CN113505664B - Fault diagnosis method for planetary gear box of wind turbine generator - Google Patents
Fault diagnosis method for planetary gear box of wind turbine generator Download PDFInfo
- Publication number
- CN113505664B CN113505664B CN202110719456.3A CN202110719456A CN113505664B CN 113505664 B CN113505664 B CN 113505664B CN 202110719456 A CN202110719456 A CN 202110719456A CN 113505664 B CN113505664 B CN 113505664B
- Authority
- CN
- China
- Prior art keywords
- domain
- sample
- semi
- source domain
- training
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- 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
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/70—Wind energy
- Y02E10/72—Wind turbines with rotation axis in wind direction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
- Wind Motors (AREA)
Abstract
The invention relates to a fault diagnosis method for a planetary gear box of a wind turbine generator, which comprises the steps of collecting vibration signals of the planetary gear box as samples, wherein the samples comprise a label-containing source domain sample, a label-free source domain sample and a target domain sample to be diagnosed; performing signal preprocessing, and converting each sample into a rapid spectral kurtosis image to obtain a marked source domain, a non-marked source domain and a target domain; constructing a deep residual semi-supervised domain generalization network structure, setting hyper-parameters required in training, taking a marked source domain and a non-marked source domain as input, training the deep residual semi-supervised domain generalization network by adopting a confrontation game mechanism for generating a confrontation network based on Wasserstein and a semi-supervised learning method based on pseudo labels, obtaining a final diagnosis model, and performing fault identification. Compared with the prior art, the method has excellent classification performance, can popularize the diagnosis model into a fault diagnosis task of the planetary gearbox with unknown rotating speed, and has high diagnosis accuracy.
Description
Technical Field
The invention relates to the field of wind turbine generator control, monitoring and diagnosis, in particular to a wind turbine generator planetary gear box fault diagnosis method.
Background
The planetary gear box has the advantages of compact structure, high power density, high transmission efficiency and the like, and is an important component in a transmission system of a wind turbine generator. During actual operation, planetary gearboxes are prone to failure and cause high maintenance costs due to dynamic load loads and frequently changing operating conditions. Therefore, the realization of accurate diagnosis of the fault of the gearbox has important significance for improving the safety and the reliability of the wind turbine generator.
In order to solve the problem that label training data contained in fault diagnosis is insufficient, transfer learning can learn knowledge from different but related domains and then transfer the knowledge to a target domain to realize a target task, so that deep transfer learning becomes a research hotspot for fault diagnosis of mechanical equipment. Deep migration learning generally trains a fault diagnosis model by using a large number of labeled source domain samples and a small number of target domain samples, and the fault diagnosis task of a target domain can be effectively realized by using the model. However, these methods trained models can only handle the diagnostic task of the target domain well and the training phase still requires target domain data. In practical application, the rotating speed of the planetary gearbox is constantly changed, and the change of the rotating speed directly causes the change of sample distribution, because various rotating speeds, even labeled samples under the condition of variable speed, cannot be comprehensively collected, the existing diagnosis model cannot identify the health state of equipment under various rotating speeds, and therefore the accuracy of model diagnosis is influenced.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provide a fault diagnosis method for a planetary gearbox of a wind turbine generator.
The purpose of the invention can be realized by the following technical scheme:
a fault diagnosis method for a planetary gear box of a wind turbine generator comprises the following steps:
s1, collecting a vibration signal of a planetary gear box as a sample, wherein the sample comprises a label-containing source domain sample, a label-free source domain sample and a target domain sample to be diagnosed;
s2, signal preprocessing is carried out, and the label source domain sample, the label-free source domain sample and the target domain sample to be diagnosed are all converted into a fast spectral kurtosis image, so that a labeled source domain omega is obtained ls No label source domain omega us And a target domain Ω;
s3, constructing a deep residual semi-supervised domain generalization network structure, and setting hyper-parameters required in training to contain a labeled source domain omega ls And a label-free source domain omega us For input, a countermeasure game mechanism for generating a countermeasure network based on Wasserstein (Waserstein) and a semi-supervised learning method based on pseudo labels are adopted to train a deep residual error semi-supervised domain generalization network to obtain a final diagnosis model;
and S4, inputting the target domain omega into the final diagnosis model for fault recognition, and outputting a diagnosis result.
Further, in the step S3, the depth residual semi-supervised domain generalization network includes a generator G, a classifier C and a discriminator D; the generator G extracts deep features from the original sample based on a depth residual error network; the classifier C adopts a Softmax classifier and uses the features extracted by the generator G to identify the health state; the discriminator D is used to estimate the Wasserstein distance between features extracted from different domains.
Further, in the step S3, the training of the deep residual semi-supervised domain generalization network includes pseudo label-based semi-supervised learning and WGAN-based domain confrontation learning, the classifier C and the discriminator D respectively adopt semi-supervised learning and domain confrontation learning training, and the generator G is trained together by using semi-supervised learning and domain confrontation learning.
Further, the loss function expression of the semi-supervised learning is as follows:
L Semi ＝L C +η(i)L PL i＝1,2,...,m
where η (i) is a pseudo-label coefficient function that varies with the number of iterations i, m is the maximum number of iterations, L C As a classifier loss function, L PL Is a pseudo tag loss function.
Further, the classifier loss function L C The expression of (a) is:
L C ＝E[-y ls log((C(G(x ls ))) T )]
in the formula (I), the compound is shown in the specification,
for samples containing labelled source domains
True label of l s Is the sample length;
the pseudo tag loss function L PL The expression of (a) is:
in the formula (I), the compound is shown in the specification,
as unlabeled source domain samples
The pseudo tag of (1).
Further, the pseudo label coefficient function η (i) is represented as:
in the formula I 1 And I 2 For iterative threshold, η 0 And η f Initial coefficients and final coefficients, respectively.
Further, the loss function expression of the domain confrontation learning is as follows:
a classifier G:
L A-G ＝-L adv
＝E[D(G(x ls ))]-E[D(G(x us ))]
a discriminator D:
in the formula, L adv For the domain antagonistic learning loss function, L GP As a gradient penalty function, x ls For samples containing a tagged source domain, x us Is a sample of unlabeled source domain, λ is a gradient penalty coefficient, μ -U [0, 1%]。
Further, the final diagnosis model comprises a generator G and a classifier C after training.
Further, in step S3, the hyper-parameters required in the training include a gradient penalty coefficient λ, a pseudo tag coefficient function η (i), and a batch size N B Number of batches n batch And maximum number of iterations n of training epochs 。
Further, the expression of the fast spectral kurtosis in the fast spectral kurtosis image is:
where < > is the time domain averaging operation, and H (t, f) is the envelope of the input signal x (t).
Compared with the prior art, the invention has the following beneficial effects:
1. according to the method, the wind turbine generator set planetary gearbox fault diagnosis is carried out by designing a deep residual semi-supervised domain generalization network (DRSDGN), and the collected diagnosis knowledge in the sample containing the label can be generalized to a diagnosis model of the unknown variable-speed sample, so that the method is more suitable for the actual application scene of the planetary gearbox with continuously-changing rotating speed during operation, and the fault diagnosis accuracy is improved.
2. The invention uses Fast Kurtogram (Fast Kurtogram) image to preprocess the vibration signal, obtains the image containing rich time-frequency information, and realizes fault diagnosis more easily than one-dimensional signal.
3. The method uses the depth residual error network (DRN) to extract the features of the image, is less prone to gradient disappearance compared with a general convolutional neural network, and can successfully extract the deep features of the image.
4. The invention introduces semi-supervised learning, can fully utilize sample data acquired in practical application, including data containing labels and data without labels, and improves the precision of the training model.
Drawings
Fig. 1 is a specific flow chart of the present embodiment.
Figure 2 is a DRSDGN network framework.
FIG. 3 is a gearbox fault simulation experiment platform.
Fig. 4 shows the state of health of the planet.
FIG. 5 is a time domain waveform of vibration of the planetary gearbox under different faults.
FIG. 6 is a graph showing the change in speed under the change in speed condition of the planetary gear box.
FIG. 7 is a Fast Kurtogram image generated from variable speed condition vibration signals.
FIG. 8 is a plot of AUC versus T3 for the set 1 experiment.
FIG. 9 is the AUC curve for T3 for experiment set 12.
Fig. 10 is a group 8 classification result confusion matrix.
FIG. 11 is the visualization result of the feature t-SNE extracted from the 8 th group of experimental networks.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.
The embodiment provides a wind turbine generator planetary gear box fault diagnosis method based on a deep residual error semi-supervised domain generalization network (DRSDGN), as shown in fig. 1, the method includes the following steps:
s1, collecting a vibration signal of a planetary gear box as a sample, wherein the sample comprises a label-containing source domain sample, a label-free source domain sample and a target domain sample to be diagnosed;
s2, performing signal preprocessing, and converting the label source domain sample, the label-free source domain sample and the target domain sample to be diagnosed into Fast Kurtogram images, thereby obtaining omega containing the label source domain ls No label source domain omega us And a target domain Ω;
s3, constructing a depth residual semi-supervised domain generalization network structure, and setting hyper-parameters required in training to contain a mark source domain omega ls And unlabeled source domain omega us For input, training a deep residual error semi-supervised domain generalization network by adopting a countermeasure game mechanism for generating a countermeasure network based on Wasserstein and a semi-supervised learning method based on a pseudo label to obtain a final diagnosis model;
and S4, inputting the target domain into the final diagnosis model to perform fault identification, outputting a diagnosis result, and then evaluating the diagnosis result by adopting multiple indexes (such as an accuracy index and an ROC (ROC) curve).
The embodiment aims to solve the problem that the fault diagnosis task of the planetary gearbox of the wind turbine generator set under variable rotating speed is difficult to complete by the existing deep migration learning method, and provides a diagnosis model capable of generalizing the diagnosis knowledge in the collected labeled sample to an unknown rotating speed sample.
The time-frequency analysis helps to reveal the characteristics of the vibration signal in the time domain and the frequency domain, the signal preprocessing process converts the vibration signal into a Fast spectral Kurtogram (Fast Kurtogram) image containing rich fault characteristics, and the spectral Kurtogram (Kurtogram) is defined as:
where < > is the time domain averaging operation, and H (t, f) is the envelope of the input signal x (t).
The structure of a deep residual semi-supervised domain generalization network (DRSDGN) is shown in fig. 2, which is based on the framework of WGAN and combines the powerful deep feature extraction capability and domain generalization theory of DRN, and only one labeled source domain and one unlabeled source domain are used in training without the need of a sample of the target domain. In the network, a generator G extracts deep features from an original sample based on a Deep Residual Network (DRN), and a classifier C performs health status identification using the extracted features. The extracted features need to be sensitive to the health status classification but not to the domain classification, so the network learning process can be divided into semi-supervised learning based on pseudo-labels and domain confrontation learning based on WGAN. And extracting domain discrimination characteristics through a semi-supervised learning guidance generator G, and diagnosing the health state of the target domain sample by using the characteristics. Domain counterstudy is used to guide the generator G to extract domain-invariant features, while training of the discriminator D aims to better estimate the Wasserstein-1 distance between features extracted from different domains. In brief, classifier C and discriminator D employ semi-supervised learning and domain confrontation learning, respectively, while generator G requires both to be trained together. Finally, a diagnosis model composed of G and C which are trained is used for realizing fault diagnosis of the target domain with unknown rotating speed.
Semi-supervised learning aims at fully utilizing available samples, training samples are labeled source domains and unlabeled source domains, and training objects are a generator G and a classifier C. In semi-supervised learning with pseudo-labels, the loss function consists of two parts: there is a supervised classification loss LC and a pseudo label iteration loss LPL. The semi-supervised learning loss function is as follows:
L Semi ＝L C +η(i)L PL i＝1,2,...,m
where η (i) is a pseudo-label coefficient function that varies as the number of iterations i varies, m is the maximum number of iterations, L C As a classifier loss function, L PL Is a pseudo tag loss function. L is C And L PL Can be expressed as:
L C ＝E[-y ls log((C(G(x ls ))) T )]
in the formula (I), the compound is shown in the specification,
for samples containing labelled source domains
True tags of l s Is the sample length;
as unlabeled source domain samples
The pseudo label of (c) is calculated as follows:
in early iterations, η (i) needs to be kept at a small value η because classifier C cannot provide sufficiently accurate labels 0 Then as the iteration progresses, the coefficient grows linearly to η f And keeping the concrete expression until the training is finishedThe following were used:
in the formula I 1 And I 2 For iterative threshold, η 0 And η f Initial coefficients and final coefficients, respectively.
In order for the generator G to be able to extract domain invariant features, WGAN-based domain antagonistic learning needs to be designed. In a conventional WGAN, a discriminator is used to estimate the Wasserstein-1 distance between the true sample probability distribution and the generated sample probability distribution, and by countertraining, G can produce better samples to minimize the Wasserstein-1 distance. Based on the above idea, in DRSDGN, the training arbiter D estimates the Wasserstein-1 distance between features extracted from the labeled source domain samples and the unlabeled source domain samples, respectively, while the training generator G extracts the domain invariant features to minimize them. The Wasserstein-1 distance is an effective measurement method for estimating distribution difference, so that the features extracted after domain confrontation training are not sensitive to domain classification. The loss functions of G and D during domain confrontation learning are as follows:
L A-G ＝-L adv
＝E[D(G(x ls ))]-E[D(G(x us ))]
in the formula, L adv For domain confrontation learning loss function, L GP As a gradient penalty function, x ls Source field sample containing label, x us Is a label-free source domain sample, and λ is a gradient penalty coefficient, μ -U [0, 1%]。
In DRSDGN, classifier C and discriminator D are trained using semi-supervised learning and domain counterlearning, respectively, while generator G is trained using both. In order to ensure that the trained generator G can extract features with both fault discrimination and domain invariance, semi-supervised learning and domain counterlearning are required to be performed simultaneously. Thus, the final classifier, the arbiter and the loss function of the classifier are as follows:
L D ＝L A-G +L Semi
L C ＝L Semi
L G ＝L A-D
in the formula, L A-D And L A-G Loss functions of the discriminator and the generator in domain confrontation learning, L, respectively Semi Is a semi-supervised learning loss function.
As shown in fig. 1, the specific steps of this embodiment include:
1) Collecting vibration signals by using a vibration acceleration sensor, carrying out domain division and respectively converting each domain into a Fast Kurtogram atlas to obtain omega containing a marker source domain ls No label source domain omega us And a target domain omega t Determining the structures of a generator G, a classifier C and a discriminator D, and setting the hyper-parameters required in training, including a gradient penalty coefficient lambda, a pseudo tag coefficient function eta (i), and a batch size N B Number of batches n batch Number of D iterations n per G iteration critic Optimization algorithm of G, D, C, maximum number of iterations n of training epochs 。
2) Using omega ls Pre-training G and C, initializing the parameters of D.
3) At omega ls Middle divided mark-containing batch B ls At Ω us Middle divided unmarked batch B us 。
4) From B ls And B us Training D in random sampling.
5) Repeating the step 4) to the maximum iteration times, and calculating B us And training G and C.
6) Repeating the steps 3) -5) to traverse the omega ls And Ω us 。
7) Repeating step 6) to a set maximum epochs.
8) Constructing a final fault diagnosis model by using the trained G and C, and using omega t The model is evaluated.
The DRSDGN specific network parameters are as follows, and the network structure is generated as shown in FIG. 2The G is a DRN comprising 9 residual error units, comprising 19 convolutional layers and 1 full-connection layer, the network input dimension is 224 multiplied by 3, in order to improve the calculation efficiency and extract the local features, the size of the convolutional kernel is set to be 3 multiplied by N ω ，N ω Is the number of input channels. In the DRN structure in fig. 2, "/2" indicates that the convolutional layer step size is 2, and to reduce the dimension of the feature mapping, 3 convolutional layers are selected in the network, where the step size is set to 2, m represents the number of convolutional kernels in the first layer, and in order to enable deep features to appear many times, the convolutional kernels are increased to 2m and 4m, and 32 is taken as m. The rest of the network structure is shown in table 1.C and D take the output of G as input, the hyper-parameters in the training process are shown in a table 2, adam optimization algorithm is used in the training of G, D and C, and the learning rate is set to be 0.001.
TABLE 1 network architecture
TABLE 2 DRSDGN hyper-parameters
The specific verification scheme is as follows:
as shown in FIG. 3, the present embodiment first establishes a wind turbine gearbox simulation platform. The test planet is mounted within the gearbox housing and the accelerometer is mounted on the gearbox housing to measure the vibration signal. The motor speed can be varied by a speed controller, and the rotational frequency can be set in the range of 0 to 60Hz. The sampling frequency of the signal is 12kHz. The fault planetary gear is sequentially normal, pitting, abrasion and tooth breakage faults from left to right as shown in FIG. 4. The health of the planetary gears includes: under normal, abrasion, pitting and broken tooth conditions, when the rotating speed of the driving motor is 1800rpm, the acquired time domain waveform of the planet wheel is shown in fig. 5 (a).
1) The fan gear box simulation platform is used for collecting data, and the rotating speeds of the motors are set to be 900rpm, 1200rpm, 1500rpm and 1800rpm in the process. The different health conditions were 100s for each speed. In addition, the load is changed, the rotation speed of the regulating motor is increased from 900rpm to 1500rpm for 100s, the rotation speed is changed as shown in FIG. 6, and the vibration waveform in 1s under the variable speed condition is shown in FIG. 5 (b).
2) In order to ensure that each sample contains abundant fault characteristics as much as possible, 100s of samples are intercepted in each group of vibration signals, and 100 sample data are generated, namely each sample contains 1s of characteristic information and 12000 data points. Fast Kurtogram analysis is adopted to convert the time domain vibration samples into time-frequency images, 400 image samples are obtained at each rotating speed, 1600 constant-speed samples and 400 variable-speed samples are obtained, and the example images under the variable-speed working condition are shown in figure 7.
3) For the acquired image samples, a sample set at the same rotation speed is defined as one domain, and the total number of the domains is five. To verify the effectiveness of the proposed method, as shown in table 3, 12 sets of planetary gear box domain generalization diagnostic experiments were designed, in each set of experiments, a labeled source domain (LS) and a label-free source domain (US) training network was used, and a plurality of target domains (T1, T2, T3) were used to evaluate the diagnostic model.
TABLE 3 domain generalization of diagnostic tasks
4) To reduce the influence of randomness, each set of experiments was repeated 10 times, the overall average accuracy was 95.24%, ROC curves were introduced to evaluate the classification performance of the models, and based on the combination of the true and predicted classes, the test samples were classified into True Positive (TP), false Positive (FP), true Negative (TN) and False Negative (FN), and the True Positive Rate (TPR) and the negative positive rate (FPR) were calculated as follows:
an ROC curve is drawn by taking FPR and TPR as horizontal and vertical coordinates, DRSDGN has the highest average accuracy in the experiment of the 1 st group, and has the lowest average accuracy in the experiment of the 12 th group, and in order to fully verify the network performance, ROC curves of the two groups of experiments for T3 classification are drawn, and AUC is calculated and is shown in fig. 8 and 9. It can be seen that in the above two experiments, the AUC value of the classifier after DRSDGN training is close to 1, and both are significantly better than those of other algorithms, which indicates that the classifier with excellent performance can be trained by the method of this embodiment, and further, the health status of the variable speed planetary gear box can be identified.
In addition, in order to further verify the effectiveness of the DRSDGN, an 8 th group of experiments with the average accuracy rate close to 95.24% are selected for visual analysis, the trained diagnostic model is used for classifying the state of the target domain sample, and a classification result confusion matrix is shown in figure 10. The characteristic dimension reduction visualization of the fully connected layer in the generator G by using a t-distribution stored neighbor embedding (t-SNE) algorithm is shown in FIG. 11.
The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations can be devised by those skilled in the art in light of the above teachings. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.
The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations can be devised by those skilled in the art in light of the above teachings. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.
Claims (6)
1. A fault diagnosis method for a planetary gear box of a wind turbine generator is characterized by comprising the following steps:
s1, collecting a vibration signal of a planetary gear box as a sample, wherein the sample comprises a label-containing source domain sample, a label-free source domain sample and a target domain sample to be diagnosed;
s2, signal preprocessing is carried out, and the label source domain sample, the label-free source domain sample and the target domain sample to be diagnosed are all converted into a rapid spectral kurtosis image, so that a labeled source domain omega is obtained ls No marker Source Domain omega us And a target domain Ω;
s3, constructing a depth residual semi-supervised domain generalization network structure, and setting hyper-parameters required in training to contain a labeled source domain omega ls And unlabeled source domain omega us For input, a countermeasure game mechanism for generating a countermeasure network based on Wasserstein and a semi-supervised learning method based on pseudo labels are adopted to train a deep residual error semi-supervised domain generalization network to obtain a final diagnosis model;
s4, inputting the target domain omega into the final diagnosis model for fault recognition, and outputting a diagnosis result;
in the step S3, the depth residual semi-supervised domain generalization network comprises a generator G, a classifier C and a discriminator D; the generator G extracts deep features from the original sample based on a depth residual error network; the classifier C adopts a Softmax classifier and uses the features extracted by the generator G to identify the health state; the discriminator D is used for estimating Wasserstein distances among the features extracted from different fields;
in the step S3, the training of the deep residual semi-supervised domain generalization network includes semi-supervised learning based on pseudo labels and domain confrontation learning based on WGAN, the classifier C and the discriminator D respectively adopt semi-supervised learning and domain confrontation learning training, and the generator G uses the semi-supervised learning and the domain confrontation learning for training together;
the loss function expression of the semi-supervised learning is as follows:
L Semi ＝L C +η(i)L PL i＝1,2,...,m
where η (i) is a pseudo-label coefficient function that varies as the number of iterations i varies, m is the maximum number of iterations, L C As a classifier loss function, L PL Is a pseudo tag loss function;
the classifier penalty function L C The expression of (a) is:
L C ＝E[-y ls log((C(G(x ls ))) T )]
in the formula (I), the compound is shown in the specification,
for samples containing labelled source domains
True label of l s Is the sample length;
the pseudo tag loss function L PL The expression of (a) is:
in the formula (I), the compound is shown in the specification,
as unlabeled source domain samples
The pseudo tag of (1).
2. The wind turbine generator set planetary gearbox fault diagnosis method according to claim 1, characterized in that the pseudo-tag coefficient function η (i) is expressed as:
in the formula I 1 And I 2 For iterative threshold, η 0 And η f Initial coefficients and final coefficients, respectively.
3. The method for diagnosing the fault of the planetary gearbox of the wind turbine generator set according to claim 1, wherein the loss function expression of the domain confrontation learning is as follows:
a classifier G:
L A-G ＝-L adv
＝E[D(G(x ls ))]-E[D(G(x us ))]
a discriminator D:
L A-D ＝L adv +λL GP
＝-E[D(G(x ls ))]+E[D(G(x us ))]+λE[(||▽D(μG(x ls )+(1-μ)G(x us ))||-1) 2 ]
in the formula, L adv For domain confrontation learning loss function, L GP As a gradient penalty function, x ls For samples containing a tagged source domain, x us Is a sample of unlabeled source domain, λ is a gradient penalty coefficient, μ -U [0, 1%]。
4. The wind turbine generator set planetary gearbox fault diagnosis method as claimed in claim 1, wherein the final diagnosis model comprises a generator G and a classifier C after training.
5. The method for diagnosing the fault of the planetary gearbox of the wind turbine generator set as claimed in claim 1, wherein in the step S3, the hyper-parameters required in training comprise a gradient penalty coefficient lambda, a pseudo tag coefficient function eta (i) and a batch size N B Number of batches n batch And maximum number of iterations n of training epochs 。
6. The method for diagnosing the fault of the planetary gearbox of the wind turbine generator set according to claim 1, wherein an expression of a fast spectral kurtosis in a fast spectral kurtosis image is as follows:
where < > is the time domain averaging operation, and H (t, f) is the envelope of the input signal x (t).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110719456.3A CN113505664B (en) | 2021-06-28 | 2021-06-28 | Fault diagnosis method for planetary gear box of wind turbine generator |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110719456.3A CN113505664B (en) | 2021-06-28 | 2021-06-28 | Fault diagnosis method for planetary gear box of wind turbine generator |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113505664A CN113505664A (en) | 2021-10-15 |
| CN113505664B true CN113505664B (en) | 2022-10-18 |
Family
ID=78010933
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110719456.3A Active CN113505664B (en) | 2021-06-28 | 2021-06-28 | Fault diagnosis method for planetary gear box of wind turbine generator |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113505664B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114021285B (en) * | 2021-11-17 | 2024-04-12 | 上海大学 | Rotary machine fault diagnosis method based on mutual local countermeasure migration learning |
| CN114526915B (en) * | 2022-01-21 | 2022-12-16 | 华南理工大学 | Early fault time point identification method based on rolling bearing characteristic distribution |
| CN114638060B (en) * | 2022-03-10 | 2023-02-17 | 重庆英科铸数网络科技有限公司 | Fault prediction method, system and electronic equipment |
| CN117875421B (en) * | 2024-03-13 | 2024-05-17 | 中国水利水电第一工程局有限公司 | Deep learning-based photovoltaic array fault diagnosis method and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111080645A (en) * | 2019-11-12 | 2020-04-28 | 中国矿业大学 | Remote sensing image semi-supervised semantic segmentation method based on generating type countermeasure network |
| CN111626317A (en) * | 2019-08-14 | 2020-09-04 | 广东省智能制造研究所 | Semi-supervised hyperspectral data analysis method based on double-flow conditional countermeasure generation network |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11538143B2 (en) * | 2018-10-26 | 2022-12-27 | Nec Corporation | Fully convolutional transformer based generative adversarial networks |
| CN109977094B (en) * | 2019-01-30 | 2021-02-19 | 中南大学 | Semi-supervised learning method for structured data |
| CN110617966A (en) * | 2019-09-23 | 2019-12-27 | 江南大学 | Bearing fault diagnosis method based on semi-supervised generation countermeasure network |
| CN111504644B (en) * | 2020-04-01 | 2021-04-09 | 西安交通大学 | Bearing fault diagnosis method for adaptive anti-domain based on embedding distinguishability condition |
| CN112183581A (en) * | 2020-09-07 | 2021-01-05 | 华南理工大学 | Semi-supervised mechanical fault diagnosis method based on self-adaptive migration neural network |
| CN112729835B (en) * | 2021-01-30 | 2022-06-03 | 温州大学 | Multi-domain semi-supervised fault diagnosis method and device for axial plunger pump bearing |
-
2021
- 2021-06-28 CN CN202110719456.3A patent/CN113505664B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111626317A (en) * | 2019-08-14 | 2020-09-04 | 广东省智能制造研究所 | Semi-supervised hyperspectral data analysis method based on double-flow conditional countermeasure generation network |
| CN111080645A (en) * | 2019-11-12 | 2020-04-28 | 中国矿业大学 | Remote sensing image semi-supervised semantic segmentation method based on generating type countermeasure network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113505664A (en) | 2021-10-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113505664B (en) | Fault diagnosis method for planetary gear box of wind turbine generator | |
| Han et al. | An adaptive spatiotemporal feature learning approach for fault diagnosis in complex systems | |
| CN109580215B (en) | Wind power transmission system fault diagnosis method based on deep generation countermeasure network | |
| CN110334764B (en) | Rotary machine intelligent fault diagnosis method based on integrated depth self-encoder | |
| CN110375987B (en) | Mechanical bearing fault detection method based on deep forest | |
| CN112200244B (en) | Intelligent detection method for anomaly of aerospace engine based on hierarchical countermeasure training | |
| CN109376620A (en) | A kind of migration diagnostic method of gearbox of wind turbine failure | |
| CN110110768B (en) | Rolling bearing fault diagnosis method based on parallel feature learning and multiple classifiers | |
| CN106874957A (en) | A kind of Fault Diagnosis of Roller Bearings | |
| CN112629851B (en) | Offshore wind turbine gearbox fault diagnosis method based on data enhancement method and image recognition | |
| CN113076834B (en) | Rotating machine fault information processing method, processing system, processing terminal, and medium | |
| CN109255333B (en) | Fault hybrid diagnosis method for rolling bearing of large wind turbine generator | |
| CN112257862B (en) | Semi-supervised identification method based on relation network marking sample expansion | |
| CN115600150A (en) | Multi-mode gearbox fault diagnosis method based on deep migration learning | |
| CN112113755A (en) | Mechanical fault intelligent diagnosis method based on deep convolution-kurtosis neural network | |
| CN116358871B (en) | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network | |
| Wang et al. | An end-to-end fault diagnostics method based on convolutional neural network for rotating machinery with multiple case studies | |
| CN116028876A (en) | Rolling bearing fault diagnosis method based on transfer learning | |
| CN114091504A (en) | Rotary machine small sample fault diagnosis method based on generation countermeasure network | |
| CN112160877A (en) | Fan bearing fault diagnosis method based on SFA and CNN | |
| CN111539381B (en) | Construction method of wind turbine bearing fault classification diagnosis model | |
| Zhang et al. | Gearbox health condition identification by neuro-fuzzy ensemble | |
| CN116625686A (en) | On-line diagnosis method for bearing faults of aero-engine | |
| CN115795397A (en) | Gearbox temperature prediction method based on 1DCNN-LSTM and BiLSTM parallel network | |
| CN115374829A (en) | Deep learning-based bearing fault diagnosis method and system |
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 |