CN115618933A - Cross-domain fault diagnosis method based on sampling convolution and counterstudy - Google Patents

Cross-domain fault diagnosis method based on sampling convolution and counterstudy Download PDF

Info

Publication number
CN115618933A
CN115618933A CN202211233006.4A CN202211233006A CN115618933A CN 115618933 A CN115618933 A CN 115618933A CN 202211233006 A CN202211233006 A CN 202211233006A CN 115618933 A CN115618933 A CN 115618933A
Authority
CN
China
Prior art keywords
domain
cross
odd
loss
fault diagnosis
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.)
Pending
Application number
CN202211233006.4A
Other languages
Chinese (zh)
Inventor
李佐勇
卢维楷
樊好义
周常恩
陈健
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.)
Minjiang University
Original Assignee
Minjiang University
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 Minjiang University filed Critical Minjiang University
Priority to CN202211233006.4A priority Critical patent/CN115618933A/en
Publication of CN115618933A publication Critical patent/CN115618933A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Complex Calculations (AREA)

Abstract

The invention relates to a cross-domain fault diagnosis method based on sampling convolution and counterstudy. Firstly, a sampling convolution network is constructed to extract feature embedding containing the correlation between the parity subsequences, then a classifier is used for fault classification, and a domain discriminator is used for discriminating the source condition of a domain. To align the feature distributions of the source and target domains, a Maximum Mean Difference (MMD) loss and a domain confrontation loss are introduced to participate in the training of the network. The invention improves the accuracy of cross-domain fault diagnosis.

Description

Cross-domain fault diagnosis method based on sampling convolution and counterstudy
Technical Field
The invention belongs to the field of time sequence analysis, and particularly relates to a cross-domain fault diagnosis method based on sampling convolution and counterstudy.
Background
As one of the key components of a rotary machine, the condition of the rotary bearing is very important for safe operation of the machine. The intelligent fault diagnosis aims to judge the type of early fault of the rotary bearing according to vibration signals collected by the sensor. The accurate and reliable fault diagnosis can prevent the whole system from being broken down, thereby ensuring the safe and reliable operation of the system. In recent years, with the development of artificial intelligence technology, intelligent fault diagnosis technology has been widely used in modern industry.
Traditional fault diagnosis methods based on machine learning, including support vector machines, logistic regression, etc., have been widely studied and applied in real scenes. However, conventional methods rely on artificial features and advanced signal processing techniques, which have limited generalization capabilities. With the explosive growth of data volume, the method based on deep learning gradually becomes a research hotspot due to the characteristic that manual features are not needed as input, and higher diagnosis precision is obtained. For example, jiang et al [1] developed a multi-scale convolutional neural network that can capture richer diagnostic information and achieve better performance. Yin et al [2] propose an optimized long-short time memory network with cosine loss function, which converts the loss from euclidean space to angular space, reducing the influence of signal strength.
The conventional fault diagnosis method assumes that the training sample and the test sample are independently and identically distributed. However, such assumptions often do not hold in practical application scenarios, e.g., the data distribution collected from the bearings under different operating conditions is widely different. Collecting sufficient training data for each different operating condition is labor intensive and expensive, which limits the realistic reference of these approaches. To solve the above problems, unsupervised Domain Adaptation (UDA), which aims to migrate knowledge from a labeled domain to an unlabeled domain, has been widely used for fault diagnosis. Existing UDA-based approaches mainly include mapping-based approaches and countermeasure-based approaches. The mapping-based approach minimizes the distribution distance of the learned features so that the feature extractor can map the source and target domains into a shared feature space. The countermeasure-based approach uses domain discriminators for countermeasure training so that the feature extractor can extract domain-invariant features.
While existing methods have achieved encouraging cross-domain fault diagnosis results, they fail to account for temporal relationships within the vibration signal, resulting in sub-optimal diagnostic accuracy. The odd-even subsequences of the time sequence are sequence pairs with unit time offset, and the relationship between the odd-even subsequences and the even-even subsequences reflects the slight change rule of the time sequence. To capture this variation law, we propose a domain-opposed-sampling convolutional network (DASCN).
Disclosure of Invention
The invention aims to provide a cross-domain fault diagnosis method based on sampling convolution and counterstudy, and accuracy of cross-domain fault diagnosis is improved.
In order to realize the purpose, the technical scheme of the invention is as follows: a cross-domain fault diagnosis method based on sampling convolution and counterstudy comprises the steps of firstly, constructing a sampling convolution network to extract feature embedding containing correlation between odd-even subsequences; then, using a classifier to classify the fault, and using a domain discriminator to discriminate the domain source condition; finally, maximum mean variance loss and domain opposition loss are introduced to align the feature distributions of the source and target domains.
Compared with the prior art, the invention has the following beneficial effects: most existing domain adaptive fault diagnosis methods fail to take into account the temporal relationship within the vibration signal, resulting in suboptimal diagnostic accuracy. The present invention designs a domain-opposed-sampling convolutional network (DASCN) to take into account the temporal relationship inside the vibration signal in the process of domain adaptation. DASCN down-samples the original signal into odd and even subsequences, extracts their feature representations using convolutional neural networks and performs feature interaction to obtain an embedded representation containing its correlation. And finally, aligning the feature distribution of the source domain and the target domain through the combined domain confrontation training and MMD module. The invention improves the accuracy of cross-domain fault diagnosis.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is an overview of DASCN.
Fig. 3 is a detailed process of the downsampling module.
Detailed Description
The technical scheme of the invention is specifically explained below with reference to the accompanying drawings.
The invention relates to a cross-domain fault diagnosis method based on sampling convolution and antagonistic learning, which comprises the following steps of firstly, constructing a sampling convolution network to extract characteristic embedding containing correlation between odd-even subsequences; then, a classifier is used for fault classification, and a domain discriminator is used for discriminating the source condition of the domain; finally, maximum mean variance loss and domain opposition loss are introduced to align the feature distributions of the source and target domains.
The following is a specific implementation process of the present invention.
The invention discloses a cross-domain fault diagnosis method based on sampling convolution and counterstudy, which is shown in figure 1. First, raw data is input into a sampled convolutional network to extract a feature representation containing internal relationship information. Then, the obtained features are input into a classification network for fault classification, and input into a domain discriminator network for domain confrontation training. In addition, MMD penalty is used to align feature distribution differences between the source and target domains.
1. Problem definition
In the present invention, we investigated the UDA fault diagnosis problem. Suppose we have a slave source domain P s Sampled tagged data sets
Figure BDA0003881674430000021
From the target domain P t Middle sampled markerless data sets
Figure BDA0003881674430000022
The two data sets have the same class space but have larger data distribution differences. Given D s And D t Our goal is to train the function f (-) to predict the samples from P t The failure category of the sample.
2. Feature extractor
Is under work [3]Inspiring, we designed an interactive network as the feature extractor Z (·). For an original signal X from a source domain or a target domain, we first split it into odd sub-sequences X using a downsampling module odd And even subsequence X even . Fig. 2 shows a detailed process of the module.
Next, we use a four-layer CNN encoder sharing weights from X odd And X even To extract feature representation. Due to the loss of information in the two sub-sequences caused by the down-sampling, feature interaction between the two sequences is achieved by learning affine transformation functions. This process can be defined as:
F′ odd ＝F odd ⊙exp(φ(F even ))+ψ(F even )
F′ even ＝F even ⊙exp(φ(F odd ))+ψ(F odd )
wherein, an is a Hadamard product; exp (·) is an exponential transformation function with e as base; f' odd And F' even Are respectively F odd And F even The feature transformation result of (1). Phi (-) and psi (-) are two projection functions that project the sub-sequence into two hidden states, for the multiplication and addition transformations of another sequence, respectively. Finally, fusing two subsequence features through nonlinear transformation to realize overall features:
F＝ReLU(W[F′ odd， F′ even ]+b)
where ReLU (. Cndot.) is a non-linear activation function, [. Cndot. ], a matrix join operation, and W and b are the trainable weight matrix and bias vector, respectively.
3. Fault classifier
To ensure basic classification functionality, the extracted features are classified using a fully connected layer with softmax activation. In the training phase, only labeled source domain data is used as input to the classifier. To minimize the difference between the prediction classes and the true labels, the cross entropy loss is considered as a classification loss:
Figure BDA0003881674430000031
where L (-) is the cross entropy loss, C (-) is the classifier, and E (-) represents the mathematical expectation.
4. Domain discriminator
Due to the large difference in distribution between the source domain and the target domain, it is difficult to obtain satisfactory diagnostic results in the target domain samples by training only with source domain samples. To address this problem, a domain discriminator composed of multiple layers of perceptrons is used to determine whether a feature is from a source domain or a target domain, and a antagonism training is performed to fool the feature extractor and the domain discriminator. We use cross-entropy loss as the domain classification loss, which can be defined as:
Figure BDA0003881674430000032
wherein D (-) represents a domain discriminator; max of Z Representation optimization Z to maximize the loss L D ；min D Representing optimization D to minimize loss L D 。
5. Feature distribution alignment
To align the feature distribution difference with the source and target domains, MMD penalty is applied to the proposed method, which penalty function can be defined as:
Figure BDA0003881674430000041
where Ω is the regenerative hilbert space.
6. Global objective function
In combination with classification loss, domain confrontation loss, and feature distribution alignment loss, the final optimization objective of the proposed method can be defined as:
L sl ＝L C +αL C +βL MMD
where α and β are adjustable parameters.
Description of Experimental data
TABLE 1
Working conditions Load torque Radial force Speed of rotation
0 0.7 1000 1500
1 0.7 1000 900
2 0.1 1000 1500
3 0.7 400 1500
The present invention verifies its validity on the published Paderborn University (PU) data set. The PU data set was collected from the padboen university platform test, including both bearing data for artificially induced damage and true damage. The present invention has four different operating conditions at different drive system speeds, radial forces on the test bearings and load torques on the drive system, as shown in table 1. According to the setting of [4], thirteen failure categories KA04, KA15, KA16, KA22, KA30, KB23, KB24, KB27, KI14, KI16, KI17, KI18, and KI21 are employed.
(II) results of the experiment
TABLE 2
Figure BDA0003881674430000042
Figure BDA0003881674430000051
Table 2 shows the experimental results of the proposed DASCN and other comparative methods. It can be observed that DASCN performs better than all comparative methods in all tasks except tasks 1-3. DASCN performs best with average accuracy over 12 tasks, 9.39% higher performance than the best comparison method CDAN + E. Notably, in tasks 1-0, the diagnostic accuracy of DASCN increased by 18.43% over the optimal baseline. These performance enhancements demonstrate the effectiveness of DASCN in cross-domain troubleshooting tasks.
Reference:
[1]Chen,Z.,Gryllias,K.,Li,W.:Mechanical fault diagnosis using convolutional neural networks and extreme learning machine.Mechanical systems and signal processing 133,106272(2019)
[2]Yin,A.,Yan,Y.,Zhang,Z.,Li,C.,Sánchez,R.V.:Fault diagnosis of wind turbine gearbox based on the optimized lstm neural network with cosine loss.Sensors 20(8),2339(2020)
[3]Liu,M.,Zeng,A.,Xu,Z.,Lai,Q.,Xu,Q.:Time series is a special sequence:Forecasting with sample convolution and interaction.arXiv preprint arXiv:2106.09305(2021)
[4]hao Z,Zhang Q,Yu X,Sun C,Wang S,Yan R,et al.Applications of unsupervised deep transfer learning to intelligent fault diagnosis:a survey and comparative study.IEEE Transactions on Instrumentation and Measurement 2021.
[5]Sun B,Saenko K.Deep coral:Correlation alignment for deep domain adaptation.In:European conference on computer vision Springer；2016.p.443–450.
[6]Long M,Cao Z,Wang J,Jordan MI.Conditional adversarial domain adaptation.Advances in neural information processing systems 2018；31.
[7]Long M,Zhu H,Wang J,Jordan MI.Deep transfer learning withjoint adaptation networks.In:International conference on machine learning PMLR；2017.p.2208–2217.。
the above are preferred embodiments of the present invention, and all changes made according to the technical solutions of the present invention that produce functional effects do not exceed the scope of the technical solutions of the present invention belong to the protection scope of the present invention.

Claims (7)

1. A cross-domain fault diagnosis method based on sampling convolution and antagonistic learning is characterized in that firstly, a sampling convolution network is constructed to extract feature embedding containing correlation between odd-even subsequences; then, using a classifier to classify the fault, and using a domain discriminator to discriminate the domain source condition; finally, maximum mean variance loss and domain opposition loss are introduced to align the feature distributions of the source and target domains.
2. The cross-domain fault diagnosis method based on sampling convolution and counterlearning of claim 1, wherein the cross-domain fault diagnosis problem is defined as follows:
suppose there is a slave source domain P s Sampled tagged data sets
Figure FDA0003881674420000011
From the target domain P t Mid-sampled unlabeled dataset
Figure FDA0003881674420000012
The two data sets have the same category space but have data distribution differences; given D s And D t The target is a training function f (·), the prediction sample is from P t Of the unlabeled data set
Figure FDA0003881674420000013
The failure category of the sample.
3. The cross-domain fault diagnosis method based on sampling convolution and counterlearning of claim 2, wherein the specific implementation manner of constructing a sampling convolution network to extract feature embedding containing the correlation between the parity subsequences is as follows:
designing an interactive network as a feature extractor Z (-) to obtain a feature value; for an original signal X from a source domain or a target domain, firstly, a downsampling module is used to split the original signal X into odd subsequences X odd And even subsequence X even ；
Four-layer CNN encoder using shared weights from X odd And X even Extracting characteristic representation; feature interaction between the two sequences is realized by learning an affine transformation function; the process is defined as:
F′ odd ＝F odd ⊙exp(φ(F even ))+ψ(F even )
F′ even ＝F even ⊙exp(φ(F odd ))+ψ(F odd )
wherein, is the Hadamard product; exp (·) is an exponential transformation function with e as base; f' odd And F' even Are respectively F odd And F even The feature transformation result of (2); phi (-) and psi (-) are two projection functions that project the sub-sequence into two hidden states for multiplication and addition transformations of another sequence, respectively;
fusing two subsequence features through nonlinear transformation to realize overall features:
F＝ReLU(W[F′ odd ,F′ even ]+b)
where ReLU (. Cndot.) is a non-linear activation function, [. Cndot.,. Cndot.) is a matrix join operation, and W and b are a trainable weight matrix and a bias vector, respectively.
4. The cross-domain fault diagnosis method based on sampling convolution and antagonistic learning as claimed in claim 3, wherein the specific implementation manner of fault classification using the classifier is as follows:
classifying the extracted features using a fully connected layer with softmax activation; in the training phase, only the tagged source domain data is used as input to the classifier; to minimize the difference between the prediction classes and the true labels, the cross entropy penalty is considered as a classification penalty:
Figure FDA0003881674420000021
where L (-) is the cross entropy loss, C (-) is the classifier, and E (-) represents the mathematical expectation.
5. The cross-domain fault diagnosis method based on sampling convolution and counterstudy is characterized in that the specific implementation manner of using the domain discriminator to discriminate the source condition of the domain is as follows:
determining whether the feature is from a source domain or a target domain using a domain discriminator composed of multiple layers of perceptrons, and performing a antagonism training to spoof the feature extractor and the domain discriminator; using cross-entropy penalties as domain classification penalties, the penalty function is defined as:
Figure FDA0003881674420000022
wherein D (-) represents a domain discriminator; max Z Representation optimization Z to maximize the loss L D ；min D Representing optimization D to minimize loss L D 。
6. The cross-domain fault diagnosis method based on sampling convolution and antagonistic learning, according to claim 5, is characterized in that the specific implementation manner of aligning the feature distribution of the source domain and the target domain by introducing the maximum average difference loss and the domain antagonistic loss is as follows:
introducing MMD penalty, aligning the feature distribution difference with the source domain and the target domain, the penalty function being defined as:
Figure FDA0003881674420000023
wherein Ω is a regenerated hilbert space.
7. The cross-domain fault diagnosis method based on sampling convolution and antagonistic learning according to claim 6, wherein a final optimization objective is defined by combining classification loss, domain antagonistic loss and feature distribution alignment loss as follows:
L s1 ＝L C +αL C +βL MMD
where α and β are adjustable parameters.
CN202211233006.4A 2022-10-10 2022-10-10 Cross-domain fault diagnosis method based on sampling convolution and counterstudy Pending CN115618933A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211233006.4A CN115618933A (en) 2022-10-10 2022-10-10 Cross-domain fault diagnosis method based on sampling convolution and counterstudy

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211233006.4A CN115618933A (en) 2022-10-10 2022-10-10 Cross-domain fault diagnosis method based on sampling convolution and counterstudy

Publications (1)

Publication Number Publication Date
CN115618933A true CN115618933A (en) 2023-01-17

Family

ID=84860215

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211233006.4A Pending CN115618933A (en) 2022-10-10 2022-10-10 Cross-domain fault diagnosis method based on sampling convolution and counterstudy

Country Status (1)

Country Link
CN (1) CN115618933A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116401532A (en) * 2023-06-07 2023-07-07 山东大学 Method and system for recognizing frequency instability of power system after disturbance

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116401532A (en) * 2023-06-07 2023-07-07 山东大学 Method and system for recognizing frequency instability of power system after disturbance
CN116401532B (en) * 2023-06-07 2024-02-23 山东大学 Method and system for recognizing frequency instability of power system after disturbance

Similar Documents

Publication Publication Date Title
Li et al. A two-stage transfer adversarial network for intelligent fault diagnosis of rotating machinery with multiple new faults
CN111161315B (en) Multi-target tracking method and system based on graph neural network
CN111709448A (en) Mechanical fault diagnosis method based on migration relation network
CN114048568A (en) Rotating machine fault diagnosis method based on multi-source migration fusion contraction framework
Yao et al. Multiscale domain adaption models and their application in fault transfer diagnosis of planetary gearboxes
CN116894187A (en) Gear box fault diagnosis method based on deep migration learning
CN110728694A (en) Long-term visual target tracking method based on continuous learning
CN112784872B (en) Cross-working condition fault diagnosis method based on open set joint transfer learning
CN116028876A (en) Rolling bearing fault diagnosis method based on transfer learning
CN115618933A (en) Cross-domain fault diagnosis method based on sampling convolution and counterstudy
CN116026593A (en) Cross-working-condition rolling bearing fault targeted migration diagnosis method and system
Wang et al. Multi-scale style generative and adversarial contrastive networks for single domain generalization fault diagnosis
CN114626454A (en) Visual emotion recognition method integrating self-supervision learning and attention mechanism
Kumar et al. Knowledge addition for improving the transfer learning from the laboratory to identify defects of hydraulic machinery
Chu et al. Illumination-guided transformer-based network for multispectral pedestrian detection
Yuan et al. A fusion TFDAN-Based framework for rotating machinery fault diagnosis under noisy labels
Chen et al. Unsupervised structure subdomain adaptation based the Contrastive Cluster Center for bearing fault diagnosis
Tian et al. A fault diagnosis method for few-shot industrial processes based on semantic segmentation and hybrid domain transfer learning
CN116740434A (en) Transformer-based cross-domain double-branch countermeasure domain adaptive image classification method
Tang et al. HmmSeNet: A Novel Single Domain Generalization Equipment Fault Diagnosis Under Unknown Working Speed Using Histogram Matching Mixup
Yang et al. Fault diagnosis of motor bearing based on improved convolution neural network based on VMD
Wen et al. Bearing fault diagnosis via fusing small samples and training multi-state siamese neural networks
Xu et al. A novel clustering based on consensus knowledge for cross-domain fault diagnoses
Espitia et al. Novel methods based on deep learning applied to condition monitoring in smart manufacturing processes
Yu et al. Towards real-time object tracking with deep Siamese network and layerwise aggregation

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