CN116522118A - Fault diagnosis method based on improved unsupervised domain self-adaptive network - Google Patents
Fault diagnosis method based on improved unsupervised domain self-adaptive network Download PDFInfo
- Publication number
- CN116522118A CN116522118A CN202310548727.2A CN202310548727A CN116522118A CN 116522118 A CN116522118 A CN 116522118A CN 202310548727 A CN202310548727 A CN 202310548727A CN 116522118 A CN116522118 A CN 116522118A
- Authority
- CN
- China
- Prior art keywords
- domain
- data
- label
- improved
- network
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000003745 diagnosis Methods 0.000 title claims abstract description 33
- 230000003044 adaptive effect Effects 0.000 claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000013508 migration Methods 0.000 claims abstract description 12
- 230000005012 migration Effects 0.000 claims abstract description 12
- 230000006870 function Effects 0.000 claims description 48
- 239000011159 matrix material Substances 0.000 claims description 25
- 230000004913 activation Effects 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 8
- 238000009826 distribution Methods 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 238000009499 grossing Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 230000001133 acceleration Effects 0.000 claims description 3
- 238000012804 iterative process Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 2
- 230000002708 enhancing effect Effects 0.000 abstract 1
- 238000003062 neural network model Methods 0.000 abstract 1
- 238000013135 deep learning Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 238000012733 comparative method Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/09—Supervised learning
-
- 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
- G06N3/094—Adversarial learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a fault diagnosis method based on an improved unsupervised domain self-adaptive network, which belongs to the field of intelligent fault diagnosis of mechanical systems and comprises the following steps: the method comprises the steps of obtaining an original vibration signal from equipment, converting the original vibration signal into frequency domain data, taking a manually damaged bearing data set as source domain data and taking a truly damaged bearing data set as target domain data according to data types; constructing a deep multi-pair domain adaptive neural network model, inputting the artificial damage bearing data and partial real damage bearing data into a network for training, and learning domain invariant features through domain countermeasure training of a feature extractor and a domain label smooth multi-domain discriminator; maximizing and enhancing the mobility of source domain data and the authenticability of target domain data through batch kernel norms; and inputting the other part of target domain data into a trained multi-contrast domain self-adaptive network to obtain a prediction label of the target domain data, and realizing migration diagnosis from the artificially damaged bearing to the truly damaged bearing.
Description
Technical Field
The invention discloses a fault diagnosis method based on an improved unsupervised domain self-adaptive network, and belongs to the field of intelligent fault diagnosis of mechanical systems.
Background
Intelligent fault diagnosis plays a vital role in current fault prediction and health management systems. In recent decades, with the advent of industrial big data, intelligent fault diagnosis technology has been greatly developed. The performance of the fault diagnosis method is greatly improved with the aid of a computer and a deep learning technology. These deep learning based fault diagnosis methods can learn meaningful features from data through a built depth model.
Although deep learning-based methods have made great progress in mechanical fault diagnosis, these methods still have some limitations. In one aspect, most current deep learning-based methods assume that samples from the source domain and samples from the target domain have the same probability distribution. However, in an actual industrial scenario, this assumption is not true. During operation, the equipment operating conditions change in real time due to equipment wear and environmental noise, which will lead to a distribution difference between the collected data and to performance degradation when training a model trained using the previously collected data to a new task. On the other hand, obtaining enough marked training samples is expensive, and the application of the intelligent fault diagnosis method in the industry is limited.
Domain adaptation is a practical technique to address the above limitations by learning domain invariant features and discriminant features to transfer knowledge learned from a labeled source domain to an unlabeled target domain. However, there are problems in the conventional studies. Most approaches focus mainly on reducing the marginal distribution difference between the source and target domains, while ignoring class information of the data. These methods cannot take advantage of dependencies between properties and classes, which is of great significance for capturing multi-modal structures of underlying distributions of data.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a fault diagnosis method based on an improved unsupervised domain self-adaptive network, which extracts domain invariant features through countermeasure training of a feature extractor and a label smoothing multi-domain discriminator; performing kernel norm minimization maximization on the batch output matrixes of the tag classifier on the source domain and the target domain respectively, so that the source domain features have mobility, and meanwhile, the diversity and the authenticability of the target domain features are enhanced; and finally, the migration diagnosis from the manual damage bearing fault data to the real damage bearing fault data is realized.
1. A fault diagnosis method based on an improved unsupervised domain adaptive network is characterized in that: the method comprises the following steps:
step 1, acquiring an original vibration signal from equipment through an acceleration sensor, converting the original vibration signal into frequency domain data, taking artificial damage bearing data as labeled source domain sample data, and taking real damage bearing data as target domain sample data which is unavailable to labels.
Step 2, constructing an improved unsupervised domain adaptive network, training the unsupervised domain adaptive network by taking source domain data and partial target domain data as inputs, wherein the unsupervised domain adaptive network comprises:
feature extractor G f The feature extractor employs lifted convolutionThe neural network comprises five layers of convolution layers, an activation function is ELU, and the feature extractor is used for extracting features of a source domain sample and a target domain sample;
label classifier G y The label classifier is a double-layer full-connection network, the first layer activation function is a ReLU, the second layer classification function is a Softmax, and the label classifier is used for performing fault classification on a source domain sample and a target domain sample.
Label smoothed multi-domain discriminatorThe domain identifier of the domain label smoothing is composed of K class domain identifiersThe source domain data and the target domain data are distributed on a multi-mode structure to be matched;
the batch core norm is minimized, wherein the batch core norm is minimized, namely, the batch core norm is maximized for the target domain batch output matrix, so that the diversity and the identifiability of the target domain characteristics are improved, the batch core norm is minimized for the source domain batch output matrix, and the mobility of the source domain characteristics is improved;
and step 3, inputting another part of target domain samples into a trained improved unsupervised domain self-adaptive network to obtain a prediction label of the target domain samples, and realizing the migration fault diagnosis from the artificial damaged bearing to the real damaged bearing.
2. The method for fault diagnosis of an improved unsupervised domain adaptive network according to claim 1, wherein in step 1, the label-available source domain samples are expressed asThe target domain sample for which the tag is not available is denoted +.>
Wherein the method comprises the steps ofIs a labeled source domain sample, n s Is the sourceDomain number of samples, ++>And->Respectively representing an ith source domain sample and a corresponding label; />Is a target domain sample without label, n t For the number of target field samples, < > and->Represented as the jth target domain sample; />And->With the same tag space, the machine is +.>And->Obeying the edge probability distributions P (X) and Q (X), respectively.
3. The method for fault diagnosis of improved unsupervised domain adaptive network according to claim 1, wherein in step 2, the feature extractor uses a lifted deep neural network as the feature extractor, firstly uses a wide-kernel convolutional layer in the first layer to extract features, and then uses a continuous small-kernel convolutional layer to deepen the network to obtain a better feature representation; secondly, the expression capability and the feature learning capability of the network are enhanced by adopting an ELU activation function after convolution operation, and the expression of the ELU is as follows:
where x represents the input characteristic and α is a positive hyper-parameter for controlling the saturation of the negative input value.
4. The method of claim 1, wherein in step 2, the batch output matrix kernel norm is minimized, and the method further comprises the step of, for the source domain batch output matrix G (x s ) Execution core norm G (x) s )|| ★ Minimizing, improving the mobility of source domain features, and batch outputting the matrix G (x t ) Execution core norm G (x) t )|| ★ Maximizing, improving diversity and discriminability of target domain features, wherein I ★ Is the core norm of the matrix.
5. The method for fault diagnosis of an improved unsupervised domain adaptive network according to claim 1, wherein in step 2, training and testing the network comprises the steps of:
1) Source domain data to be taggedAnd partially unlabeled target domain data +.>Input the feature extractor G f Extracting depth features to obtain output features expressed as G f (x i );
2) Output G of the feature extractor f (x i ) Input depth discrimination tag classifier G y Obtaining a K-dimensional vectorWherein->Representing the probability that the sample belongs to the fault class k, the loss function of the label classifier is expressed in terms of a cross entropy loss function as:
wherein I is an index function;
3) Constructing individual classes of domain discriminators in the multi-domain discriminatorThe goal is to resolve the feature representation G f (x i ) Whether from the source domain or the target domain, based on the thought of attention, will +.>Representing discriminator as probability->The corresponding data should be given a weight of how much attention, and therefore with probability +.>The weighted features represent G f (x i ) As category discriminator->Obtaining a predicted value of each type of data field tag, expressed as 1 or 0;
after obtaining the domain label, setting an error rate epsilon for the training set based on the two-class cross entropy loss function, and using the probability of 1-epsilon to make G f (x i ) In the iterative process, the influence caused by label errors is reduced, the convergence process is accelerated, and the loss of the multi-domain discriminator described by label smooth cross entropy is as follows:
wherein d i Data point x i Domain tag of d' i =(1-ε)d i +ε(1-d i );
4) For a size of B t Is output as G (x) t ) The loss function of the core norm maximization is expressed as:
likewise, batchOutput matrix G (x s ) In the size of B s The source domain sample of the tag classifier is obtained after the full connection layer of the tag classifier, and the migration performance of the source domain feature is improved by adopting a minimized kernel norm, and a loss function is expressed as follows:
further, the loss function of the batch output matrix kernel norm minimization is:
5) Constructing an objective function combining the fault classification loss, domain classification loss and batch kernel norm minimization, the objective function expressed as:
wherein the method comprises the steps ofLambda and eta are trade-off parameters for the total loss function;
6) Updating parameters of the improved unsupervised domain adaptive network according to the objective function, training the network by setting different super parameter optimization strategies for different modules, and repeatedly executing the steps 1) to 5) until the specified training times are reached, wherein updating the parameters of the unsupervised domain adaptive network is performed by the following formula
Wherein mu 1 、μ 2 Sum mu 3 The different learning rates are indicated to be different,representing the differential operator, θ f 、θ y And theta d Respectively represent the feature extractor and the label classifierAnd a network parameter of a label-smoothed multi-domain discriminator;
7) And testing the trained improved unsupervised domain self-adaptive network by using another part of target domain samples, and outputting a target domain sample fault prediction label.
Compared with the prior art, the invention has the beneficial effects that: the invention provides a domain label smooth multi-contrast domain self-adaptive training method, which considers the multi-modal structure of data and fully distributes Ji Yuanyu and target domain data; introducing maximum minimization of batch kernel norms, so that the source domain features are more migratable while the target domain features have diversity and authenticability; a training strategy for setting different super parameters for different modules is used for independently optimizing a main network and a branch network, so that under fitting and over fitting are avoided.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention.
Fig. 2 is a diagram of an improved unsupervised domain adaptive network architecture according to the present invention.
FIG. 3 is a graph of training process accuracy for the present invention and the comparative method.
FIG. 4 is a graph of t-SNE feature dimension reduction visualization of the present invention and comparative method on migration tasks, (a) present invention, (b) ENT, (c) MADA, (d) DANN.
Detailed Description
The following describes the embodiments of the present invention further with reference to the drawings and detailed description.
1. A fault diagnosis method based on an improved unsupervised domain adaptive network is characterized in that: the method comprises the following steps:
step 1, acquiring an original vibration signal from equipment through an acceleration sensor, converting the original vibration signal into frequency domain data, taking artificial damage bearing data as labeled source domain sample data, and taking real damage bearing data as target domain sample data which is unavailable to labels.
Step 2, constructing an improved unsupervised domain adaptive network, training the unsupervised domain adaptive network by taking source domain data and partial target domain data as inputs, wherein the unsupervised domain adaptive network comprises:
feature extractor G f The feature extractor adopts a lifted convolutional neural network, comprises five layers of convolutional layers, has an ELU as an activation function, and is used for extracting features of a source domain sample and a target domain sample;
label classifier G y The label classifier is a double-layer full-connection network, the first layer activation function is a ReLU, the second layer classification function is a Softmax, and the label classifier is used for performing fault classification on a source domain sample and a target domain sample.
Label smoothed multi-domain discriminatorThe domain identifier of the domain label smoothing is composed of K class domain identifiersThe source domain data and the target domain data are distributed on a multi-mode structure to be matched;
the batch core norm is minimized, wherein the batch core norm is minimized, namely, the batch core norm is maximized for the target domain batch output matrix, so that the diversity and the identifiability of the target domain characteristics are improved, the batch core norm is minimized for the source domain batch output matrix, and the mobility of the source domain characteristics is improved;
and step 3, inputting another part of target domain samples into a trained improved unsupervised domain self-adaptive network to obtain a prediction label of the target domain samples, and realizing the migration fault diagnosis from the artificial damaged bearing to the real damaged bearing.
2. The method for fault diagnosis of an improved unsupervised domain adaptive network according to claim 1, wherein in step 1, the label-available source domain samples are expressed asThe target domain sample for which the tag is not available is denoted +.>
Wherein the method comprises the steps ofIs a labeled source domain sample, n s For the number of source field samples, < > and->And->Respectively representing an ith source domain sample and a corresponding label; />Is a target domain sample without label, n t For the number of target field samples, < > and->Represented as the jth target domain sample;and->Having the same label space, the machine will experience different environmental and operating conditions during its service life, and thereforeAnd->Obeying the edge probability distributions P (X) and Q (X), respectively.
3. The method for fault diagnosis of improved unsupervised domain adaptive network according to claim 1, wherein in step 2, the feature extractor uses a lifted deep neural network as the feature extractor, firstly uses a wide-kernel convolutional layer in the first layer to extract features, and then uses a continuous small-kernel convolutional layer to deepen the network to obtain a better feature representation; secondly, the expression capability and the feature learning capability of the network are enhanced by adopting an ELU activation function after convolution operation, and the expression of the ELU is as follows:
where x represents the input characteristic and α is a positive hyper-parameter for controlling the saturation of the negative input value.
4. The method of claim 1, wherein in step 2, the batch output matrix kernel norm is minimized, and the method further comprises the step of, for the source domain batch output matrix G (x s ) Execution core norm G (x) s )|| ★ Minimizing, improving the mobility of source domain features, and batch outputting the matrix G (x t ) Execution core norm G (x) t )|| ★ Maximizing, improving diversity and discriminability of target domain features, wherein I ★ Is the core norm of the matrix.
5. The method for fault diagnosis of an improved unsupervised domain adaptive network according to claim 1, wherein in step 2, training and testing the network comprises the steps of:
1) Source domain data to be taggedAnd partially unlabeled target domain data +.>Input the feature extractor G f Extracting depth features to obtain output features expressed as G f (x i );
2) Output G of the feature extractor f (x i ) Input depth discrimination tag classifier G y Obtaining a K-dimensional vectorWherein->Representing the probability that the sample belongs to the fault class k, the loss function of the label classifier is expressed in terms of a cross entropy loss function as:
wherein I is an index function;
3) Constructing individual classes of domain discriminators in the multi-domain discriminatorThe goal is to resolve the feature representation G f (x i ) Whether from the source domain or the target domain, based on the thought of attention, will +.>Representing discriminator as probability->The corresponding data should be given a weight of how much attention, and therefore with probability +.>The weighted features represent G f (x i ) As category discriminator->Obtaining a predicted value of each type of data field tag, expressed as 1 or 0;
after obtaining the domain label, setting an error rate epsilon for the training set based on the two-class cross entropy loss function, and using the probability of 1-epsilon to make G f (x i ) In the iterative process, the influence caused by label errors is reduced, the convergence process is accelerated, and the loss of the multi-domain discriminator described by label smooth cross entropy is as follows:
wherein d i Data point x i Domain tag of d' i =(1-ε)d i +ε(1-d i );
4) For a size of B t Is output as G (x) t ) The loss function of the core norm maximization is expressed as:
likewise, a batch output matrix G (x s ) In the size of B s The source domain sample of the tag classifier is obtained after the full connection layer of the tag classifier, and the migration performance of the source domain feature is improved by adopting a minimized kernel norm, and a loss function is expressed as follows:
further, the loss function of the batch output matrix kernel norm minimization is:
5) Constructing an objective function combining the fault classification loss, domain classification loss and batch kernel norm minimization, the objective function expressed as:
wherein the method comprises the steps ofLambda and eta are trade-off parameters for the total loss function;
6) Updating parameters of the improved unsupervised domain adaptive network according to the objective function, training the network by setting different super parameter optimization strategies for different modules, and repeatedly executing the steps 1) to 5) until the specified training times are reached, wherein updating the parameters of the unsupervised domain adaptive network is performed by the following formula
Wherein mu 1 、μ 2 Sum mu 3 The different learning rates are indicated to be different,representing microDivision operator, θ f 、θ y And theta d Network parameters respectively representing a feature extractor, a tag classifier and a tag smoothing multi-domain discriminator;
7) And testing the trained improved unsupervised domain self-adaptive network by using another part of target domain samples, and outputting a target domain sample fault prediction label.
A specific application example procedure is given below while verifying the validity of the present invention in engineering applications.
The method for identifying the true damage bearing fault is implemented according to the following specific steps, and the implementation flow is shown in figure 1:
1) Construction of data sets
The experiment is carried out by using a data set of the university of Pade Boen, germany, the damage state of the bearing comprises faults of an inner ring and an outer ring, the bearing is in healthy operation and respectively corresponds to data labels 0, 1 and 2, the bearing faults are respectively caused by artificial damage and real damage, the artificial damage mainly causes bearing pitting corrosion through motor engraving, the real damaged bearing is obtained through an accelerated life test bed, the main purpose of the experiment is to identify the real damaged bearing through an artificial damaged bearing fault data training model, and 100 source domain samples and 80 target domain samples are obtained through the step 1.
2) Construction of network model
According to said step 2, a network is constructed as shown in fig. 2.
3) Training of models
And 3, carrying out migration diagnosis of the artificial damaged bearing to the real damaged bearing under the same damage degree by using the rotating speed of 1500rpm, training the model by using the source domain sample and part of the target domain sample, setting the iteration number to be 80, setting the batch training size to be 128, continuously converging model parameters in the iteration process until the process is ended after 80 times, and converging the model total loss function at the moment.
To verify the effectiveness of the present invention, a comparison was made between MADA networks having the same basic network parameters as the present invention and DANN networks using a cross entropy function as the loss function, with the exception of the network structure network as in the present experiment, but with the substitution of the maximum of the core norms of the batch output matrix to the minimum of Entropy (ENT), and multiple sets of migration tasks were tested. The model training process accuracy is shown in figure 3.
3) Model testing
The method for diagnosing the bearing faults based on the depth multi-pair domain self-adaption is superior to other traditional methods and similar methods, can extract discriminant features, can finish detection of fault modes, can extract domain invariant features and can realize migration of fault diagnosis knowledge.
Claims (5)
1. A fault diagnosis method based on an improved unsupervised domain adaptive network is characterized in that: the method comprises the following steps:
step 1, acquiring an original vibration signal from equipment through an acceleration sensor, converting the original vibration signal into frequency domain data, taking artificial damage bearing data as labeled source domain sample data, and taking real damage bearing data as target domain sample data which is unavailable to labels.
Step 2, constructing an improved unsupervised domain adaptive network, training the unsupervised domain adaptive network by taking source domain data and partial target domain data as inputs, wherein the unsupervised domain adaptive network comprises:
feature extractor G f The feature extractor adopts a lifted convolutional neural network, comprises five layers of convolutional layers, has an ELU as an activation function, and is used for extracting features of a source domain sample and a target domain sample;
label classifier G y The label classifier is a double-layer full-connection network, the first layer activation function is a ReLU, the second layer classification function is a Softmax, and the label classifier is used for performing fault classification on a source domain sample and a target domain sample;
label smoothed multi-domain discriminatorThe domain identifier of the domain label smoothing is composed of K classesDomain discriminatorThe source domain data and the target domain data are distributed on a multi-mode structure to be matched;
the batch core norm is minimized, wherein the batch core norm is minimized, namely, the batch core norm is maximized for the target domain batch output matrix, so that the diversity and the identifiability of the target domain characteristics are improved, the batch core norm is minimized for the source domain batch output matrix, and the mobility of the source domain characteristics is improved;
and step 3, inputting another part of target domain samples into a trained improved unsupervised domain self-adaptive network to obtain a prediction label of the target domain samples, and realizing the migration fault diagnosis from the artificial damaged bearing to the real damaged bearing.
2. The method for fault diagnosis of an improved unsupervised domain adaptive network according to claim 1, wherein in step 1, the label-available source domain samples are expressed asThe target domain sample for which the tag is not available is denoted +.>
Wherein the method comprises the steps ofIs a labeled source domain sample, n s For the number of source field samples, < > and->And->Respectively representing an ith source domain sample and a corresponding label; />Is a sample of the target domain without a tag,n t for the number of target field samples, < > and->Represented as the jth target domain sample; />And->With the same tag space, the machine is +.>And->Obeying the edge probability distributions P (X) and Q (X), respectively.
3. The method for fault diagnosis of improved unsupervised domain adaptive network according to claim 1, wherein in step 2, the feature extractor uses a lifted deep neural network as the feature extractor, firstly uses a wide-kernel convolutional layer in the first layer to extract features, and then uses a continuous small-kernel convolutional layer to deepen the network to obtain a better feature representation; secondly, the expression capability and the feature learning capability of the network are enhanced by adopting an ELU activation function after convolution operation, and the expression of the ELU is as follows:
where x represents the input characteristic and is a positive hyper-parameter for controlling the saturation of the negative input value.
4. A method of fault diagnosis for an improved unsupervised domain adaptive network as claimed in claim 3, characterized in that in step 2, the batch output matrix kernel norm is maximized for the source domain batch output matrix G (x s ) ExecutingLine core norms ||G (x s )‖ ★ Minimizing, improving the mobility of source domain features, and batch outputting the matrix G (x t ) Execution core norm ||G (x t )‖ ★ Maximizing, increasing diversity and authenticability of target domain features, wherein II ★ Is the core norm of the matrix.
5. The method for fault diagnosis of an improved unsupervised domain adaptive network according to claim 4, wherein in step 2, training and testing the network comprises the steps of:
1) Source domain data to be taggedAnd partially unlabeled target domain data +.>Input the feature extractor G f Extracting depth features to obtain output features expressed as G f (x i );
2) Output G of the feature extractor f (x i ) Input depth discrimination tag classifier G y Obtaining a K-dimensional vectorWherein->Representing the probability that the sample belongs to the fault class k, the loss function of the label classifier is expressed in terms of a cross entropy loss function as:
wherein I is an index function;
3) Constructing individual classes of domain discriminators in the multi-domain discriminatorThe goal is to resolve the feature representation G f (x i ) Whether from the source domain or the target domain, based on the thought of attention, will +.>Representing discriminator as probability->The corresponding data should be given a weight of how much attention, and therefore with probability +.>The weighted features represent G f (x i ) As category discriminator->Obtaining a predicted value of each type of data field tag, expressed as 1 or 0;
after obtaining the domain label, setting an error rate epsilon for the training set based on the two-class cross entropy loss function, and using the probability of 1-epsilon to make G f (x i ) In the iterative process, the influence caused by label errors is reduced, the convergence process is accelerated, and the loss of the multi-domain discriminator described by label smooth cross entropy is as follows:
wherein d i Data point x i Domain tag of d' i =(1-ε)d i +ε(1-d i );
4) For a size of B t Is output as G (x) t ) The loss function of the core norm maximization is expressed as:
likewise, a batch output matrix G (x s ) In the size of B s The source domain sample of the tag classifier is obtained after the full connection layer of the tag classifier, and the migration performance of the source domain feature is improved by adopting a minimized kernel norm, and a loss function is expressed as follows:
further, the loss function of the batch output matrix kernel norm minimization is:
5) Constructing an objective function combining the fault classification loss, domain classification loss and batch kernel norm minimization, the objective function expressed as:
wherein the method comprises the steps ofLambda and eta are trade-off parameters for the total loss function;
6) Updating parameters of the improved unsupervised domain adaptive network according to the objective function, training the network by setting different super parameter optimization strategies for different modules, and repeatedly executing the steps 1) to 5) until the specified training times are reached, wherein updating the parameters of the unsupervised domain adaptive network is performed by the following formula
Wherein mu 1 、μ 2 Sum mu 3 The different learning rates are indicated to be different,representing the differential operator, θ f 、θ t And theta d Network parameters respectively representing a feature extractor, a tag classifier and a tag smoothing multi-domain discriminator;
7) And testing the trained improved unsupervised domain self-adaptive network by using another part of target domain samples, and outputting a target domain sample fault prediction label.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310548727.2A CN116522118A (en) | 2023-05-16 | 2023-05-16 | Fault diagnosis method based on improved unsupervised domain self-adaptive network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310548727.2A CN116522118A (en) | 2023-05-16 | 2023-05-16 | Fault diagnosis method based on improved unsupervised domain self-adaptive network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116522118A true CN116522118A (en) | 2023-08-01 |
Family
ID=87404533
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310548727.2A Pending CN116522118A (en) | 2023-05-16 | 2023-05-16 | Fault diagnosis method based on improved unsupervised domain self-adaptive network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116522118A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117009794A (en) * | 2023-09-27 | 2023-11-07 | 山东大学 | Machine fault diagnosis method and system based on unsupervised subdomain self-adaption |
CN117370851A (en) * | 2023-08-31 | 2024-01-09 | 西南交通大学 | Self-adaptive input length bearing fault diagnosis method based on unsupervised transfer learning |
-
2023
- 2023-05-16 CN CN202310548727.2A patent/CN116522118A/en active Pending
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117370851A (en) * | 2023-08-31 | 2024-01-09 | 西南交通大学 | Self-adaptive input length bearing fault diagnosis method based on unsupervised transfer learning |
CN117370851B (en) * | 2023-08-31 | 2024-04-16 | 西南交通大学 | Self-adaptive input length bearing fault diagnosis method based on unsupervised transfer learning |
CN117009794A (en) * | 2023-09-27 | 2023-11-07 | 山东大学 | Machine fault diagnosis method and system based on unsupervised subdomain self-adaption |
CN117009794B (en) * | 2023-09-27 | 2023-12-15 | 山东大学 | Machine fault diagnosis method and system based on unsupervised subdomain self-adaption |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN116522118A (en) | Fault diagnosis method based on improved unsupervised domain self-adaptive network | |
CN109186973B (en) | Mechanical fault diagnosis method of unsupervised deep learning network | |
CN112883994B (en) | Rotating machinery variable working condition fault diagnosis method with balanced distribution adaptation | |
CN110647830B (en) | Bearing fault diagnosis method based on convolutional neural network and Gaussian mixture model | |
CN114048568B (en) | Rotary machine fault diagnosis method based on multisource migration fusion shrinkage framework | |
CN113469219B (en) | Rotary machine fault diagnosis method under complex working condition based on element transfer learning | |
CN115127814A (en) | Unsupervised bearing fault diagnosis method based on self-adaptive residual error countermeasure network | |
CN115510965A (en) | Bearing imbalance fault diagnosis method based on generated data fusion | |
CN115600150A (en) | Multi-mode gearbox fault diagnosis method based on deep migration learning | |
CN111222442A (en) | Electromagnetic signal classification method and device | |
CN116451150A (en) | Equipment fault diagnosis method based on semi-supervised small sample | |
CN111931814A (en) | Unsupervised anti-domain adaptation method based on intra-class structure compactness constraint | |
CN115165366A (en) | Variable working condition fault diagnosis method and system for rotary machine | |
CN110363214A (en) | A kind of contact condition recognition methods of the robotic asssembly based on GWA-SVM | |
CN116186641A (en) | Cross-equipment rolling bearing small sample fault diagnosis method based on feature fusion | |
CN111582373A (en) | Radiation source identification method based on weighted migration extreme learning machine algorithm | |
CN115791179A (en) | Self-supervision bearing fault diagnosis method based on clustering algorithm | |
CN115358337A (en) | Small sample fault diagnosis method and device and storage medium | |
CN112001115B (en) | Soft measurement modeling method of semi-supervised dynamic soft measurement network | |
CN113537306A (en) | Image classification method based on progressive growth element learning | |
CN112541966A (en) | Face replacement method based on reconstruction and network generation | |
CN115791174B (en) | Rolling bearing abnormality diagnosis method, system, electronic equipment and storage medium | |
CN116089812A (en) | Fault diagnosis method based on semi-supervised contrast domain generalization intelligent model | |
CN116026590A (en) | Self-supervision bearing fault diagnosis method | |
Cui et al. | Few-shot Learning for Rolling Bearing Fault Diagnosis Based on Residual Convolutional Neural Network |
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 |