fault type labels, and then vibration signals of a bearing to be diagnosed are sent into the trained model to predict the health condition of the bearing.
The existing bearing fault diagnosis methods mostly assume that the environment of model deployment is the same as the environment of model training, i.e. the vibration signals in the training data set and the test data set have the same data distribution. However, the factors of cross-machine-set, variable working condition, sensor position difference and the like in practical application cause different distribution characteristics of training data in a source domain and test data in a target domain. This difference in data fields results in a significant reduction in the ability of the diagnostic model to predict failure after deployment. Therefore, the cross-domain fault diagnosis technology has practical significance for solving the problem of bearing health condition monitoring in the actual production environment.
At present, cross-domain bearing fault diagnosis is mainly studied around the situation of a single source domain and a single target domain, and the basic principle is that bearing fault information is mined from the single source domain and effectively migrated to the target domain so as to ensure the diagnosis performance of a diagnosis model in the target domain. In practice, however, the bearing vibration signals used for model training are typically from different environments, i.e., the training data is a collection of multiple data sources, and in such a multi-source domain scenario, performing cross-domain bearing fault diagnostics presents a greater challenge. On one hand, the data distribution difference degree of each source domain is different from that of the target domain, so that the migration contribution degree of the training data of each source domain to the fault semantic information is inconsistent. On the other hand, there is also a difference in data distribution among a plurality of data sources, resulting in difficulty in extracting common semantic information in source domain data. Practice proves that the two challenges cannot be solved by simply expanding the existing Shan Yuanyu cross-domain diagnostic method. Starting from the two aspects of a model structure and a learning strategy, comprehensively considering the distribution difference in a multi-source domain and the distribution difference between a source domain and a target domain, the end-to-end fault diagnosis method based on the deep learning network design is expected to provide an effective solution.
Through the above analysis, the problems and defects existing in the prior art are as follows:
(1) Most of the existing rolling bearing fault diagnosis methods assume that vibration signal data in a training set and vibration signal data in a test set come from the same distribution, but the assumption is difficult to be established in a practical application scene, and the distribution difference of training and test samples can obviously reduce the diagnosis accuracy of a traditional diagnosis model.
(2) In practical application, there is a problem of cross-domain diagnosis of migration from multiple source domains to target domains, and the difficulty is that the distribution of mixed multi-source domain data and the distribution difference from each source domain to target domain are distributed, while the existing cross-domain diagnosis method is less involved in the situation, and it is difficult to solve the problem by simply combining or expanding the existing domain cross-domain model.
Disclosure of Invention
Aiming at the problems existing in the prior art, the invention provides a bearing fault type prediction method and a system based on multi-source domain transfer learning.
The invention is realized in such a way that the bearing fault type prediction method based on the multi-source domain transfer learning comprises the following steps: outputting the prediction labels of N source domain samples and the comprehensive source domain samples by using the fault classification model and the fault classification comprehensive model; the countermeasure learning network learns the domain invariant features of the comprehensive source domain and the N source domains by using the fault classification comprehensive model through countermeasure training; outputting a prediction label of the target domain sample by using the fault classification comprehensive model; the countermeasure learning network learns the domain invariant features of the comprehensive source domain and the target domain through countermeasure training, and outputs the prediction labels of the target domain samples through the fault classification comprehensive model.
Further, the bearing fault type prediction method based on multi-source domain transfer learning comprises the following steps:
step one, acquiring a vibration signal data set; synthesizing the N source domains into a comprehensive source domain, wherein the comprehensive source domain comprises all samples and sample labels of the N source domains; respectively inputting samples and sample labels of N source domains into N ResNet50 deep convolutional neural networks to perform supervised learning and training to obtain feature extractors and classifiers of N fault classification models, and outputting prediction labels of the samples of the N source domains through the classifiers of the N fault classification models; inputting the comprehensive source domain samples and the sample labels into a ResNet50 deep convolutional neural network to perform supervised learning and training to obtain a feature extractor and a classifier of a fault classification comprehensive model, and outputting a prediction label of the comprehensive source domain samples through the classifier of the fault classification comprehensive model;
step two, respectively inputting the prediction labels of the comprehensive source domain samples in the step one and the prediction labels of the N source domain samples in a pairwise countermeasure learning network for countermeasure training, reducing the field difference between the comprehensive source domain and the N source domains, and enabling the fault classification comprehensive model to learn the domain invariant features between the comprehensive source domain and the N source domains;
Step three, inputting the target domain sample into the fault classification comprehensive model in the step two for unsupervised learning training, and outputting a prediction label of the target domain sample through a classifier of the fault classification comprehensive model;
and step four, inputting the prediction label of the target domain sample in the step three and the prediction label of the comprehensive source domain into the countermeasure learning network in the step two to perform countermeasure training, reducing the domain difference between the comprehensive source domain and the target domain, enabling the fault classification comprehensive model to learn the domain invariant features between the comprehensive source domain and the target domain, and finally outputting the prediction label of the target domain sample by the target domain sample through a classifier of the fault classification comprehensive model.
Further, in the first step, the vibration signal data set includes N source domains and one target domain, and the fault type sets in the source domain and the target domain are the same; the data and the tag in the source domain are known, while the only data in the target domain is known, without any tag information; the source domain sample consists of several vibration signals of known type and the target domain sample consists of several vibration signals of unknown type.
The source domain and the target domain have a common fault type, and different distributions are included between the source domain samples and the target domain samples. The source domain dataset is noted as Representing the ith vibration signal in the jth source domain sample, +.>Is->Fault type label corresponding to vibration signal, N s j The j-th source domain sample number. Synthesizing N source domains into a comprehensive source domain data set, which is marked as +.>Wherein (1)>Representing the i-th vibration signal in the integrated source domain sample,>is->Fault type label corresponding to vibration signal +.>To synthesize the number of source domain samples. The target domain dataset is noted asWherein (1)>Represents the i-th vibration signal in the target domain sample,/->For the target domain fault type label to be predicted, N t For target domain samplesA number. For the jth source domain, feature extractor G is used, which corresponds to the fault classification model j Extracting sample characteristics->Will be->Classifier F fed into corresponding fault classification model j Sample prediction tag for obtaining jth source domain +.>And through calculation of the cross entropy loss function, the accuracy of the prediction label of the corresponding source domain output by the fault classification model classifier is improved. Obtaining L by averaging and calculating cross entropy loss function results trained by N fault classification models corresponding to source domains C The calculation method is as follows:
for the integrated source domain, its corresponding fault classification integrated model feature extractor G N+1 Extracting sample featuresSample characteristics->Input corresponding classifier F N+1 Obtaining a sample prediction label of the comprehensive source domainAnd performing supervised learning training on the comprehensive source domain samples and the sample labels by using a cross entropy loss function, and improving the accuracy of the comprehensive source domain prediction labels output by the fault classification comprehensive model classifier through calculation of the cross entropy loss function. The cross entropy loss function calculation result of the fault classification comprehensive model on the comprehensive source domain training is obtained +.>The calculation method is as follows:
in order to model the prediction label of the domain invariant feature, a feature extractor is provided, wherein G1 represents a fault classification model and a fault classification comprehensive extraction domain feature, G2 represents a fault classification model and a fault classification comprehensive extraction domain difference feature, F1 represents a fault classification model and a fault classification comprehensive domain feature classification classifier, F2 represents a fault classification model and a fault classification comprehensive domain difference feature classification classifier, and the domain feature extracted by the feature extractor G1 is input into the domain feature prediction label obtained by classifying the domain feature by the classifier F1The domain difference features extracted by the feature extractor G2 are input into a classifier F2 to be classified to obtain a prediction tag of the domain difference features +.>The two expressions are described aboveAnd->Predictive labels for domain invariant features of N source domain samples obtained by subtraction Prediction tag of domain invariant feature of integrated source domain sample can be obtained in the same way>The calculation method is as follows:
extracting the difference features of the constraint domains, extracting N source domain samples from the fault classification model, extracting the constraint of the domain difference features of the comprehensive source domain samples from the fault classification comprehensive model, and using a loss function L G A computing implementation, wherein K represents the number of individual source domain vibration signal fault types, in the following manner:
extracting constraint of domain difference characteristics of comprehensive source domain and target domain for fault classification comprehensive model and utilizing loss functionThe calculation is realized in the following way:
in the second step, the predictive labels of the comprehensive source domain samples obtained in the first step and the predictive labels of the N source domain samples are respectively input into the same countermeasure learning network for training, wherein the countermeasure learning network is realized by introducing a discriminator D h So that the feature extractor is opposed to the discriminator D h Predictive labels responsible for outputting N source domain samples to fault classification modelPredictive tag for outputting comprehensive source domain samples with fault classification comprehensive model>Proceeding zoneDividing; feature extractor G using fault classification synthesis model N+1 The extraction of domain invariant features of the comprehensive source domain sample and N source domain samples is realized, so that the fault classification comprehensive model outputs a prediction label of the comprehensive source domain sample +. >Deception discriminator D h . The predictive labels of the comprehensive source domain samples output by the fault classification comprehensive model and the predictive labels of the N source domain samples output by the fault classification model are respectively used as a discriminator D h To perform a contrast loss function +.>The calculation of (2) is as follows:
for N countermeasures against loss functionAveraged to L adv The calculation method is as follows:
in order to model predictive labels for identifying domain invariant features, let D1 represent a discriminator for identifying domain features in an countermeasure learning network, D1 represent a discriminator for identifying domain difference features in the countermeasure learning network, and apply predictive labels for domain featuresInputting discriminator D1 to obtain the result of discriminating domain feature prediction tag +.>Predictive tag for characterizing domain differences->Inputting the result of the discriminator D2 to obtain the discrimination domain difference characteristic predictive labelPredictive tag identifying domain invariant features of N source domain samples +.>Two expressions>And->The subtraction is calculated, and the predictive label for identifying the domain invariant feature of the comprehensive source domain sample can be obtained in the same way>The calculation method is as follows:
using a loss function L D The influence of the constraint domain difference feature on the discriminator is calculated, the discrimination capability of the discriminator on the domain invariant feature between N source domains and the comprehensive source domain is enhanced, and the calculation mode is as follows:
Further, in step three, for the target domain sample, the fault classification comprehensive model feature extractor G N+1 Extracting target domain sample characteristics, and extracting the target domain sample characteristics asFault classification comprehensive model classifier F N+1 Through input of fault classification comprehensive model classifierAnd outputting the prediction label of the target domain sample.
In the fourth step, the prediction label of the target domain sample in the third step and the prediction label of the comprehensive source domain sample are input into the countermeasure learning network in the second step to train, and the discriminator D h Predictive labels responsible for comprehensive source domain samplesAnd predictive tag of target domain sample +.>Distinguishing; the extraction of the domain invariant features between the comprehensive source domain sample and the target domain sample is realized by using the fault classification comprehensive model, so that the fault classification comprehensive model outputs a prediction label of the comprehensive source domain sample>Spoofing the discriminator; the predictive label of the comprehensive source domain sample output by the comprehensive fault classification model and the predictive label of the target domain sample output by the comprehensive fault classification model are used as a discriminator D h Is to perform the fight loss function>The calculation method is as follows:
using a loss function L Dt Computing the influence of the implementation constraint domain difference features on the discriminator, enhancing the discriminator D h For the discrimination capability of the domain invariant feature between the comprehensive source domain sample and the target domain sample, the calculation mode is as follows:
through continuous iterative training, the fault classification comprehensive model learns to extract domain invariant features between comprehensive source domain samples and target domain samples, L all The integrated loss function defines the total loss function as follows:
by the above formula (14). And updating the network weight parameters by the total loss function through continuous iterative training of an SGD random gradient descent method, so that the unlabeled type target domain samples output predicted sample labels through the fault classification comprehensive model.
Another object of the present invention is to provide a bearing fault type prediction system based on multi-source domain transfer learning, which applies the bearing fault type prediction method based on multi-source domain transfer learning, the bearing fault type prediction system based on multi-source domain transfer learning includes:
the sample prediction label output module is used for outputting the prediction labels of the N source domain samples and the comprehensive source domain samples by utilizing the fault classification model and the fault classification comprehensive model;
the countermeasure training module is used for enabling the fault classification comprehensive model to learn domain invariant features of the comprehensive source domain and the N source domains through countermeasure training by using a countermeasure learning network;
The fault prediction module is used for outputting a prediction label of the target domain sample by using the fault classification comprehensive model, and the fault classification comprehensive model learns the domain invariant features of the comprehensive source domain and the target domain by countermeasure training.
Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program when executed by the processor causes the processor to execute the steps of the bearing fault type prediction method based on multi-source domain transfer learning.
Another object of the present invention is to provide a computer readable storage medium storing a computer program, which when executed by a processor, causes the processor to perform the steps of the bearing failure type prediction method based on multi-source domain transfer learning.
The invention further aims to provide an information data processing terminal which is used for realizing the bearing fault type prediction system based on multi-source domain transfer learning.
In combination with the technical scheme and the technical problems to be solved, the technical scheme to be protected has the following advantages and positive effects:
First, aiming at the technical problems in the prior art and the difficulty of solving the problems, the technical problems solved by the technical proposal of the invention are analyzed in detail and deeply by tightly combining the technical proposal to be protected, the results and data in the research and development process, and the like, and some technical effects brought after the problems are solved have creative technical effects. The specific description is as follows:
the technical scheme of the invention has the advantages that:
1. under the same experimental data conditions, a higher recognition rate can be obtained through the ResNet50 deep convolutional neural network.
2. The identification method for constructing domain difference features in the source domain and the target domain respectively improves the reliability of the ResNet50 deep convolutional neural network on domain invariant feature extraction.
3. Modeling the differential feature module of the identification domain is helpful to better describe the distance between the hyperplane of the domain and the ideal decision boundary of the anti-learning network, thereby facilitating the domain invariant feature extraction between the source domain and the target domain.
The biggest bright point of the invention is that the ResNet50 deep convolutional neural network and domain difference feature modeling are applied to extract the features of the source domain and the target domain, firstly, the domain invariant features between the comprehensive source domain and N source domains are learned by using the antagonism learning network, and after the difference between the source domains is eliminated, the domain invariant features between the target domain and the comprehensive source domain are extracted, so that the multi-source domain knowledge is migrated to the target domain. According to the method and the device, the label of the target domain vibration signal sample can be directly obtained through iteration, so that the influence of domain migration problem caused by domain differences existing among different source domains in the multi-source domain is effectively relieved.
Secondly, the technical scheme is regarded as a whole or from the perspective of products, and the technical scheme to be protected has the following technical effects and advantages:
the novel bearing fault type prediction method based on multi-source domain transfer learning based on multi-source domain countermeasure learning, provided by the invention, applies domain invariant features of N source domains learned by a countermeasure learning network, and then utilizes domain invariant features of N source domains and target domains learned by the countermeasure learning network to realize multi-source domain knowledge transfer to the target domain. According to the method, the more and more accurate target domain sample labels are obtained through continuous iterative training by utilizing the fault classification model and the fault classification comprehensive model, so that the influence caused by domain differences among different source domains in the multi-source domain migration problem is effectively relieved.
Thirdly, as inventive supplementary evidence of the claims of the present invention, the following important aspects are also presented:
(1) The expected benefits and commercial values after the technical scheme of the invention is converted are as follows:
the invention can be used for technical transformation of enterprises comprising industrial mobile equipment. Particularly, in the production process of a large number of bearing components in enterprises such as petrochemical industry, manufacturing industry and the like, the technology is expected to remarkably improve the monitoring performance of the bearing health condition, and timely and effectively early warning is carried out in face of complex environments, so that production accidents and production pauses caused by bearing faults are avoided, potential production risks of the enterprises are reduced, and the production quality improvement and efficiency improvement of the enterprises are promoted.
(2) The technical scheme of the invention fills the technical blank in the domestic and foreign industries:
for the domain migration problem in the bearing fault diagnosis field, the current technical scheme mainly focuses on a Shan Yuanyu migration scene, but for a multi-source domain migration scene which is more practical, attention is not paid. Aiming at the actual situation that the performance of the bearing fault diagnosis technology in the complex production environment is reduced in the presence of the multi-source domain migration situation, the invention provides the practical problem that the multi-source domain migration is unavoidable in bearing fault diagnosis because training data in the actual production environment are usually from different data sources. The invention provides an effective technical scheme aiming at bearing fault diagnosis under the condition of multi-source domain migration by adopting advanced tools such as an anti-learning network, migration learning and the like around the overall target of migrating fault semantic knowledge of the multi-source domain to the target domain, and is beneficial to improving the effect of bearing health condition monitoring under the actual complex production environment.
(3) The technical scheme of the invention solves the technical problems that people are always desirous of solving but are not successful all the time:
the domain invariant features of N source domains are learned by the countermeasure learning network, and then the domain invariant features of N source domains and the target domain are learned by the countermeasure learning network, so that the multi-source domain knowledge is transferred to the target domain. The invention is directed to a bearing fault type prediction method based on multi-source domain migration learning of multi-source domain countermeasure learning with unknown class target domain data, which is also required to be continuously migrated to a target domain in a countermeasure learning network, a fault classification model and a fault classification comprehensive model are constructed and trained to accurately predict source domain and target domain samples, and the countermeasure learning network can adjust domain differences among the multi-source domains to cause negative migration in the domain adaptation process. The invention further aims at the problem that the data distribution of a plurality of source domains is inconsistent with that of a plurality of target domains, adopts the countermeasure learning network integrated based on the thought of generating countermeasure learning, realizes the effective migration of knowledge, and improves the fault diagnosis precision and the robustness of the fault classification comprehensive model.
The invention also respectively models the fault classification model and the fault classification comprehensive model as well as the classifier and the discriminator used in the antagonism learning network, and can lead the parameters of model training to be more accurate through the modeling of the domain difference characteristics, thereby effectively improving the prediction result of the fault classification comprehensive model on the target domain sample type.
(4) The technical scheme of the invention overcomes the technical bias:
research on cross-domain bearing fault diagnosis technology has been focused on single-source domain migration for a long time, and concerns on multi-source domain migration in a practical environment are insufficient. Meanwhile, the existing multi-source domain migration technical scheme only stays on the domain migration problem between the source domain and the target domain, and ignores the multi-source domain knowledge migration failure problem caused by inconsistent data distribution between the source domain. The invention comprehensively considers the distribution difference between the multi-source domains and between the source domain and the target domain in the model design process, reduces the difference between different domains by constructing the antagonism learning network and the loss function, and opens up a new visual angle for multi-source domain migration.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments of the present invention will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flowchart of a bearing fault type prediction method based on multi-source domain transfer learning provided by an embodiment of the invention;
fig. 2 is a schematic diagram of a bearing fault type prediction method based on multi-source domain transfer learning according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the following examples in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Aiming at the problems existing in the prior art, the invention provides a bearing fault type prediction method and a system based on multi-source domain transfer learning, and the invention is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, the bearing fault type prediction method based on multi-source domain transfer learning provided by the embodiment of the invention includes the following steps:
s101, outputting N source domain samples and predictive labels of the comprehensive source domain samples by using a fault classification model and a fault classification comprehensive model;
s102, enabling the fault classification comprehensive model to learn domain invariant features of a comprehensive source domain and N source domains through countermeasure training by a countermeasure learning network;
S103, outputting a prediction label of the target domain sample by using the fault classification comprehensive model;
s104, the countermeasure learning network learns the domain invariant features of the comprehensive source domain and the target domain through countermeasure training by the fault classification comprehensive model, and outputs the prediction labels of the target domain samples through the fault classification comprehensive model.
As a preferred embodiment, as shown in fig. 2, the bearing fault type prediction method based on multi-source domain transfer learning provided by the embodiment of the present invention includes the following steps:
s1: a known vibration signal dataset is utilized, the vibration signal dataset comprising N source domains and 1 target domain, and the set of fault types in the source domains and the target domains being the same. Both the data and the tag in the source domain are known, while only the data in the target domain is known, without any tag information. The N source domains are synthesized into 1 comprehensive source domain, and the comprehensive source domain contains all samples of the N source domains and sample labels. The N source domain samples and the sample labels are respectively input into N ResNet50 deep convolutional neural networks to perform supervised learning to train out feature extractors and classifiers of N fault classification models, and the N source domain samples can be output through the classifiers of the N fault classification models. And the comprehensive source domain sample and the sample label are input into a ResNet50 deep convolutional neural network to perform supervised learning to train a feature extractor and a classifier of a fault classification comprehensive model, and the classifier of the fault classification comprehensive model can output a prediction label of the comprehensive source domain sample.
S2: and (2) inputting the prediction labels of the comprehensive source domain samples in the step (S1) and the prediction labels of the N source domain samples into a countermeasure learning network in a pairwise manner to perform countermeasure training, so that the domain difference between the comprehensive source domain and the N source domains is reduced, and the domain invariant features between the comprehensive source domain and the N source domains are learned by the fault classification comprehensive model.
S3: and (2) inputting the target domain sample into the fault classification comprehensive model in the step (S2) for performing unsupervised learning training, and outputting a prediction label of the target domain sample through a classifier of the fault classification comprehensive model.
S4: inputting the prediction label of the target domain sample in the step S3 and the prediction label of the comprehensive source domain into the countermeasure learning network in the step S2 for countermeasure training, reducing the domain difference between the comprehensive source domain and the target domain, enabling the fault classification comprehensive model to learn the domain invariant features between the comprehensive source domain and the target domain, and finally enabling the target domain sample to output the accurate prediction label of the target domain sample through the classifier of the fault classification comprehensive model.
In step S1 provided in the embodiment of the present invention, the source domain sample is composed of a plurality of vibration signals of known types, and the target domain sample is composed of a plurality of vibration signals of unknown types.
In step S1 provided by the embodiment of the present invention, there are N source domains and one target domain, where the source domain and the target domain have a common fault type, and there are different distributions between the source domain samples and the target domain samples. The source domain dataset is noted asWherein->Representing the ith vibration signal in the jth source domain sample, +.>Is->Fault type label corresponding to vibration signal, N s j For the jth source domain sample number, the integrated source domain dataset synthesized by N source domains is recorded as +.>Wherein->Wherein->Representing the i-th vibration signal in the integrated source domain sample,>is->Fault type label corresponding to vibration signal +.>To synthesize the number of source domain samples. The target domain dataset is marked +.>Wherein->Representing the ith vibration signal in the target domain sample, where N t For the number of target domain samples, the fault type label of the target domainPredicting; for the jth source domain, feature extractor G is used, which corresponds to the fault classification model j Extracting sample characteristics->Will be->Classifier F fed into corresponding fault classification model j Sample prediction label for obtaining jth source domainBy calculating the cross entropy loss function, the fault classification model is improvedThe classifier outputs the accuracy of the predictive label corresponding to the source domain. Obtaining L by averaging and calculating cross entropy loss function results trained by N fault classification models corresponding to source domains C The calculation method is as follows:
for the integrated source domain, its corresponding fault classification integrated model feature extractor G N+1 Extracting sample featuresSample characteristics->Input corresponding classifier F N+1 Obtaining a sample prediction label of the comprehensive source domainAnd performing supervised learning training on the comprehensive source domain samples and the sample labels by using a cross entropy loss function, and improving the accuracy of the comprehensive source domain prediction labels output by the fault classification comprehensive model classifier through calculation of the cross entropy loss function. The cross entropy loss function calculation result of the fault classification comprehensive model on the comprehensive source domain training is obtained +.>The calculation method is as follows:
in order to model the prediction labels of the domain invariant features, a feature extractor is provided, wherein G1 represents a fault classification model and a fault classification comprehensive extraction domain feature, G2 represents a fault classification model and a fault classification comprehensive extraction domain difference feature, F1 represents a fault classification model and a fault classification comprehensive classification domain feature classification classifier, F2 represents a fault classification model and a fault classification comprehensive extraction domain feature classifierThe obstacle classification synthesizes the classifier for classifying the domain difference features, and inputs the domain features extracted by the feature extractor G1 into the classifier F1 to classify the obtained prediction labels of the domain features The domain difference features extracted by the feature extractor G2 are input into a classifier F2 to be classified to obtain a prediction tag of the domain difference features +.>The two expressions are described aboveAnd->Predictive labels for domain invariant features of N source domain samples obtained by subtractionPrediction tag of domain invariant feature of integrated source domain sample can be obtained in the same way>The calculation method is as follows:
in order to make the fault classification model and the fault classification comprehensive model extract domain difference features as little as possible, so that the extraction of domain difference features is further restricted, the constraint of extracting the domain difference features of the N source domain samples from the fault classification model and the domain difference features of the comprehensive source domain samples from the fault classification comprehensive model can be used as the loss function L G Computational implementation, where K represents a source domain vibration signal fault classThe number of patterns is calculated as follows:
for the constraint that the fault classification comprehensive model extracts domain difference characteristics of a comprehensive source domain and a target domain, a loss function can be usedThe calculation is realized in the following way:
the constraint of the formula (5) improves the extraction capacity of the fault classification model and the fault classification comprehensive model to the domain invariant features.
In step S2 provided by the embodiment of the present invention, the prediction labels of the comprehensive source domain samples obtained in step S1 are respectively input into the same countermeasure learning network with the prediction labels of the N source domain samples to train, and the countermeasure learning network is implemented by introducing a discriminator D h So that the feature extractor is opposed to the discriminator D h Predictive labels responsible for outputting N source domain samples to fault classification modelPredictive label for outputting comprehensive source domain samples by comprehensive fault classification modelFeature extractor G for discriminating and fault classifying comprehensive model N+1 The extraction of the domain invariant features of the comprehensive source domain sample and the N source domain samples is required to be realized as far as possible, so that the fault classification comprehensive model outputs a prediction label of the comprehensive source domain sample +.>Capable of deceiving discriminator D h . The predictive labels of the comprehensive source domain samples output by the fault classification comprehensive model and the predictive labels of the N source domain samples output by the fault classification model are respectively used as a discriminator D h To perform a contrast loss function +.>The calculation of (2) is as follows:
for the N countermeasures loss functionsAveraged to L adv The calculation method is as follows:
in order to model predictive labels for identifying domain invariant features, let D1 represent a discriminator for identifying domain features in an countermeasure learning network, D1 represent a discriminator for identifying domain difference features in the countermeasure learning network, and apply predictive labels for domain featuresInputting discriminator D1 to obtain the result of discriminating domain feature prediction tag +.>Predictive tag for characterizing domain differences- >Inputting the result of the discriminator D2 to obtain the discrimination domain difference characteristic predictive labelPredictive tag identifying domain invariant features of N source domain samples +.>Two expressions>And->The subtraction is calculated, and the predictive label for identifying the domain invariant feature of the comprehensive source domain sample can be obtained in the same way>The calculation method is as follows:
to further constrain the impact of domain difference features on the discriminator, enhancing the discriminator's discrimination capability for domain invariant features between N source domains and the integrated source domain, the loss function L can be used D And (5) computing. The calculation method is as follows:
in step S3 provided by the embodiment of the present invention, for the target domain sample, the fault classification comprehensive model feature extractor G N+1 Extracting target domain sample characteristics, and extracting the target domain sample characteristics as a fault classification comprehensive model classifier F N+1 Through input of fault classification comprehensive model classifierThe predictive label of the target domain sample can be output.
Implementation of the inventionIn example provided step S4, training the predictive label of the target domain sample in step S3 and the predictive label of the integrated source domain sample in the countermeasure learning network in step S2, and the discriminator D h Predictive labels responsible for comprehensive source domain samples And predictive tag of target domain sample +.>The method is characterized in that a distinction is made, and the fault classification comprehensive model extracts domain invariant features between a comprehensive source domain sample and a target domain sample as far as possible, so that the fault classification comprehensive model outputs a prediction label +_ of the comprehensive source domain sample>The authenticator can be spoofed. The predictive label of the comprehensive source domain sample output by the comprehensive fault classification model and the predictive label of the target domain sample output by the comprehensive fault classification model are used as a discriminator D h To perform a contrast loss function +.>The calculation method is as follows:
to further constrain the impact of domain difference features on the discriminator, the discriminator D is enhanced h For the discrimination capability of the domain invariant feature between the integrated source domain sample and the target domain sample, the loss function L can be used Dt The calculation is realized in the following way:
through continuous iterative training, the fault classification comprehensive model learns the domain between the comprehensive source domain sample and the target domain sampleExtracting invariant features, and synthesizing the loss function in the formula to define L all The total loss function is as follows:
by the above formula (14). The total loss function is trained through continuous iteration of an SGD random gradient descent method, and the weight parameters of the network are updated, so that the unlabeled type target domain sample can output a corresponding sample label accurately predicted through a fault classification comprehensive model.
According to the method, the predictive labels of unknown vibration signals in the target domain are obtained mainly through fault classification models and fault classification comprehensive models and countermeasure learning network training learning in the multi-source domain and the target domain. Firstly N source domains and an integrated source domain respectively obtain N source domain samples and predictive labels of the integrated source domain samples through a fault classification model and a fault classification integrated model, the N source domain samples and the predictive labels of the integrated source domain samples are respectively input into an countermeasure learning network to perform countermeasure training learning to obtain domain discrimination probability vectors between the integrated source domain samples and the N source domain samples, then a target domain is input into the countermeasure learning network to perform countermeasure training learning to obtain domain discrimination probability vectors between the integrated source domain samples and the target domain samples through the fault classification integrated model to obtain domain invariant features between multiple source domains and the target domain samples, and the target domain samples can obtain predictive labels of unknown type vibration signals in the target domain through the fault classification integrated model.
Source domain difference problem: while there is a certain difference in data attributes between different source domains, a core problem of multi-source domain knowledge migration to a target domain is how to learn attribute features shared by the multi-source domains through a network, so as to eliminate the difference between the different source domains. For this purpose we have chosen a method of generating a challenge to extract common features that exist between the domains, and when training, the source domain consists of several known types of vibration data and the target domain consists of several unknown types of vibration data. The method comprises the steps of learning a target domain with unknown source domain and sample label by using a fault classification model, a fault classification comprehensive model and an countermeasure learning network, achieving multi-source domain knowledge transfer learning to the target domain, starting N source domain and comprehensive source domain to obtain N source domain samples and comprehensive source domain samples prediction labels through the fault classification model and the fault classification comprehensive model, inputting the N source domain samples label prediction labels and the comprehensive source domain sample prediction labels into the countermeasure learning network to perform countermeasure training learning to obtain domain discrimination probability vectors between the comprehensive source domain samples and the N source domain samples, restricting the domain discrimination probability vectors by a loss function, so that the fault classification comprehensive model is reduced to extract domain difference characteristics between the comprehensive source domain samples and the N source domain samples, and then obtaining the target domain samples prediction labels through the fault classification comprehensive model, inputting the target domain sample prediction labels and the comprehensive source domain samples into the countermeasure learning network to perform countermeasure training to obtain domain discrimination probability vectors between the comprehensive source domain samples and the target domain samples, restricting the domain discrimination probability vectors by the loss function, and reducing the domain discrimination probability vectors between the comprehensive source domain samples and the target domain samples, and obtaining the failure classification characteristic of the target domain sample can not be obtained by the target domain classification comprehensive domain.
The bearing fault type prediction system based on multi-source domain transfer learning provided by the embodiment of the invention comprises the following components:
the sample prediction label output module is used for outputting the prediction labels of the N source domain samples and the comprehensive source domain samples by utilizing the fault classification model and the fault classification comprehensive model;
the countermeasure training module is used for enabling the fault classification comprehensive model to learn domain invariant features of the comprehensive source domain and the N source domains through countermeasure training by using a countermeasure learning network;
the fault prediction module is used for outputting a prediction label of the target domain sample by using the fault classification comprehensive model, and the fault classification comprehensive model learns the domain invariant features of the comprehensive source domain and the target domain by countermeasure training.
In order to prove the inventive and technical value of the technical solution of the present invention, this section is an application example on specific products or related technologies of the claim technical solution.
The invention uses multi-source domain and target domain data for training. During training, the source domain is composed of a plurality of vibration data of known types, and the target domain is composed of a plurality of vibration data of unknown types. And learning the target domain with unknown source domain and sample label by using a fault classification model, a fault classification comprehensive model and an countermeasure learning network to realize multi-source domain knowledge transfer learning to the target domain, obtaining N source domain samples and prediction labels of the comprehensive source domain samples by using the fault classification model and the fault classification comprehensive model respectively, using the obtained source domain and the prediction labels of the comprehensive source domain to enable the countermeasure learning network to learn domain invariant features between the comprehensive source domain and the N source domains first, obtaining a target domain sample prediction label by using the fault classification comprehensive model after the difference between the source domains is eliminated, and then learning and extracting domain invariant features between the target domain and the comprehensive source domain by using the countermeasure learning network to realize multi-source domain knowledge transfer to the target domain.
The vibration data label prediction method is used for carrying out experiments on a CWRU bearing vibration database, and comparing and analyzing experimental results with other unknown vibration data label prediction methods.
A multisource domain migration bearing fault diagnosis experiment is conducted based on a rolling bearing database CWRU of Kassi university in the United states so as to verify the effectiveness of the proposed method in cross-domain bearing fault diagnosis. The bearing data acquisition test board consists of an execution motor, a torque measuring sensor, a torque sensor/encoder, a power machine, a dynamometer and control electronics.
Table 1 bearing parameter table for eight vibration data fields in CWRU bearing database
The test bearing is used for supporting a motor shaft, and three single-point faults, namely an outer ring fault, an inner ring fault and a ball fault, are introduced into the test bearing by adopting electric spark machining. Vibration data is collected using an accelerometer placed at twelve point positions at the drive end and the fan end of the motor housing. The sampling frequency of the driving end is 12KHz, and the adopted frequency of the fan end is 48KHz. Vibration signals are collected under four motor load conditions of 0, 1, 2 and 3 horsepower. As shown in table 1, the vibration data is divided into 8 data fields, respectively A, B, C, D, E, F, G, H, according to the vibration signal sampling points and the motor load, wherein A, B, C, D is derived from the driving end data, and E, F, G, H is derived from the fan end data.
By selecting one of the 8 data fields as a target field and the data fields of the other different ends as source fields, the invention constructs 8 fault diagnosis tasks for multi-source field migration, and specific information of the data fields and the fault diagnosis tasks is shown in table 2. The present invention employs the following data processing method, each vibration data sample is a 4000-dimensional vector from 4 bearing condition types, namely health, outer race failure, inner race failure, and ball failure. The algorithm takes initial data of a source domain and a target domain as input, and obtains a fault state prediction label of the target bearing through a multi-source domain field self-adaptive model. The list of multi-source domain migration bearing fault diagnosis tasks is shown in table 2.
TABLE 2 Multi-Source Domain migration bearing failure diagnosis task List
Under the same experimental setup, 5 widely accepted and representative methods were chosen as references, SSTCA, trAdaboost, SGF, GFK and MSDGIFI, the bearing failure type prediction accuracy is shown in table 3.
Table 3 results of multisource domain migration bearing fault diagnosis experiments at experimental set-up I
The invention applies the principle of the antagonism learning network to learn the domain invariant features between the multi-source domain sample and the target domain sample on the multi-source domain and the target domain. The multi-source domain sample, the sample label and the target domain sample can obtain the label of the unknown type target domain sample through the fault classification model, the fault classification comprehensive model and the anti-learning network iterative training, so that the influence of the source domain difference problem is effectively relieved, and the multi-source domain knowledge is migrated to the target domain. The invention uses the classification recognition method, and the results in the table 3 show that the recognition rate of the method provided by the invention is higher than that of other 5 methods, and experiments prove that the method can effectively relieve the influence caused by the domain difference among different source domains in the multi-source domain migration problem.
It should be noted that the embodiments of the present invention can be realized in hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or special purpose design hardware. Those of ordinary skill in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such as provided on a carrier medium such as a magnetic disk, CD or DVD-ROM, a programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The device of the present invention and its modules may be implemented by hardware circuitry, such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., as well as software executed by various types of processors, or by a combination of the above hardware circuitry and software, such as firmware.
The foregoing is merely illustrative of specific embodiments of the present invention, and the scope of the invention is not limited thereto, but any modifications, equivalents, improvements and alternatives falling within the spirit and principles of the present invention will be apparent to those skilled in the art within the scope of the present invention.

Claims (9)

1. The bearing fault type prediction method based on the multi-source domain transfer learning is characterized by comprising the following steps of:
step one, acquiring a vibration signal data set; synthesizing the N source domains into a comprehensive source domain, wherein the comprehensive source domain comprises all samples and sample labels of the N source domains; respectively inputting samples and sample labels of N source domains into N ResNet50 deep convolutional neural networks to perform supervised learning and training to obtain feature extractors and classifiers of N fault classification models, and outputting prediction labels of the samples of the N source domains through the classifiers of the N fault classification models; inputting the comprehensive source domain samples and the sample labels into a ResNet50 deep convolutional neural network to perform supervised learning and training to obtain a feature extractor and a classifier of a fault classification comprehensive model, and outputting a prediction label of the comprehensive source domain samples through the classifier of the fault classification comprehensive model;
step two, respectively inputting the prediction labels of the comprehensive source domain samples in the step one and the prediction labels of the N source domain samples in a pairwise countermeasure learning network for countermeasure training, reducing the field difference between the comprehensive source domain and the N source domains, and enabling the fault classification comprehensive model to learn the domain invariant features between the comprehensive source domain and the N source domains;
Step three, inputting the target domain sample into the fault classification comprehensive model in the step two for unsupervised learning training, and outputting a prediction label of the target domain sample through a classifier of the fault classification comprehensive model;
and step four, inputting the prediction label of the target domain sample in the step three and the prediction label of the comprehensive source domain into the countermeasure learning network in the step two to perform countermeasure training, reducing the domain difference between the comprehensive source domain and the target domain, enabling the fault classification comprehensive model to learn the domain invariant features between the comprehensive source domain and the target domain, and finally outputting the prediction label of the target domain sample by the target domain sample through a classifier of the fault classification comprehensive model.
2. The method for predicting the type of bearing failure based on multi-source domain transfer learning of claim 1, wherein in the first step, the vibration signal data set includes N source domains and one target domain, and the sets of failure types in the source domain and the target domain are the same; the data and the tag in the source domain are known, while the only data in the target domain is known, without any tag information; the source domain sample consists of a plurality of vibration signals of known types, and the target domain sample consists of a plurality of vibration signals of unknown types;
The source domain and the target domain have a common fault type, and different distributions are included between the source domain sample and the target domain sample; the source domain dataset is noted as Representing the ith vibration signal in the jth source domain sample, +.>Is->Fault type label corresponding to vibration signal, N s j The j-th source domain sample number; synthesizing N source domains into a comprehensive source domain data set, and recording asWherein (1)>Representing the i-th vibration signal in the integrated source domain sample,>is->Fault type label corresponding to vibration signal +.>The number of the source domain samples is synthesized; the target domain dataset is marked +.>Wherein (1)>Represents the i-th vibration signal in the target domain sample,/->For the target domain fault type label to be predicted, N t The number of samples for the target domain; for the jth source domain, feature extractor G is used, which corresponds to the fault classification model j Extracting sample characteristics->Will be->Classifier F fed into corresponding fault classification model j Sample prediction label for obtaining jth source domainThrough calculation of the cross entropy loss function, the accuracy of the prediction label of the corresponding source domain output by the fault classification model classifier is improved; obtaining L by averaging and calculating cross entropy loss function results trained by N fault classification models corresponding to source domains C The calculation method is as follows:
For the integrated source domain, its corresponding fault classification integrated model feature extractor G N+1 Extracting sample featuresSample characteristics->Input corresponding classifier F N+1 Sample prediction tag of comprehensive source domain is obtained>Performing supervised learning training on the comprehensive source domain samples and sample labels by using a cross entropy loss function, and improving the accuracy of the prediction labels of the comprehensive source domain output by the fault classification comprehensive model classifier through calculation of the cross entropy loss function; the cross entropy loss function calculation result of the fault classification comprehensive model on the comprehensive source domain training is obtained +.>The calculation method is as follows:
in order to model predictive labels of domain invariant features, let G1 be j And G1 N+1 Feature extractor, G2, for extracting domain features respectively representing fault classification model and fault classification comprehensive model j And G2 N+1 Feature extractor for extracting domain difference features respectively representing fault classification model and fault classification comprehensive model, F1 j And F1 N+1 Classifier for classifying domain features respectively representing fault classification model and fault classification comprehensive model, F2 j And F2 N+1 Respectively represent fault classification modelsClassifier for classifying domain difference features by model and fault classification comprehensive model and feature extractor G1 j The extracted domain features are input into classifier F1 j Predictive tags to derive domain featuresFeature extractor G2 j The extracted domain difference features are input into a classifier F2 j Predictive tag for obtaining domain difference features +.>The two expressions +.>And->Subtracting to obtain predictive label of domain invariant feature of N source domain samples +.>Prediction tag of domain invariant feature of integrated source domain sample can be obtained in the same way>The calculation method is as follows:
where j=1,..
Extracting constraint domain difference features, extracting N source domain samples from a fault classification model and extracting constraint of domain difference features of comprehensive source domain samples from a fault classification comprehensive modelBy loss function L G A computing implementation, wherein K represents the number of individual source domain vibration signal fault types, in the following manner:
extracting constraint of domain difference characteristics of comprehensive source domain and target domain for fault classification comprehensive model and utilizing loss functionThe calculation is realized in the following way:
3. the method for predicting bearing failure type based on multi-source-domain transfer learning as claimed in claim 1, wherein in the second step, the prediction labels of the integrated source-domain samples obtained in the first step are respectively input into the same countermeasure learning network as the prediction labels of the N source-domain samples for training, and the countermeasure learning network is implemented by introducing a discriminator D h So that the feature extractor is opposed to the discriminator D h Predictive labels responsible for outputting N source domain samples to fault classification modelPredictive tag for outputting comprehensive source domain samples with fault classification comprehensive model>Distinguishing; feature extractor G using fault classification synthesis model N+1 The extraction of domain invariant features of the comprehensive source domain sample and N source domain samples is realized, so that the fault classification comprehensive model outputs a prediction label of the comprehensive source domain sample +.>Deception discriminator D h The method comprises the steps of carrying out a first treatment on the surface of the The predictive labels of the comprehensive source domain samples output by the fault classification comprehensive model and the predictive labels of the N source domain samples output by the fault classification model are respectively used as a discriminator D h To perform a contrast loss function +.>The calculation of (2) is as follows:
countering loss functionThe domain discrimination loss of the prediction label of the comprehensive source domain sample and the prediction label of each source domain sample is represented, so that the one-by-one countermeasure learning of the fault classification comprehensive model and the fault classification models corresponding to N source domains is realized;
for N countermeasures against loss functionAveraged to L adv The calculation method is as follows:
in order to model predictive labels for identifying domain invariant features, let D1 represent a discriminator for identifying domain features in an countermeasure learning network, D2 represent a discriminator for identifying domain difference features in the countermeasure learning network, and apply predictive labels for domain features Inputting discriminator D1 to obtain the result of discriminating domain feature prediction tag +.>Predictive tag for characterizing domain differences->Inputting the result of the discriminator D2 to obtain the discrimination domain difference characteristic predictive labelPredictive tag identifying domain invariant features of N source domain samples +.>Two expressions>And->The subtraction is calculated, and the predictive label for identifying the domain invariant feature of the comprehensive source domain sample can be obtained in the same way>The calculation method is as follows:
using a loss function L D The influence of the constraint domain difference feature on the discriminator is calculated, the discrimination capability of the discriminator on the domain invariant feature between N source domains and the comprehensive source domain is enhanced, and the calculation mode is as follows:
4. the method for predicting bearing failure type based on multi-source domain transfer learning of claim 1, wherein in step three, for the target domain sample, the failure classification integrated model feature extractor G N+1 Extracting target domain sample characteristics, and extracting the target domain sample characteristics as a fault classification comprehensive model classifier F N+1 Through input of fault classification comprehensive model classifierAnd outputting the prediction label of the target domain sample.
5. The method for predicting bearing failure type based on multi-source domain transfer learning as claimed in claim 1, wherein in the fourth step, the prediction label of the target domain sample in the third step and the prediction label of the integrated source domain sample are input to the countermeasure learning network in the second step to train, and the discriminator D is used for training h Predictive labels responsible for comprehensive source domain samplesAnd predictive tag of target domain sample +.>Distinguishing; the extraction of the domain invariant features between the comprehensive source domain sample and the target domain sample is realized by using the fault classification comprehensive model, so that the fault classification comprehensive model outputs a prediction label of the comprehensive source domain sampleSpoofing the discriminator; the predictive label of the comprehensive source domain sample output by the comprehensive fault classification model and the predictive label of the target domain sample output by the comprehensive fault classification model are used as a discriminator D h Input execution of (a) againstLoss function->The calculation method is as follows:
using loss functionsComputing the influence of the implementation constraint domain difference features on the discriminator, enhancing the discriminator D h For the discrimination capability of the domain invariant feature between the comprehensive source domain sample and the target domain sample, the calculation mode is as follows:
through continuous iterative training, the fault classification comprehensive model learns to extract domain invariant features between comprehensive source domain samples and target domain samples, L all The integrated loss function defines the total loss function as follows:
through the formula (14), the total loss function is trained through continuous iteration of the SGD random gradient descent method, and the network weight parameters are updated, so that the unlabeled type target domain samples output predicted sample labels through the fault classification comprehensive model.
6. A bearing failure type prediction system based on multi-source domain transfer learning applying the bearing failure type prediction method based on multi-source domain transfer learning according to any one of claims 1 to 5, characterized in that the bearing failure type prediction system based on multi-source domain transfer learning comprises:
the sample prediction label output module is used for outputting the prediction labels of the N source domain samples and the comprehensive source domain samples by utilizing the fault classification model and the fault classification comprehensive model;
the countermeasure training module is used for enabling the fault classification comprehensive model to learn domain invariant features of the comprehensive source domain and the N source domains through countermeasure training by using a countermeasure learning network;
the fault prediction module is used for outputting a prediction label of the target domain sample by using the fault classification comprehensive model, and the fault classification comprehensive model learns the domain invariant features of the comprehensive source domain and the target domain by countermeasure training.
7. A computer device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the bearing failure type prediction method based on multi-source domain transfer learning as claimed in any one of claims 1 to 6.
8. A computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the bearing failure type prediction method based on multi-source domain transfer learning as claimed in any one of claims 1 to 5.
9. An information data processing terminal for implementing the bearing failure type prediction system based on multi-source domain transfer learning as claimed in claim 6.
CN202310251823.0A 2023-03-14 2023-03-14 Bearing fault type prediction method and system based on multi-source domain transfer learning Active CN116304820B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310251823.0A CN116304820B (en) 2023-03-14 2023-03-14 Bearing fault type prediction method and system based on multi-source domain transfer learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310251823.0A CN116304820B (en) 2023-03-14 2023-03-14 Bearing fault type prediction method and system based on multi-source domain transfer learning

Publications (2)

Publication Number Publication Date
CN116304820A CN116304820A (en) 2023-06-23
CN116304820B true CN116304820B (en) 2023-12-01

Family

ID=86820070

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310251823.0A Active CN116304820B (en) 2023-03-14 2023-03-14 Bearing fault type prediction method and system based on multi-source domain transfer learning

Country Status (1)

Country Link
CN (1) CN116304820B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116859247B (en) * 2023-07-26 2024-06-14 哈尔滨工业大学 Electric vehicle lithium battery SOH estimation method based on multi-source transfer learning
CN116910632B (en) * 2023-09-14 2023-12-12 硕橙（厦门）科技有限公司 Radio frequency fingerprint identification method, device, equipment and medium based on transfer learning
CN117315487A (en) * 2023-11-02 2023-12-29 北京林业大学 Wild animal image cross-domain identification method based on deep migration learning
CN117933102B (en) * 2024-03-22 2024-06-14 华侨大学 MKTL-based multipoint frequency domain vibration response prediction method, MKTL-based multipoint frequency domain vibration response prediction device, MKTL-based multipoint frequency domain vibration response prediction equipment and medium
CN118035766A (en) * 2024-04-12 2024-05-14 太原理工大学 Variable working condition bearing fault diagnosis method based on similarity countermeasure and contrast learning
CN118094367A (en) * 2024-04-25 2024-05-28 广东石油化工学院 Passive field self-adaptive bearing fault classification method and system

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113537292A (en) * 2021-06-18 2021-10-22 杭州电子科技大学 Multi-source domain adaptation method based on tensor high-order mutual attention mechanism
CN114004252A (en) * 2021-10-13 2022-02-01 苏州大学 Bearing fault diagnosis method, device and equipment
CN115307910A (en) * 2022-08-08 2022-11-08 广东石油化工学院 Fault monitoring device and monitoring method for simulating gear of rotating unit
CN115328691A (en) * 2022-10-14 2022-11-11 山东大学 Fault diagnosis method, system, storage medium and equipment based on model difference
CN115479775A (en) * 2022-10-08 2022-12-16 重庆大学 Cross-bearing migration diagnosis method based on depth separable migration learning network
CN115562488A (en) * 2022-10-08 2023-01-03 上海前瞻创新研究院有限公司 Motor imagery brain-computer interface communication method, device, system, medium and equipment
CN115712859A (en) * 2022-11-22 2023-02-24 杭州电子科技大学 Multi-source domain equipment fault diagnosis method based on refined feature decoupling

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10776693B2 (en) * 2017-01-31 2020-09-15 Xerox Corporation Method and system for learning transferable feature representations from a source domain for a target domain
CN109710636B (en) * 2018-11-13 2022-10-21 广东工业大学 Unsupervised industrial system anomaly detection method based on deep transfer learning
US11526807B2 (en) * 2020-03-20 2022-12-13 Beijing Didi Infinity Technology And Development Co., Ltd. Machine learning systems and methods with source-target adaptation

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113537292A (en) * 2021-06-18 2021-10-22 杭州电子科技大学 Multi-source domain adaptation method based on tensor high-order mutual attention mechanism
CN114004252A (en) * 2021-10-13 2022-02-01 苏州大学 Bearing fault diagnosis method, device and equipment
CN115307910A (en) * 2022-08-08 2022-11-08 广东石油化工学院 Fault monitoring device and monitoring method for simulating gear of rotating unit
CN115479775A (en) * 2022-10-08 2022-12-16 重庆大学 Cross-bearing migration diagnosis method based on depth separable migration learning network
CN115562488A (en) * 2022-10-08 2023-01-03 上海前瞻创新研究院有限公司 Motor imagery brain-computer interface communication method, device, system, medium and equipment
CN115328691A (en) * 2022-10-14 2022-11-11 山东大学 Fault diagnosis method, system, storage medium and equipment based on model difference
CN115712859A (en) * 2022-11-22 2023-02-24 杭州电子科技大学 Multi-source domain equipment fault diagnosis method based on refined feature decoupling

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"Prediction Consistency Guided Convolutional Neural Networks for Cross-Domain Bearing Fault Diagnosis";SONGSONG WU.et al;《IEEE》;全文 *
"Classification Fusion for Bearing Fault Diagnosis with Multi-source Domain Shift";Huang Mu-sheng.et al;《Science publishing group》;全文 *

Also Published As

Publication number Publication date
CN116304820A (en) 2023-06-23

Similar Documents

Publication Publication Date Title
CN116304820B (en) Bearing fault type prediction method and system based on multi-source domain transfer learning
Zhao et al. Deep multi-scale convolutional transfer learning network: A novel method for intelligent fault diagnosis of rolling bearings under variable working conditions and domains
Chen et al. Deep transfer learning for bearing fault diagnosis: A systematic review since 2016
Zhang et al. Intelligent fault diagnosis under varying working conditions based on domain adaptive convolutional neural networks
Kuang et al. Class-imbalance adversarial transfer learning network for cross-domain fault diagnosis with imbalanced data
Feng et al. Similarity-based meta-learning network with adversarial domain adaptation for cross-domain fault identification
Li et al. Whitening-Net: A generalized network to diagnose the faults among different machines and conditions
Zhao et al. Adaptive open set domain generalization network: Learning to diagnose unknown faults under unknown working conditions
Yang et al. Diagnosis of incipient fault based on sliding-scale resampling strategy and improved deep autoencoder
Wen et al. A novel deep clustering network using multi-representation autoencoder and adversarial learning for large cross-domain fault diagnosis of rolling bearings
Yu et al. PCWGAN-GP: A new method for imbalanced fault diagnosis of machines
Zhang et al. A novel multiscale lightweight fault diagnosis model based on the idea of adversarial learning
Gu et al. An improved sensor fault diagnosis scheme based on TA-LSSVM and ECOC-SVM
Kumar et al. The Importance of Feature Processing in Deep‐Learning‐Based Condition Monitoring of Motors
CN114429152A (en) Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption
Shi et al. A reliable feature-assisted contrastive generalization net for intelligent fault diagnosis under unseen machines and working conditions
CN112926642A (en) Multi-source domain self-adaptive intelligent mechanical fault diagnosis method and system
Li et al. Intelligent fault diagnosis of aeroengine sensors using improved pattern gradient spectrum entropy
Wan et al. Bearing fault diagnosis based on multisensor information coupling and attentional feature fusion
CN116305119A (en) APT malicious software classification method and device based on predictive guidance prototype
Yu et al. A universal transfer network for machinery fault diagnosis
Sun et al. Fault diagnosis of rolling bearings based on CNN and LSTM networks under mixed load and noise
CN114626461A (en) Cross-domain target detection method based on domain self-adaptation
CN113884300A (en) Rolling bearing fault diagnosis method for deep anti-migration learning
Lu et al. A zero-shot intelligent fault diagnosis system based on EEMD

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant