CN116337448A - Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention - Google Patents

Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention Download PDF

Info

Publication number
CN116337448A
CN116337448A CN202310176104.7A CN202310176104A CN116337448A CN 116337448 A CN116337448 A CN 116337448A CN 202310176104 A CN202310176104 A CN 202310176104A CN 116337448 A CN116337448 A CN 116337448A
Authority
CN
China
Prior art keywords
domain
network
bearing
loss
data
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
CN202310176104.7A
Other languages
Chinese (zh)
Inventor
张勋兵
袁海飞
李聪聪
王斌
王继伟
洪麒麟
李晓炜
张凤勇
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.)
Xuzhou XCMG Excavator Machinery Co Ltd
Original Assignee
Xuzhou XCMG Excavator Machinery Co Ltd
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 Xuzhou XCMG Excavator Machinery Co Ltd filed Critical Xuzhou XCMG Excavator Machinery Co Ltd
Priority to CN202310176104.7A priority Critical patent/CN116337448A/en
Publication of CN116337448A publication Critical patent/CN116337448A/en
Pending legal-status Critical Current

Links

Images

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
    • G06N3/04—Architecture, e.g. interconnection topology
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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)
  • Data Exchanges In Wide-Area Networks (AREA)

Abstract

The invention discloses a method, a device and a storage medium for diagnosing a fault of a transfer learning bearing based on width multi-scale space-time attention, wherein the method comprises the following steps: carrying out multi-layer wavelet packet decomposition processing on the bearing vibration signals to obtain signal components of the vibration signals in different frequency domains, and inputting a pre-trained bearing fault classification model to obtain bearing fault classification results; the bearing fault classification model comprises a feature extraction network, a label classification network and a domain discrimination network, wherein the feature extraction network utilizes convolution kernels with different sizes to extract multi-scale features of signal components, and a higher weight is given to the fault features on space and channels by using an attention mechanism; constraining the distribution difference of the data characteristics of the source domain and the target domain by using CORAL; and a four-layer convolution structure is used for replacing a full connection layer of the CDAN domain discrimination network, so that the information fusion capability of the two domains is enhanced. The invention combines multiscale feature extraction, CORAL distribution adaptation and condition countermeasure migration technology to construct the bearing fault diagnosis model, and can improve the accuracy of diagnosis.

Description

Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention
Technical Field
The invention relates to the technical field of bearing fault diagnosis, in particular to a method and a device for diagnosing a fault of a transfer learning bearing based on width multi-scale space-time attention and a storage medium.
Background
The traditional fault diagnosis generally needs effective signal analysis means and priori knowledge as guidance, and cannot meet the requirement of large data development in the industrial field. The occurrence of deep learning provides a new means for the bearing fault diagnosis field, and the strong characteristic extraction capability of the deep semantic features of the original bearing vibration signals are automatically extracted, so that the influence of human experience is reduced, and the method is more suitable for processing complex bearing vibration data. However, most of the training and testing data used by current deep learning models are collected from the same rotating machine under the same operating conditions, however, in actual industrial scenarios, this assumption is almost impossible to meet. Therefore, the existing bearing fault diagnosis technology still has the following defects:
(1) The traditional fault requires prior knowledge of an expert and is easily influenced by subjective factors;
(2) Bearing fault data collected in an engineering actual environment are often label-free data, and if a supervised learning or semi-supervised learning method is adopted, the classification accuracy is not ideal;
(3) Although the accuracy of the convolutional neural network-based method is obviously improved compared with that of the traditional method, the overall recognition accuracy is not high because the fault data in the variable working condition environment does not meet the assumption of independent same distribution. Meanwhile, the model has the problems of poor stability, weak noise resistance and the like;
(4) In the prior art, although a certain preprocessing operation is implemented on the acquired data, the depth fault characteristics of the bearing are not fully extracted, and the fault classification accuracy of the model is affected;
(5) By adopting a feature learning method, although a corresponding feature migration network is constructed, the distribution of multi-source fault feature data in a high-dimensional feature space cannot be sufficiently reduced, so that the fault classification accuracy is not ideal;
(6) For the whole bearing input data, certain feature redundancy exists when the bearing input data enters a feature extraction network, so that subsequent key feature information is lost, and the fault classification accuracy is low.
(7) For the whole bearing input data, when entering the domain discriminator network, shallow key information features are easy to ignore, which results in insufficient integration capability of two-domain key feature information, thereby reducing the overall network performance.
Disclosure of Invention
The invention aims to provide a method, a device and a storage medium for diagnosing a migration learning bearing fault based on width multi-scale space-time attention, which are used for realizing training of a bearing fault diagnosis model by combining multi-scale feature extraction, a CBAM attention module and a migration resisting technology, and improving the accuracy of bearing fault diagnosis. The technical scheme adopted by the invention is as follows.
In one aspect, the present invention provides a bearing fault diagnosis method, including:
acquiring bearing vibration signal data;
carrying out multi-layer wavelet packet decomposition processing on the obtained bearing vibration signal data to obtain signal components of the bearing vibration signal in different frequency domains;
inputting the signal component of the bearing vibration signal data into a pre-trained bearing fault classification model to obtain a bearing fault classification result;
the bearing fault classification model comprises a feature extraction network, a label classification network and a domain discrimination network, wherein the feature extraction network comprises a convolution layer, an attention module and a correlation alignment algorithm module, the convolution layer is provided with a plurality of convolution kernels and is used for extracting multi-scale features of the signal components, and the attention module is used for carrying out weighting processing on the extracted key features; the correlation alignment algorithm module is used for restricting the distribution of the source domain and target domain characteristics in a high-dimensional space.
Optionally, the bearing fault classification model further includes a convolution layer with a large convolution kernel, for reducing feature redundancy of the input signal component, and an output end of the convolution layer is connected to the feature extraction network. The convolution layer preferably employs a 32 x 1 convolution kernel.
Optionally, the bearing fault classification model adopts a CDAN (Conditional Adversarial Domain Adaptation Network) network, and a full connection layer of the domain discrimination network is replaced by a convolution layer. The four-layer convolution structure can be used for replacing the original full-connection layer, so that the network depth can be expanded, and the two-domain information integration capability can be enhanced.
Optionally, the feature extraction network includes four convolution kernels of 3×1, 5×1, 7×1, and 9×1.
Optionally, the pre-training method of the bearing fault classification model includes:
acquiring bearing vibration signals under various working conditions;
normalizing the obtained bearing vibration signal;
acquiring source domain data with fault classification labels and target domain data without fault classification labels based on the bearing vibration signals after normalization processing;
performing multi-layer wavelet packet decomposition processing on the source domain data and the target domain data to obtain signal components of the data in different frequency domains;
inputting the signal components of the source domain data and the signal components of the target domain data into the feature extraction network respectively, performing multi-feature extraction, weighting key features through an attention module, and restricting the distribution of the source domain features and the target domain features in a high-dimensional space through a correlation alignment algorithm module to obtain the source domain features and the target domain features;
sending the extracted source domain features into a label classification network for fault classification, and calculating a cross entropy loss function to obtain label classification loss; the source domain features and the target domain features are sent into a domain judging network together, and the second-order statistical feature distance loss and domain judging device loss function of the source domain and the target domain are calculated;
and training a bearing fault classification model by minimizing the label classification loss, the second-order statistical feature distance loss of the source domain and the target domain and maximizing the domain discriminator loss, and the back propagation optimization feature extraction network, the label classification network and the domain discrimination network.
Optionally, the processing of decomposing the obtained bearing vibration signal data by using a multi-layer wavelet packet is as follows: and performing four-layer wavelet packet decomposition processing on the obtained bearing vibration signal data.
Optionally, the second order statistical feature distance loss of the source domain and the target domain is expressed as:
Figure BDA0004100899220000031
in the method, in the process of the invention, I.I. | F Represents the Frobenius norm, D s Representing source domain data, D t Representing target domain data, d representing sample dimensions, C s Representing a source domain covariance matrix, C t Representing the target domain covariance matrix.
Optionally, the minimizing tag classification loss is expressed as:
Figure BDA0004100899220000032
the maximum domain arbiter loss is expressed as:
Figure BDA0004100899220000033
in θ f A representative feature extraction network parameter vector,
Figure BDA0004100899220000034
representing the optimal solution of the feature extraction network parameter vector at saddle points, θ y Representing mapping parameters->
Figure BDA0004100899220000035
Representing the optimal solution of the mapping parameters at saddle points, < >>
Figure BDA0004100899220000036
Representing domain arbiter loss, θ d Representing domain arbiter network parameters, arg min E represents a minimization loss function.
Optionally, a gradient inversion layer is arranged between the domain discrimination network and the feature extraction network for automatically inverting during back propagation, so that domain discriminator loss, label classification loss and L CORAL The loss gradient is the same;
the loss function of the bearing failure classification model is expressed as:
Figure BDA0004100899220000041
wherein N represents a positive integer, d i Domain tag representing ith training sample, L y Representing tag predictive loss, G y Representing tag predictor functions, G f Representing a feature extraction network function, x i Training samples representing source and target domains, y i Representing a label sample corresponding to a source domain, lambda represents a meta-parameter, L d Representing domain discrimination loss, G d Representing domain classification network functions, R λ Representing forward transmissionBroadcasting function, theta f Representing feature extraction network parameter vector, θ d Representing domain-classified network parameters, L CORAL Representing the second order statistical feature distance loss of the source domain and the target domain.
In a second aspect, the present invention provides a bearing failure diagnosis apparatus comprising:
the to-be-diagnosed data acquisition module is configured to acquire bearing vibration signal data;
the signal processing module is configured to perform multi-layer wavelet packet decomposition processing on the acquired bearing vibration signal data to obtain signal components of the bearing vibration signal in different frequency domains;
the fault classification module is configured to input signal components of the bearing vibration signal data into a pre-trained bearing fault classification model to obtain a bearing fault classification result;
the bearing fault classification model comprises a feature extraction network, a label classification network and a domain discrimination network, wherein the feature extraction network comprises a convolution layer, an attention module and a correlation alignment algorithm module, the convolution layer is provided with a plurality of convolution kernels and is used for extracting multi-scale features of the signal components, and the attention module is used for carrying out weighting processing on the extracted key features; the correlation alignment algorithm module is used for restricting the distribution of the source domain and target domain characteristics in a high-dimensional space.
Optionally, the signal processing module performs four-layer wavelet packet decomposition processing on the obtained bearing vibration signal data;
the bearing fault classification model adopts a CDAN network, and a full connection layer of a domain discrimination network is replaced by a convolution layer; the CDAN network also comprises a convolution layer with a large convolution kernel, wherein the convolution layer is used for reducing the feature redundancy of the input signal component, and the output end of the convolution layer is connected with the feature extraction network; a gradient turnover layer is arranged between the domain judging network and the characteristic extracting network and is used for automatically inverting during counter propagation, so that domain judging device loss, label classifying loss and L CORAL The loss gradient is the same;
the pre-training method of the bearing fault classification model comprises the following steps:
acquiring bearing vibration signals under various working conditions;
normalizing the obtained bearing vibration signal;
acquiring source domain data with fault classification labels and target domain data without fault classification labels based on the bearing vibration signals after normalization processing;
performing multi-layer wavelet packet decomposition processing on the source domain data and the target domain data to obtain signal components of the data in different frequency domains;
inputting the signal components of the source domain data and the signal components of the target domain data into the feature extraction network respectively, performing multi-feature extraction, weighting key features through an attention module, and restricting the distribution of the source domain features and the target domain features in a high-dimensional space through a correlation alignment algorithm module to obtain the source domain features and the target domain features;
sending the extracted source domain features into a label classification network for fault classification, and calculating a cross entropy loss function to obtain label classification loss; the source domain features and the target domain features are sent into a domain judging network together, and the second-order statistical feature distance loss and domain judging device loss function of the source domain and the target domain are calculated;
and training a bearing fault classification model by minimizing the label classification loss, the second-order statistical feature distance loss of the source domain and the target domain and maximizing the domain discriminator loss, and the back propagation optimization feature extraction network, the label classification network and the domain discrimination network.
In a third aspect, the present invention provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the bearing fault diagnosis method as described in the first aspect.
Advantageous effects
Compared with the prior art, the invention has the following advantages and advances:
(1) The wavelet packet decomposition is adopted to process the acquired bearing vibration signal data, so that the time-frequency domain bearing fault characteristic data under different scales can be obtained, and the efficiency of subsequent feature extraction is improved; in addition, the signal components after wavelet packet decomposition are processed through a layer of large convolution check, so that the feature redundancy of bearing faults can be reduced, and the subsequent feature extraction efficiency and feature reliability are further improved;
(2) Based on a CDAN network, the fault characteristics between two domains conveyed by the characteristic extractor cannot be resolved through the confusion domain discriminator, so that the fault characteristics and the fault characteristics reach Nash balance, and the problem of low classification accuracy caused by overlarge data distribution difference between the two domains is solved through the distribution between Ji Yuanyu and a target domain;
(3) The characteristic extraction network of the CDAN network adopts a plurality of different convolution kernels to extract component characteristic information with different frequencies, so that the integrity of the characteristic information can be ensured, and meanwhile, a CBAM module is utilized to endow key fault characteristics with higher weight to improve the performance of the characteristic extraction network;
(4) Based on a feature extraction network, the key features given with weights are constrained to be distributed in a high-dimensional space by CORAL, so that the distribution difference of two-domain data is reduced;
(5) The domain countermeasure network of the CDAN network utilizes a four-layer convolution structure to replace the full-connection layer to expand the network depth, so that the phenomena of gradient disappearance and the like caused by layer number limitation are reduced, the integration capability of fault information is enhanced, and the overall classification performance of the model is improved.
Drawings
FIG. 1 is a schematic diagram of a CDAN network model architecture for bearing fault diagnosis according to the present invention;
FIGS. 2 a-2 d are t-SNE graphs of various fault diagnosis models in a B2-B0 migration mode of the rotational speed data set, wherein the fault diagnosis model corresponding to FIG. 2a is WMSRCIDANN, the fault diagnosis model corresponding to FIG. 2B is WMSRCDANN, the fault diagnosis model corresponding to FIG. 2c is WMSRIDANN, and the fault diagnosis model corresponding to FIG. 2d is WCIDANN;
fig. 3 a-3 d are graphs showing the classification results of confusion matrices corresponding to various fault diagnosis models in the migration mode of the rotational speed data set B2-B0, wherein the fault diagnosis model corresponding to fig. 3a is WMSRCIDANN, the fault diagnosis model corresponding to fig. 3B is WMSRCDANN, the fault diagnosis model corresponding to fig. 3c is WMSRIDANN, and the fault diagnosis model corresponding to fig. 3d is WCIDANN;
fig. 4 is a schematic diagram showing comparison of migration fault classification accuracy obtained by multiple fault diagnosis models.
Detailed Description
Noun interpretation in relation to the invention
Multiscale network: different convolution kernels are different in emphasis on extracting bearing fault deep characteristic information, so that different convolution kernels can be selected to fully extract fault characteristic information in each time-frequency component.
CBAM attention module: the system is a lightweight general module and consists of a channel attention module and a space attention module. The research is widely applied to visual semantics and natural language tasks, and the CBAM attention mechanism can enrich the feature extractor network and improve the network classification performance by giving higher weight to the extracted key fault features on the space and the channel.
CORAL: the method is a micro-loss function for minimizing the correlation between the source domain and the target domain, and the distribution difference of the two-domain data in a high-dimensional space is reduced by adjusting the second-order statistical feature distance of the source domain and the target domain distribution.
CDAN network: the CDAN uses GAN game ideas for research and is widely used in different natural language tasks, including: text classification, machine translation, and question-answering systems. Since CDAN has achieved great success in visual semantics, researchers have attempted to apply it in fault diagnosis. The CDAN architecture is based on GAN, and the mechanism is mainly used for solving the problem of overlarge data distribution difference of two domains. The main advantage of this over feature migration is that it is not necessary to choose different metrology parameter distances. In addition, due to the self characteristics of the CDAN, the CDAN can maintain good model stability, and meanwhile, the problem of low model classification accuracy caused by different distance parameters is avoided.
The technical conception of the invention is as follows: aiming at the source domain and target domain signals of the bearing vibration signals with overlarge data distribution difference, the integrity and reliability of feature extraction are improved by combining the multi-scale feature extraction, the CBAM attention mechanism, the CORAL and the CDAN network, and meanwhile, the accuracy of fault diagnosis results is improved.
Further description is provided below in connection with the drawings and the specific embodiments.
Example 1
The embodiment introduces a bearing fault diagnosis method, which includes:
acquiring bearing vibration signal data;
carrying out multi-layer wavelet packet decomposition processing on the obtained bearing vibration signal data to obtain signal components of the bearing vibration signal in different frequency domains;
inputting the signal component of the bearing vibration signal data into a pre-trained CDAN network to obtain a bearing fault classification result;
referring to fig. 1, in this embodiment, the CDAN network includes a feature extraction network, a label classification network and a domain discrimination network, where the feature extraction network has four convolution kernels of 3×1, 5×1, 7×1 and 9×1, and is used to extract multi-scale features of the signal components, and simultaneously, in combination with CBAM attention mechanisms, assign higher weights to the extracted key fault features of the bearing on space and channels, and the key features after being assigned weights constraint the distribution of the source domain and the target domain data features in a high-dimensional space by using CORAL, so as to reduce the distribution difference of the two domain data; a convolution layer of 32 x 1 large convolution kernel is also provided before this to reduce the feature redundancy of the input signal components. The full connection structure of the domain discrimination network is replaced by a four-layer convolution structure, so that the integration capability of the two-domain information is enhanced.
In this embodiment, the pretraining method of the CDAN network specifically includes the following steps.
S1, acquiring bearing vibration signals of various fault states under various working conditions, wherein the signal length can be set to 1024; and carrying out normalization processing on the obtained bearing vibration signals to obtain source domain data and target domain data for model training, wherein the source domain data is provided with a fault classification label, and the target domain data is provided with a fault classification label.
S2, performing four-layer wavelet packet decomposition processing on the source domain data and the target domain data obtained in the S1, and performing multi-scale refinement analysis on the bearing vibration signals by utilizing wavelet packet decomposition to obtain node coefficients of different layers, wherein the obtained wavelet packet node coefficients represent time-frequency domain characteristics of the bearing vibration signals under different scales.
S3, sending the bearing vibration signals obtained after wavelet packet decomposition processing to a characteristic extraction network of a CDAN network, firstly extracting fault characteristics through a layer of large convolution kernel, then fully extracting fault characteristics of source domain data and target domain data through four different convolution kernels of 3×1, 5×1, 7×1 and 9×1, and fully extracting bearing information characteristic emphasis points are different due to different receptive fields of different convolution kernels, so that the fully extracting bearing fault information characteristics by selecting four convolution kernels with different size scales can ensure the full extraction of bearing fault characteristics, simultaneously, integrating and screening rich fault characteristic information with different scales on channels and spaces by utilizing CBAM, endowing high weight to key information after integrating and screening, restraining the distribution of the characteristics after endowing weight in a high-dimensional space by utilizing CORAL, reducing the distribution difference of two-domain data, and improving the performance of the characteristic extraction network.
S4, calculating second-order statistical feature distance loss according to the extracted source domain features and the target domain features; the extracted source domain features are sent into a label classification network to carry out fault classification, fault classification labels of source domain data are utilized, and a cross entropy loss function, namely a label classification loss function, is calculated; the source domain features and the target domain features are sent into the domain discrimination network together, the fully connected network structure in the domain discriminator is replaced by a convolution structure, the network depth is expanded, and the integration capability of fault information is enhanced. And calculating a domain arbiter loss function, and enabling the domain arbiter to be incapable of distinguishing the source of the fault feature by minimizing the domain arbiter classification accuracy to confuse the source domain fault feature and the target domain fault feature. The label classifier aims to accurately classify source domain data, the domain discriminator functions as the discriminator functions in the GAN, and the label classifier aims to accurately classify depth features between two domains conveyed by the feature extractor as much as possible and obtain a domain classification result.
In the CDAN network, f=G is extracted through the characteristics f (x；θ f ) The extracted source domain feature vector and target domain feature vector pass through two branches: a first partThe branches are the source domain features entering the tag classifier G y (x；θ y ) Completing source domain data classification; another branch is the entry of source and target domain features into the domain classifier network G d (x；θ d ) And obtaining domain classification results. In the training process, the network wants to realize that the loss of the tag classifier is minimum, so that the tag classifier can accurately classify source domain data, and on the other hand, in order to finish migration, the domain classifier must be confused, so that the depth characteristics between two domains cannot be accurately judged, and the maximization of the error of the domain classifier is realized; thus, the goal minimizing label classification loss during training is expressed as:
Figure BDA0004100899220000081
the maximum domain arbiter loss is expressed as:
Figure BDA0004100899220000091
in θ f A characteristic network parameter vector is represented and,
Figure BDA0004100899220000092
representing the optimal solution of the characteristic network parameter vector at saddle points, θ y Representing mapping parameters->
Figure BDA0004100899220000093
Representing the optimal solution of the mapping parameters at saddle points, < >>
Figure BDA0004100899220000094
Representing domain arbiter loss, argminE represents the minimization of the loss function.
The CORAL loss is expressed as:
Figure BDA0004100899220000095
in the method, in the process of the invention, I.I. | F Representing Frobenuus norm, D s Representing source domain data, D t Representing target domain data, d representing sample dimensions, C s Representing a source domain covariance matrix, C t Representing the target domain covariance matrix.
The CDAN network of the embodiment adds a gradient turnover layer in front of the domain classifier for automatically inverting during counter propagation, so that the domain classification loss and the label classifier loss gradient direction are ensured to be the same. The CDAN loss function is expressed as follows:
Figure BDA0004100899220000096
wherein G is f (·；θ y ) Is the characteristic parameter extracted by the characteristic extraction network; g d (·；θ d ) Is a domain classification network and parameters therein; l (L) y And L d Representing tag classification loss and domain discrimination loss, respectively.
R in forward propagation after addition of gradient inversion layer λ (x) =x, automatic inversion dR in back propagation λ And/dx= - λi, so the above formula can be rewritten as:
Figure BDA0004100899220000097
to this end, minimizing tag classification loss by calculation
Figure BDA0004100899220000098
L of deep features of source domain and target domain CORAL Loss and Domain discriminator loss->
Figure BDA0004100899220000101
Feature extraction network, label classification network and domain discrimination network of back propagation optimized CDAN network until L CORAL 、L CORAL And->
Figure BDA0004100899220000102
The iteration times of the method reach the target requirement, and the CDAN network is completedAnd training the collaterals to obtain a fault diagnosis model.
When the method is practically applied to fault diagnosis of the bearing, the preprocessed bearing time-frequency domain data is input into a CDAN fault diagnosis model with complete training, and the fault type of the bearing can be obtained through a multi-feature extraction network and a label classification network.
The specific implementation effect of this embodiment is as follows:
(1) Introduction to data acquisition
Bearing fault diagnosis experiment platform has gathered the bearing data of 600rpm, 800rpm, three kinds of rotational speeds of 1000rpm under 50kHz sampling frequency, and each data contains inner circle, outer lane, rolling element, four kinds of bearing types of health, and the type mark is shown in table 1. The three working condition data sets are respectively a record data set B0, a data set B1 and a data set B2 from low to high according to different rotating speeds.
Table 1:
Figure BDA0004100899220000103
(2) Performing result analysis
And (3) processing the data in the table 1 by a signal processing module and a fault classification module to obtain an analysis and diagnosis result.
In order to verify the migration diagnosis effect of the migration learning bearing fault diagnosis method (herein denoted as WMSRCIDANN model) based on the width multi-scale space-time attention provided by the invention between different working conditions. Verification was performed by four sets of ablation experiments, with the model settings described below:
1) WMSRCDANN the same network structure as WMSRCIDANN feature extraction, but without modification of the domain arbiter.
2) WMSRIDANN no CBAM attention mechanism is added compared to WMSRCIDANN.
3) WRCIDANN: compared to WMSRCIDANN, without adding multiscales, resnet18 is used as a feature extraction network.
The results of the four model experiments are shown in table 2, from which the following conclusions can be drawn:
1) Although WMSRCDANN classification diagnosis results have a smaller gap compared with WMSRCIDANN network models, WMSRCDANN model stability is worse than WMSRCIDANN, classification accuracy is up to 99.31%, and the minimum is 96.25%, which indicates that convolution operation not only increases depth of a network, but also retains important parameters through weight sharing, so that the domain of bearing data can be better judged.
2) WMSRCIDANN network classification accuracy is higher than WMSRIDANN under each migration working condition, so that the key information characteristics of the bearing can be better reserved by adding the CBAM attention mechanism into the WMSR network, and the key fault information extraction capability of the WMSR network is improved.
3) The WMSRCIDANN model still keeps the highest fault classification average accuracy in 6 migration working conditions, and the average accuracy is 99.20%. It can be seen that the WMSRCIDANN model has good generalization ability in the faulty bearing dataset.
Table 2: WMSRCIDANN model and fault classification results of each ablation model
Figure BDA0004100899220000111
In order to compare the self-adaptive capacity of four model feature extraction networks to extract depth feature samples, we analyzed the sample clustering effect of each model using a t-SNE visualization method. FIGS. 2 a-2 d are graphs of t-SNE visual signature analysis diagnostic results for various diagnostic models in a rotational speed dataset-B0 migration mode. As can be seen from the t-SNE diagram, the WMSRCIDANN method can effectively increase the inter-class distance of the deep features of the variable-working-condition bearing and reduce the intra-class distance compared with other three methods. Again, the WMSRCIDANN model proved to have better sample clustering effect.
In order to analyze the classification result of each label in the B2-B0 working condition migration mode in further detail, the classification result of each model B2-B0 working condition migration is drawn into a confusion matrix, and the results are shown in fig. 3 a-3 d. Compared with other three network models, the WMSRCIDANN network label classification error number is the least, so that the WMSRCIDANN network model provided by the invention can fully extract the key information characteristics of bearing faults, enhance the integration capability of two-domain information and improve the resolution capability of the domain to which the two-domain data belong.
To demonstrate the superiority of the proposed model of the present invention, it was compared with CNN, adaBN, CORAL, MK-MMD 4 methods. The experimental results are shown in table 3 and fig. 4. It can be seen from table 3 and fig. 4 that the average classification accuracy of the WMSRCIDANN network is far higher than that of the other four network models, and the migration classification accuracy under six working conditions is also higher than that of the other five models. In particular, compared with CNN, the working condition migration average classification accuracy of the WMSRCIDANN network is 21.63% higher than that of the CNN. Therefore, under the variable working condition, WMSRCIDANN can better extract key depth fault characteristics reflecting vibration signals, reduce the difference of depth characteristic distribution of data in a source domain and a target domain, better judge the domain to which the data belongs, enhance the information integration capability and improve the fault classification accuracy of the data in the two domains.
Table 3: WMSRCIDANN table for comparing diagnostic results with existing models
Figure BDA0004100899220000121
Example 2
Based on the same inventive concept as embodiment 1, this embodiment introduces a bearing failure diagnosis apparatus including:
the to-be-diagnosed data acquisition module is configured to acquire bearing vibration signal data;
the signal processing module is configured to perform multi-layer wavelet packet decomposition processing on the acquired bearing vibration signal data to obtain signal components of the bearing vibration signal in different frequency domains;
the fault classification module is configured to input signal components of the bearing vibration signal data into a pre-trained bearing fault classification model to obtain a bearing fault classification result;
the bearing fault classification model comprises a feature extraction network, a label classification network and a domain discrimination network, wherein the feature extraction network comprises a convolution layer, an attention module and a correlation alignment algorithm module, the convolution layer is provided with a plurality of convolution kernels and is used for extracting multi-scale features of the signal components, the attention module utilizes a CBAM attention mechanism to give higher weight to the extracted bearing key fault features on a space and a channel, so that the feature extraction network is enriched, and the correlation alignment algorithm module utilizes CORAL to restrict the distribution of source domain and target domain data features in a high-dimensional space, so that the distribution difference of two-domain data is reduced.
The specific implementation of each functional module is described in embodiment 1, and is not repeated herein, specifically, in this embodiment:
the signal processing module is used for carrying out four-layer wavelet packet decomposition processing on the obtained bearing vibration signal data;
the bearing fault classification model adopts a CDAN network, a domain discrimination network replaces a full-connection layer by a four-layer convolution structure, the network depth is expanded, and the two-domain information integration capability is enhanced. The CDAN network also comprises a convolution layer with a large convolution kernel, wherein the convolution layer is used for reducing the feature redundancy of the input signal component, and the output end of the convolution layer is connected with the feature extraction network; a gradient turnover layer is arranged between the domain judging network and the characteristic extracting network and is used for automatically inverting during counter propagation, so that the domain judging device loss, the label classifying loss and the L CORAL The loss gradient is the same;
the pretraining method of the CDAN network comprises the following steps:
acquiring bearing vibration signals under various working conditions;
normalizing the obtained bearing vibration signal;
acquiring source domain data with fault classification labels and target domain data without fault classification labels based on the bearing vibration signals after normalization processing;
performing multi-layer wavelet packet decomposition processing on the source domain data and the target domain data to obtain signal components of the data in different frequency domains;
inputting the signal components of the source domain data and the signal components of the target domain data into the feature extraction network respectively, performing multi-feature extraction, weighting key features through an attention module, and restricting the distribution of the source domain features and the target domain features in a high-dimensional space through a correlation alignment algorithm module to obtain the source domain features and the target domain features;
sending the extracted source domain features into a label classification network for fault classification, and calculating a cross entropy loss function to obtain label classification loss; the source domain features and the target domain features are sent into a domain judging network together, and the second-order statistical feature distance loss and domain judging device loss function of the source domain and the target domain are calculated;
and training a bearing fault classification model by minimizing the label classification loss, the second-order statistical feature distance loss of the source domain and the target domain and maximizing the domain discriminator loss, and the back propagation optimization feature extraction network, the label classification network and the domain discrimination network.
Example 3
The present embodiment describes a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the bearing failure diagnosis method as described in embodiment 1.
It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The embodiments of the present invention have been described above with reference to the accompanying drawings, but the present invention is not limited to the above-described embodiments, which are merely illustrative and not restrictive, and many forms may be made by those having ordinary skill in the art without departing from the spirit of the present invention and the scope of the claims, which are all within the protection of the present invention.

Claims (10)

1. A bearing fault diagnosis method is characterized by comprising the following steps:
acquiring bearing vibration signal data;
carrying out multi-layer wavelet packet decomposition processing on the obtained bearing vibration signal data to obtain signal components of the bearing vibration signal in different frequency domains;
inputting the signal component of the bearing vibration signal data into a pre-trained bearing fault classification model to obtain a bearing fault classification result;
the bearing fault classification model comprises a feature extraction network, a label classification network and a domain discrimination network, wherein the feature extraction network comprises a convolution layer, an attention module and a correlation alignment algorithm module, the convolution layer is provided with a plurality of convolution kernels and is used for extracting multi-scale features of the signal components, and the attention module is used for carrying out weighting processing on the extracted key features; the correlation alignment algorithm module is used for restricting the distribution of the source domain and target domain characteristics in a high-dimensional space.
2. The method of claim 1, wherein the bearing fault classification model further comprises a convolution layer having a large convolution kernel for reducing redundancy of features of the input signal component, and an output of the convolution layer is connected to the feature extraction network;
and/or the bearing fault classification model adopts a CDAN network, and the full connection layer of the domain discrimination network is replaced by a convolution layer;
and/or the feature extraction network comprises four convolution kernels of 3×1, 5×1, 7×1, 9×1.
3. The bearing failure diagnosis method according to claim 1 or 2, characterized in that the pre-training method of the bearing failure classification model comprises:
acquiring bearing vibration signals under various working conditions;
normalizing the obtained bearing vibration signal;
acquiring source domain data with fault classification labels and target domain data without fault classification labels based on the bearing vibration signals after normalization processing;
performing multi-layer wavelet packet decomposition processing on the source domain data and the target domain data to obtain signal components of the data in different frequency domains;
inputting the signal components of the source domain data and the signal components of the target domain data into the feature extraction network respectively, performing multi-feature extraction, weighting key features through an attention module, and restricting the distribution of the source domain features and the target domain features in a high-dimensional space through a correlation alignment algorithm module to obtain the source domain features and the target domain features;
sending the extracted source domain features into a label classification network for fault classification, and calculating a cross entropy loss function to obtain label classification loss; the source domain features and the target domain features are sent into a domain judging network together, and the second-order statistical feature distance loss and domain judging device loss function of the source domain and the target domain are calculated;
and training a bearing fault classification model by minimizing the label classification loss, the second-order statistical feature distance loss of the source domain and the target domain and maximizing the domain discriminator loss, and the back propagation optimization feature extraction network, the label classification network and the domain discrimination network.
4. The bearing failure diagnosis method according to claim 1, wherein the processing of the acquired bearing vibration signal data by multi-layer wavelet packet decomposition is: and performing four-layer wavelet packet decomposition processing on the obtained bearing vibration signal data.
5. A bearing failure diagnosis method according to claim 3, wherein the second order statistical feature distance loss of the source domain and target domain is expressed as:
Figure FDA0004100899200000021
in the formula. F Represents the Frobenius norm, D s Representing source domain data, D t Representing target domain data, d representing sample dimensions, C s Representing a source domain covariance matrix, C t Representing the target domain covariance matrix.
6. A bearing failure diagnosis method according to claim 3, wherein the minimized tag classification loss is expressed as:
Figure FDA0004100899200000022
the maximum domain arbiter loss is expressed as:
Figure FDA0004100899200000023
in θ f A representative feature extraction network parameter vector,
Figure FDA0004100899200000024
representing the optimal solution of the feature extraction network parameter vector at saddle points, θ y Representing mapping parameters->
Figure FDA0004100899200000025
Representing the optimal solution of the mapping parameters at saddle points, < >>
Figure FDA0004100899200000026
Representing domain arbiter loss, θ d Representing domain arbiter network parameters, argminE represents the minimization of the loss function.
7. A bearing fault diagnosis method according to claim 3, wherein a gradient inversion layer is provided between the domain discrimination network and the feature extraction network for automatically inverting upon counter-propagation, so that domain discrimination loss, label classification loss and L CORAL The loss gradient is the same;
the loss function of the bearing failure classification model network is expressed as:
Figure FDA0004100899200000031
wherein N represents a positive integer, d i Domain tag representing ith training sample, L y Representing tag predictive loss, G y Representing a tag pre-classifier function, G f Representing a feature extraction network function, x i Training samples representing source and target domains, y i Representing a label sample corresponding to a source domain, lambda represents a meta-parameter, L d Representing domain discrimination loss, G d Representing domain classification network functions, R λ Representing the forward propagation function, θ f Representing feature extraction network parameter vector, θ d Representing domain arbiter network parameters, L CORAL Representing the second order statistical feature distance loss of the source domain and the target domain.
8. A bearing failure diagnosis apparatus, comprising:
the to-be-diagnosed data acquisition module is configured to acquire bearing vibration signal data;
the signal processing module is configured to perform multi-layer wavelet packet decomposition processing on the acquired bearing vibration signal data to obtain signal components of the bearing vibration signal in different frequency domains;
the fault classification module is configured to input signal components of the bearing vibration signal data into a pre-trained bearing fault classification model to obtain a bearing fault classification result;
the bearing fault classification model comprises a feature extraction network, a label classification network and a domain discrimination network, wherein the feature extraction network comprises a convolution layer, an attention module and a correlation alignment algorithm module, the convolution layer is provided with a plurality of convolution kernels and is used for extracting multi-scale features of the signal components, and the attention module is used for carrying out weighting processing on the extracted key features; the correlation alignment algorithm module is used for restricting the distribution of the source domain and target domain characteristics in a high-dimensional space.
9. The bearing fault diagnosis apparatus according to claim 8, wherein the signal processing module performs four-layer wavelet packet decomposition processing on the acquired bearing vibration signal data;
the bearing fault classification model adopts a CDAN network, and a full connection layer of a domain discrimination network is replaced by a convolution layer; the CDAN network further includes a convolution layer having a large convolution kernelThe output end of the convolution layer is connected with the feature extraction network; a gradient turnover layer is arranged between the domain judging network and the characteristic extracting network and is used for automatically inverting during counter propagation, so that domain judging device loss, label classifying loss and L CORAL The loss gradient is the same;
the pre-training method of the bearing fault classification model comprises the following steps:
acquiring bearing vibration signals under various working conditions;
normalizing the obtained bearing vibration signal;
acquiring source domain data with fault classification labels and target domain data without fault classification labels based on the bearing vibration signals after normalization processing;
performing multi-layer wavelet packet decomposition processing on the source domain data and the target domain data to obtain signal components of the data in different frequency domains;
inputting the signal components of the source domain data and the signal components of the target domain data into the feature extraction network respectively, performing multi-feature extraction, weighting key features through an attention module, and restricting the distribution of the source domain features and the target domain features in a high-dimensional space through a correlation alignment algorithm module to obtain the source domain features and the target domain features;
sending the extracted source domain features into a label classification network for fault classification, and calculating a cross entropy loss function to obtain label classification loss; the source domain features and the target domain features are sent into a domain judging network together, and the second-order statistical feature distance loss and domain judging device loss function of the source domain and the target domain are calculated;
and training a bearing fault classification model by minimizing the label classification loss, the second-order statistical feature distance loss of the source domain and the target domain and maximizing the domain discriminator loss, and the back propagation optimization feature extraction network, the label classification network and the domain discrimination network.
10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the bearing fault diagnosis method according to any one of claims 1-8.
CN202310176104.7A 2023-02-28 2023-02-28 Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention Pending CN116337448A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310176104.7A CN116337448A (en) 2023-02-28 2023-02-28 Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310176104.7A CN116337448A (en) 2023-02-28 2023-02-28 Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention

Publications (1)

Publication Number Publication Date
CN116337448A true CN116337448A (en) 2023-06-27

Family

ID=86876664

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310176104.7A Pending CN116337448A (en) 2023-02-28 2023-02-28 Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention

Country Status (1)

Country Link
CN (1) CN116337448A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116894190A (en) * 2023-09-11 2023-10-17 江西南昌济生制药有限责任公司 Bearing fault diagnosis method, device, electronic equipment and storage medium
CN116893924A (en) * 2023-09-11 2023-10-17 江西南昌济生制药有限责任公司 Equipment fault processing method, device, electronic equipment and storage medium

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116894190A (en) * 2023-09-11 2023-10-17 江西南昌济生制药有限责任公司 Bearing fault diagnosis method, device, electronic equipment and storage medium
CN116893924A (en) * 2023-09-11 2023-10-17 江西南昌济生制药有限责任公司 Equipment fault processing method, device, electronic equipment and storage medium
CN116894190B (en) * 2023-09-11 2023-11-28 江西南昌济生制药有限责任公司 Bearing fault diagnosis method, device, electronic equipment and storage medium
CN116893924B (en) * 2023-09-11 2023-12-01 江西南昌济生制药有限责任公司 Equipment fault processing method, device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
CN113554089B (en) Image classification countermeasure sample defense method and system and data processing terminal
Lin et al. Deep unknown intent detection with margin loss
Arras et al. Explaining predictions of non-linear classifiers in NLP
CN116337448A (en) Method, device and storage medium for diagnosing faults of transfer learning bearing based on width multi-scale space-time attention
Chen et al. Rethinking data augmentation for robust visual question answering
CN113596007B (en) Vulnerability attack detection method and device based on deep learning
CN107945210B (en) Target tracking method based on deep learning and environment self-adaption
Adhao et al. Feature selection using principal component analysis and genetic algorithm
CN112417206B (en) Weak supervision video time interval retrieval method and system based on two-branch proposed network
CN112235434B (en) DGA network domain name detection and identification system fusing k-means and capsule network thereof
Zhang et al. Cagfuzz: coverage-guided adversarial generative fuzzing testing for image-based deep learning systems
Abeyagunasekera et al. LISA: Enhance the explainability of medical images unifying current XAI techniques
CN114547299A (en) Short text sentiment classification method and device based on composite network model
Ilić et al. Active learning using a self-correcting neural network (ALSCN)
Schneider et al. Detecting anomalies within time series using local neural transformations
Tang et al. Boundary regression model for joint entity and relation extraction
Liu et al. A dual-branch balance saliency model based on discriminative feature for fabric defect detection
CN113779938B (en) System and method for generating coherent stories based on visual and theme cooperative attention
CN111738455B (en) Fault diagnosis method and system based on integration domain self-adaptation
CN112261028B (en) DGA botnet domain name detection method based on capsule network and k-means
Lapertot et al. Supervised learning of hierarchical image segmentation
CN114722400A (en) Side channel vulnerability detection method, system, medium, equipment and terminal
Wen et al. DCN: detector-corrector network against evasion attacks on deep neural networks
CN115309854A (en) Countermeasure sample generation method and device and computer equipment
Qiu et al. Graph Convolution and Self Attention Based Non-maximum Suppression

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