CN115099270A - Bearing fault diagnosis method under variable load based on sub-domain adaptive countermeasure network - Google Patents
Bearing fault diagnosis method under variable load based on sub-domain adaptive countermeasure network Download PDFInfo
- Publication number
- CN115099270A CN115099270A CN202210684731.7A CN202210684731A CN115099270A CN 115099270 A CN115099270 A CN 115099270A CN 202210684731 A CN202210684731 A CN 202210684731A CN 115099270 A CN115099270 A CN 115099270A
- Authority
- CN
- China
- Prior art keywords
- domain
- sample
- target
- target domain
- source domain
- 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
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
Landscapes
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- General Physics & Mathematics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a bearing fault diagnosis method under variable load based on a sub-domain adaptive countermeasure network, and belongs to the field of rotary machine fault diagnosis. The invention utilizes the one-dimensional convolution subdomain adaptive countermeasure network model to extract domain invariant features, establishes an effective cross-load rolling bearing fault diagnosis model, realizes unsupervised fault diagnosis on a target domain, relieves the problem of rare fault data marking information in practical application, and improves the effectiveness and accuracy of rolling bearing fault diagnosis, thereby ensuring the safety and reliability of modern rotating machinery operation and the rapidness and accuracy of equipment component supply.
Description
Technical Field
The invention belongs to the field of fault diagnosis of rotary machines, and particularly relates to a bearing fault diagnosis method under variable load based on a subdomain adaptive countermeasure network.
Background
Rolling bearings are common parts of rotating machinery, and are prone to wear and failure when working in a severe environment for a long time. The rolling bearing fault identification is researched, the fault position and the fault degree can be found in time, the safe and stable operation of equipment is guaranteed, and meanwhile, clear and quick requirements are provided for manufacturing enterprises.
The rolling bearing operating state information is often reflected by its vibration signal, and with the development of data-driven failure diagnosis methods, research based on the vibration signal of the rolling bearing has become a hot spot. Diagnosis based on the traditional signal processing method has been studied in a large quantity, and vibration data is processed by utilizing a time-frequency domain analysis technology so as to extract weak signals capable of reflecting faults. Machine learning has been widely applied in the field of fault diagnosis, the traditional machine learning method is mainly established on the basis of manually selecting and constructing vibration signal feature vectors, a diagnosis model is trained by using the selected features, the model mainly comprises a support vector machine, K neighbor, a decision tree and the like, and the diagnosis effect of the method is highly related to the feature vector constitution from the aspect of action mechanism. The deep learning method can establish a vibration signal end-to-end model, autonomously learns characteristics to identify a health state, is widely applied to a Convolutional Neural Network (CNN) in deep learning, has the characteristics of parameter sharing, sparse connection, translation invariance and the like, and is small in data processing amount, low in training difficulty and remarkable in effect. Meanwhile, based on the one-dimensional representation of the vibration signal, the one-dimensional CNN is more suitable for the application scene of the invention.
The deep learning method requires that the training sample and the test sample have the same feature space and probability distribution, however, in an actual fault diagnosis scene, due to the working environment and load change of the rotary machine, the training data and the test data often have distribution difference; and after the fault occurs, the labeled data is difficult to obtain, and the fault data under certain operating loads is unsupervised. This reduces the model diagnosis effect, and the model generalization capability cannot be ensured, which brings a challenge to the fault diagnosis.
Disclosure of Invention
The invention aims to provide a variable-load bearing fault diagnosis method based on a subdomain adaptive countermeasure network, aiming at the defects of the prior art. Firstly, collecting bearing vibration signals under different operation loads as source domain and target domain data, and performing Variational Modal Decomposition (VMD) on the vibration signals as signal pretreatment aiming at the characteristics of the vibration signals; secondly, inputting the VMD source domain and target domain data into a one-dimensional convolution subdomain adaptive countermeasure network, shortening the characteristic space distance of the source domain and the target domain, extracting inter-domain invariant characteristics, realizing subdomain characteristic alignment between the source domain and the target domain, and realizing knowledge transfer from a supervised source domain to an unsupervised target domain; and finally, carrying out fault diagnosis on the unsupervised signal of the target domain by using the trained model.
The purpose of the invention is realized by the following technical scheme: a bearing fault diagnosis method under variable load based on a subdomain adaptive countermeasure network comprises the following steps:
(1) collecting vibration signals of a rolling bearing under different operating loads in a multi-health state to serve as a bearing fault vibration signal database, taking the vibration signal under the load with known fault information as a source domain signal, and taking the vibration signal under the load with unknown fault information as a target domain signal;
(2) carrying out non-overlapping equal-length division on the multiple health state signals of the source domain and the target domain;
(3) respectively carrying out variation modal decomposition on the divided source domain signal and target domain signal, and forming a source domain data set D by the intrinsic modal function group obtained by decomposition S And a target domain data set D T ；
(4) Establishing a fault diagnosis model based on a one-dimensional convolution subdomain adaptive countermeasure network; the model includes a feature extractor G F Label sorter G Y Sum domain discriminator G D (ii) a Feature extractor G F The method is used for implicitly extracting data features and adopts a one-dimensional convolution neural network; label classifier G Y For use inSource domain data is classified, connected with a feature extractor G F (ii) a Domain discriminator G D For discriminating whether the data is from the source domain or the target domain, the feature extractor G is also connected F (ii) a Feature extractor G F And domain discriminator G D A gradient inversion layer is arranged between the two layers;
(5) inputting a source domain signal with known fault information and a target domain signal with unknown fault information into a one-dimensional convolution subdomain adaptation countermeasure network for model training, and fixing model parameters;
(6) in the online diagnosis process, a vibration signal under the load of a section of target domain is collected, variation modal decomposition is carried out, the vibration signal is input into a trained model, and a fault diagnosis result is output.
Further, the state-of-health vibration signals in the step (1) comprise vibration acceleration signals of a plurality of states of health of the bearing, such as normal state, inner ring fault, outer ring fault, rolling body fault and different degradation degrees.
Further, in the step (3), performing variational modal decomposition on the divided source domain signal and target domain signal respectively, setting VMD parameters, and taking the same value for the decomposition number when performing VMD on signals in different health states in order to ensure the consistent data dimensionality; forming a source domain data set D by using the Intrinsic Mode Functions (IMFs) obtained by decomposition as samples S With the target domain data set D T ：
Wherein
And
respectively representing the ith source domain sample and the target domain sample,
sample data tag, n, representing the source domain s And n t Respectively representing the number of samples in a source domain and a target domain; the feature space of both domains is the same as the class space, i.e. X S ＝X T And Y S ＝Y T But the two domain data distributions are different, namely P (X) S )≠P(X T )。
Further, in the step (4), establishing a fault diagnosis model based on the one-dimensional convolution subdomain adaptation countermeasure network specifically includes:
(4-1) the feature extractor is composed of three convolution pooling modules, the first two pooling layers adopt maximum pooling operation, and the last pooling layer adopts average pooling; in order to prevent the network from being over-fitted, a dropout layer is introduced; the source domain samples and the target domain samples share a feature extractor;
(4-2) the label classifier comprises two full-connection layers, the number of neurons in the last layer is the same as the number of categories, and the output of the last full-connection layer is mapped to a (0,1) interval by adopting a softmax function, which can be expressed as:
wherein mu i Is the output value of the ith node, and C is the number of output nodes, namely the number of categories;
(4-3) the domain discriminator confuses the source domain data with the target domain data, and a gradient inversion layer (GRL) exists between the feature extractor and the domain discriminator, expressed as follows:
R λ (x)＝x
R λ (x) Is a pseudo function defining gradient inversion layer, x is network parameter, lambda is constant parameter, the proportion of control domain discriminator in network is selected according to experience, I is unit matrix, andin the transmission process, parameters are automatically negated before the loss of the domain discriminator is transmitted to the feature extractor, so that the training targets of the feature extractor and the domain discriminator are opposite, and a confrontation relation is formed between the feature extractor and the domain discriminator; the domain discriminator comprises two full connection layers, the neuron number of the last layer is 1, and whether the sample is from a source domain or a target domain is judged.
Further, in the step (5), a known fault information source domain sample and an unknown fault information target domain sample are simultaneously input, and the hyperparameters such as iteration times, an optimizer, a learning rate and a sample batch size are set for model training; the model optimization objective in the training process is as follows:
(5-1) the first optimization objective is to minimize the fault classification loss function on the source domain data set, and hope that the label classifier outputs a correct prediction label; the label classifier loss function adopts a cross entropy loss function, and the cross entropy loss function is defined as follows:
wherein p ═ p 0 ,...,p C-1 ]Is a probability distribution, element p i Indicates the probability of a sample belonging to class i, y ═ y 0 ,...,y C-1 ]Is a one-hot representation of the specimen label, y when the specimen belongs to the i-th class i 1, otherwise y i 0; the term for this part of the loss function is:
J F,Y ＝J ce (G Y (G F (x s )),y s )
wherein x s As source domain samples, y s Tagging of source domain samples with J ce (. is a cross entropy loss function;
(5-2) a second optimization goal is that the domain discriminator wants to be able to distinguish samples from the source domain and the target domain, but the features that the feature extractor wants to extract can confuse the domain discriminator, so that the domain discriminator cannot judge the source of the samples, and the two compete in the game; given a source domain sample domain label of 1 and a target domain sample domain label of 0, the optimization objective can be written as:
wherein n is s And n t The number of samples in the source domain and the target domain,
and
respectively source domain and target domain samples;
(5-3) minimizing the feature representation distance of the source domain and target domain data sets after feature extraction by the aid of a third optimization target, and hopefully learning domain-independent features; distance was measured using local maximum-mean difference of fused sample-level weights (sample-level weight LMMD, slw-LMMD), slw-LMMD being defined as follows:
wherein p is (c) And q is (c) Distribution of a source domain data class c and a target domain data set class c, respectively, z s And z t The output of the instances on the source domain and the target domain, respectively, through a feature extractor, E c The finger expects from the category c,
finger according to distribution p (c) In the hope of expectation,
according to distribution q (c) In the expectation that,
to define the symbol, H is the regenerated kernel Hilbert space, φ (-) is a function that maps the original data to H; the probability that the sample belongs to the class c is predicted in a soft label mode due to unsupervised data in the target domain; sThe sample level weight of the target domain is considered by the lw-LMMD on the basis of the LMMD, and the association degree and the similarity degree of each sample in the target domain and the source domain are different, so that different weights are considered for each sample in the target domain during weight calculation, and the inter-domain alignment is facilitated; each sample is then dependent on the weight ω c Belong to each class, so that the unbiased estimate of slw-LMMD is:
wherein
And
the ith instance of the source domain and the jth instance of the target domain are respectively output through the feature extractor,
and
respectively represent
And
the weight belonging to the category c is,
for the output of the target domain samples passing through the domain discriminator, since the source domain sample domain label is 1,
the output range is 0 to 1,
the larger the value, the more similar the target domain sample and the source domain sample are, the higher the weight should be in the inter-domain alignment; y is ic Is a vector y i For the c element, the source domain data uses a real one-hot tag to calculate the weight, and the target domain data uses the softmax output of the tag classifier to calculate the weight; slw-LMMD considers sample level weight, calculates instance mapping value on each category, calculates distance on each subdomain respectively, and then calculates expected value on the category; the term for this part of the loss function is:
(5-4) due to the existence of the gradient inversion layer between the feature extractor and the domain discriminator, the training process is simplified, the aim of resisting training can be achieved in one cycle, and the total optimization target can be written as a schematic expression:
minJ＝J F,Y +αJ F -λJ F,D
wherein alpha is a weight hyperparameter characterizing the distance metric slw-LMMD, and lambda is actually a parameter of a gradient inversion layer, and due to the existence of GRL, the optimization target of a part of the network actually belonging to the domain discriminator is still minJ F,D 。
The invention has the beneficial effects that:
(1) the invention provides a variable-load bearing fault diagnosis method based on a sub-domain adaptive countermeasure network, which is more suitable for an actual fault occurrence scene, can face the unsupervised challenge of rolling bearing fault data, and utilizes fault marking information under different loads to realize cross-domain fault diagnosis of a bearing.
(2) The invention can utilize fine-grained information of fault data and a soft label method to divide the sub-domain of the fault data of the unmarked target domain, thereby realizing sub-domain alignment represented by the data characteristics of the source domain and the target domain.
(3) The method can consider the sample level weight of each sample of the target domain, so that the samples have different importance degrees in the alignment process among the sub-domains, and the inter-domain feature alignment is more facilitated.
(4) The vibration signal is used as a kind of time sequence data, meanwhile, the characteristics of signal layer periodicity, impact property and the like are kept, and the vibration signal has time-frequency domain properties. The invention uses the variational modal decomposition as the data preprocessing, can adaptively decompose the signal into the limited inherent modal function, realizes the frequency domain division of the signal, has better sparsity and effectiveness of the decomposition, and is beneficial to the subsequent characteristic extraction step.
Drawings
FIG. 1 is a flow chart of a bearing fault diagnosis method under variable load based on a subdomain adaptation countermeasure network in the invention.
Fig. 2 is a diagram of a one-dimensional convolutional subfield adaptation countermeasure network architecture.
FIG. 3 is a graph of the convergence of the loss function during network training.
FIG. 4 is a model test result confusion matrix.
Detailed Description
The invention relates to a bearing fault diagnosis method under variable load based on a sub-domain adaptive countermeasure network, which aims at the fault diagnosis problem of a rotating machine. The invention is further elucidated with reference to the figures and embodiments.
The effectiveness of the method is illustrated by combining bearing fault data of the university of western university of reservoir, and the main components of the test bed are as follows: 1.5KW motor, torque sensor/decoder, power tester and electronic controller. As shown in FIG. 1, the implementation steps of the present invention are explained in detail as follows:
the first step is as follows: the method comprises the steps of collecting vibration acceleration signals of a bearing to be detected, wherein the vibration acceleration signals can reflect various health states of the bearing at a specific position under different loads. The data is vibration acceleration data of a bearing at a driving end, which is acquired by an acceleration sensor arranged above a bearing seat, and the sampling frequency is 12 kHz; the motor has four load conditions of 0,1, 2 and 3HP, a vibration signal under 0HP is used as a source domain signal, label information is known, a vibration signal under 3HP is used as a target domain signal, and label information is unknown; the bearing has four faults of normal, inner ring fault, rolling body fault and outer ring fault, and the inner ring, the rolling body and the outer ring respectively contain three faults with fault diameters of 7mil, 14mil and 21mil, so the bearing has 10 health states in total.
The second step is that: setting the length of each sample to be 1024, dividing the vibration signal of each health state into 100 samples by adopting a non-overlapping sampling mode, and then respectively having 1000 samples containing 10 health states in the source domain and the target domain.
The third step: and performing Variational Modal Decomposition (VMD) on the divided source domain signal and the divided target domain signal respectively. Setting VMD parameters, setting a penalty factor alpha of 2000, a noise tolerance tau of 0, a mode number K of 5, a direct current component DC of 0, and a convergence precision epsilon of 1 x 10 -7 And initializing the center frequency mode to be uniform initialization. In order to ensure consistent data dimension, when VMD is carried out on signals under different health states, the parameter K takes the same value of 5. The decomposed IMFs are used as samples to form a source domain data set D S With the target domain data set D T ：
Wherein
And
respectively representing the ith source domain sample and the target domain sample,
sample data tag, n, representing the source domain s And n t Respectively representing the number of samples in a source domain and a target domain; the feature space of both domains is the same as the class space, i.e. X S ＝X T And Y S ＝Y T But the two domain data distributions are different, namely P (X) S )≠P(X T )。
The fourth step: data set D S And D T The method comprises the steps of dividing a training set and a test set into a source domain and a target domain according to a ratio of 7:3, wherein 70 samples of each health state belong to the training set, 30 samples belong to the test set, the source domain comprises 700 training set samples and 300 test set samples, and the target domain comprises 700 training set samples and 300 test set samples. The training set is used for model training, and the testing set is used for predicting the diagnostic effect of the model. The individual data field data set details are shown in table 1.
TABLE 1 Rolling bearing data set information
The fifth step: establishing a fault diagnosis model based on a one-dimensional convolution subdomain adaptive countermeasure network, wherein the model can be divided into three parts: feature extractor G F Label sorter G Y Domain discriminator G D The network structure is shown in fig. 2, and specifically includes:
the feature extractor is used to implicitly extract data features, employing a one-dimensional convolutional neural network. The feature extractor is composed of three convolution pooling modules, the first two pooling layers are operated in a maximal pooling mode, and the last pooling layer is operated in an average pooling mode. To prevent the network from overfitting, a dropout layer is introduced. The source domain and target domain samples share a feature extractor.
The label classifier is used for classifying the source domain data, and the connection feature extractor comprises two full-connection layers, wherein the number of neurons in the last layer is the same as the number of classes, and is set to be 10. The output of the last fully-connected layer is mapped to the (0,1) interval using the softmax function, which can be expressed as:
wherein mu i And C is the output value of the ith node, and the number of output nodes, namely the number of categories.
The domain discriminator is used to discriminate whether the data is from the source domain or the target domain, and the feature extractor is also connected, but since the invention is aimed at the domain discriminator confusing the source domain and the target domain data, there is a gradient inversion layer (GRL) between the feature extractor and the domain discriminator, expressed as follows:
R λ (x)＝x
R λ (x) The method is characterized in that a pseudo function of a gradient inversion layer is defined, x is a network parameter, lambda is a constant parameter, the proportion of a domain discriminator acting in a network is controlled, the pseudo function is generally selected according to experience, I is an identity matrix, and in the process of back propagation, the parameters are automatically inverted before the loss of the domain discriminator is propagated to a feature extractor, so that the training targets of the feature extractor and the domain discriminator are opposite, and a confrontation relation is formed between the feature extractor and the domain discriminator. The domain discriminator comprises two full connection layers, the neuron number of the last layer is 1, and whether the sample is from a source domain or a target domain is judged.
The network parameters are shown in table 2.
TABLE 2 one-dimensional convolution subdomain adaptation countermeasure network parameters
And a sixth step: inputting a marked source domain training set and an unmarked target domain training set, setting the iteration times to be 50 and the sample batch size to be 64, selecting an Adam optimizer, setting the initial learning rate to be 0.005, adjusting the learning rate according to exponential decay in the iteration process, and carrying out model training. The model optimization objective in the training process includes three components, specifically as follows:
the first optimization objective is to minimize the fault classification loss function on the source domain data set, with the hope that the label classifier outputs the correct prediction label. The label classifier loss function adopts a cross entropy loss function, and the cross entropy loss function is defined as follows:
wherein p ═ p 0 ,...,p C-1 ]Is a probability distribution, element p i Indicates the probability of a sample belonging to class i, y ═ y 0 ,...,y C-1 ]Is a one-hot representation of the specimen label, y when the specimen belongs to the i-th class i 1, otherwise y i 0. The partial loss function term is:
J F,Y ＝J ce (G Y (G F (x s )),y s )
wherein x is s As source domain samples, y s Tagging of source domain samples J ce (. cndot.) is a cross entropy loss function.
The second optimization objective is that the domain discriminator wants to be able to distinguish samples from the source domain and the target domain, but the features that the feature extractor wants to extract can confuse the domain discriminator, so that it cannot judge the source of the samples, and the two compete against the game. Given a source domain sample domain label of 1 and a target domain sample domain label of 0, the optimization objective can be written as:
wherein n is s And n t The number of samples in the source domain and the target domain respectively,
and
source domain and target domain samples, respectively.
The third optimization objective is to minimize the feature representation distance of the source domain and target domain data sets after feature extraction, and hope to learn domain-independent features. Distance was measured using local maximum-mean difference of fused sample-level weights (sample-level weight LMMD, slw-LMMD), slw-LMMD being defined as follows:
wherein p is (c) And q is (c) Distribution of a source domain data class c and a target domain data set class c, respectively, z s And z t The output of the instances on the source domain and the target domain, respectively, through a feature extractor, E c Refers to the expectation over the category c,
finger according to distribution p (c) In the hope of expectation,
according to distribution q (c) In the hope of expectation,
to define the notation, H is the regenerated kernel Hilbert space and φ (-) is a function that maps the original data to H. And due to unsupervised data in the target domain, predicting the probability that the sample belongs to the class c by adopting a soft label mode. slw-LMMD considers the sample level weight of the target domain on the basis of the LMMD, and the association degree and the similarity degree of each sample in the target domain and the source domain sample are different, so that different weights are given to each sample in the target domain during weight calculation, and the inter-domain alignment is facilitated. Each sample is then dependent on the weight ω c Belong to each class, so that the unbiased estimate of slw-LMMD is:
wherein
And
the ith instance of the source domain and the jth instance of the target domain are respectively output through the feature extractor,
and
respectively represent
And
the weight belonging to the category c is,
for the output of the target domain samples passing through the domain discriminator, since the source domain samples have a domain label of 1,
the output range is 0 to 1,
the larger, the greater the degree of similarity that indicates this target domain sample is to the source domain sample, the higher the weight should be in inter-domain alignment. y is ic Is a vector y i The c element of (1), the source domain data uses the true one-hot tag to compute weights, and the target domain data uses the softmax output of the tag classifier to compute weights. slw-LMMD considers sample level weight, calculates example mapping value on each category, calculates distance on each subdomain respectively, and calculates expectation value on the category. And (3) introducing a nuclear skill to perform actual calculation:
where k (,) selects a gaussian kernel function. The term for this part of the loss function is:
therefore, due to the existence of the gradient inversion layer between the feature extractor and the domain discriminator, the training process is simplified, the aim of resisting training can be achieved in one cycle, and the total optimization target can be written as a schematic formula:
minJ＝J F,Y +αJ F -λJ F,D
wherein alpha is a weight hyperparameter characterized by representing the distance measure slw-LMMD, and lambda is actually a parameter of the gradient inversion layer in the fifth step, and due to the existence of GRL, the optimization target of the partial network actually belonging to the domain discriminator is still minJ F,D . The parameter α and the parameter λ progressively increase as the number of network iterations increases, as follows:
wherein τ is a coefficient for controlling the variation speed of the parameter α and the parameter λ, 10 is taken in this embodiment; epoch is the current iteration round and Nepoch is the total iteration number.
The training process loss function convergence curve is shown in fig. 3.
The seventh step: inputting the target domain test set into the trained model for testing, outputting a fault diagnosis result, and only concerning the output result of the target domain test set data after passing through the feature extractor and the label classifier. The model test result confusion matrix is shown in fig. 4, which shows that the invention can obtain more accurate diagnosis results. In order to illustrate the feasibility and the advantages of the invention, two groups of schemes are compared, firstly, the sub-domain adaptation is not carried out, and the network is trained by utilizing the source domain data; the second is a Deep Adaptation Network (DAN) of the transfer learning classical method, and the diagnosis accuracy ratio is shown in Table 3.
TABLE 3 comparison of accuracy rates
| Domains-free adaptation | DAN | The method of the invention | |
| Rate of accuracy | 70.67％ | 83.67％ | 98.67％ |
The above-described embodiments are intended to illustrate rather than to limit the invention, and any modifications and variations of the present invention are within the spirit of the invention and the scope of the claims.
Claims (5)
1. A bearing fault diagnosis method under variable load based on a subdomain adaptive countermeasure network is characterized by comprising the following steps:
(1) acquiring multiple health state vibration signals of a rolling bearing under different operation loads to serve as a bearing fault vibration signal database, taking the vibration signals under the loads with known fault information as source domain signals, and taking the vibration signals under the loads with unknown fault information as target domain signals;
(2) carrying out non-overlapping equal-length division on the multiple health state signals of the source domain and the target domain;
(3) respectively carrying out variation modal decomposition on the divided source domain signal and target domain signal, and forming a source domain data set D by the intrinsic modal function group obtained by decomposition S With the target domain data set D T ；
(4) Establishing a fault diagnosis model based on a one-dimensional convolution subdomain adaptive countermeasure network; the model includes a feature extractor G F Label classifier G Y Sum domain discriminator G D (ii) a Feature extractor G F The method is used for implicitly extracting data features and adopts a one-dimensional convolution neural network; label classifier G Y For classifying source domain data, a feature extractor G is connected F (ii) a Domain discriminator G D For discriminating whether the data is from the source domain or the target domain, the feature extractor G is also connected F (ii) a Feature extractor G F And domain discriminator G D A gradient inversion layer is arranged between the two layers;
(5) inputting a source domain signal with known fault information and a target domain signal with unknown fault information into a one-dimensional convolution subdomain adaptation countermeasure network for model training, and fixing model parameters;
(6) in the online diagnosis process, a vibration signal under the load of a section of target domain is collected, variation modal decomposition is carried out, the vibration signal is input into a trained model, and a fault diagnosis result is output.
2. The method for diagnosing the bearing fault under the variable load based on the subdomain adaptive countermeasure network as claimed in claim 1, wherein the state-of-health vibration signals in the step (1) comprise vibration acceleration signals of a plurality of states of health of the bearing with normal state, inner ring fault, outer ring fault, rolling element fault and different degradation degrees.
3. The method for diagnosing the bearing fault under the variable load based on the subdomain adaptive countermeasure network as claimed in claim 1, wherein in the step (3), the divided source domain signal and the divided target domain signal are respectively subjected to variation modal decomposition, VMD parameters are set, and in order to ensure that the data dimensions are consistent, when VMD is carried out on the signals under different health states, the decomposition numbers take the same value; forming a source domain data set D by using the Intrinsic Mode Functions (IMFs) obtained by decomposition as samples S And a target domain data set D T ：
Wherein
And
respectively representing the ith source domain sample and the target domain sample,
sample data tag, n, representing the source domain s And n t Respectively representing the number of samples in a source domain and a target domain; the feature space of both domains is the same as the class space,namely X S ＝X T And Y S ＝Y T But the two domain data distributions are different, namely P (X) S )≠P(X T )。
4. The variable-load bearing fault diagnosis method based on the subdomain adaptive countermeasure network as claimed in claim 1, wherein in the step (4), the establishment of the fault diagnosis model based on the one-dimensional convolution subdomain adaptive countermeasure network specifically comprises:
(4-1) the feature extractor is composed of three convolution pooling modules, the first two pooling layers adopt maximum pooling operation, and the last pooling layer adopts average pooling; in order to prevent the network from being over-fitted, a dropout layer is introduced; the source domain samples and the target domain samples share a feature extractor;
(4-2) the label classifier comprises two full-connection layers, the number of neurons in the last layer is the same as the number of categories, and the output of the last full-connection layer is mapped to a (0,1) interval by using a softmax function, which can be expressed as:
wherein mu i The number of the output nodes is C, namely the number of categories;
(4-3) the domain discriminator confuses the source domain data with the target domain data, and a gradient inversion layer (GRL) exists between the feature extractor and the domain discriminator, expressed as follows:
R λ (x)＝x
R λ (x) The method is characterized in that a pseudo function of a gradient inversion layer is defined, x is a network parameter, lambda is a constant parameter, the proportion of a domain discriminator acting in a network is controlled, the pseudo function is generally selected according to experience, I is an identity matrix, and in the process of back propagation, the parameter is automatically inverted before the loss of the domain discriminator is propagated to a feature extractor, so that the feature extraction is carried outThe training targets of the domain discriminator and the domain discriminator are opposite, so that a confrontation relation is formed between the two; the domain discriminator comprises two full connection layers, the neuron number of the last layer is 1, and whether the sample is from a source domain or a target domain is judged.
5. The variable-load bearing fault diagnosis method based on the sub-domain adaptive countermeasure network according to claim 1, characterized in that in step (5), known fault information source domain samples and unknown fault information target domain samples are input simultaneously, and the model training is performed by setting the iteration times, the optimizer, the learning rate, the sample batch size and other hyperparameters; the model optimization objective in the training process is as follows:
(5-1) the first optimization objective is to minimize the fault classification loss function on the source domain data set, and hope that the label classifier outputs a correct prediction label; the label classifier loss function adopts a cross entropy loss function, and the cross entropy loss function is defined as follows:
wherein p ═ p 0 ,...,p C-1 ]Is a probability distribution, element p i Indicates the probability of a sample belonging to class i, y ═ y 0 ,...,y C-1 ]Is a one-hot representation of the specimen label, y when the specimen belongs to the i-th class i 1, otherwise y i 0; the term for this part of the loss function is:
J F,Y ＝J ce (G Y (G F (x s )),y s )
wherein x s As source domain samples, y s Tagging of source domain samples with J ce (. is a cross entropy loss function;
(5-2) a second optimization goal is that the domain discriminator wants to be able to distinguish samples from the source domain and the target domain, but the features that the feature extractor wants to extract can confuse the domain discriminator, so that the domain discriminator cannot judge the source of the samples, and the two compete in the game; given a source domain sample domain label of 1 and a target domain sample domain label of 0, the optimization objective can be written as:
wherein n is s And n t The number of samples in the source domain and the target domain respectively,
and
respectively source domain and target domain samples;
(5-3) minimizing the feature representation distance of the source domain and target domain data sets after feature extraction by the aid of a third optimization target, and hopefully learning domain-independent features; distance was measured using local maximum-mean difference of fused sample-level weights (sample-level weight LMMD, slw-LMMD), slw-LMMD being defined as follows:
wherein p is (c) And q is (c) Distribution of a source domain data class c and a target domain data set class c, respectively, z s And z t The output of the instances on the source domain and the target domain, respectively, through a feature extractor, E c The finger expects from the category c,
finger according to distribution p (c) In the hope of expectation,
according to distribution q (c) In the expectation that,
for defining the symbol, H is a regenerated nuclear Hilbert spacePhi (·) is a function that maps the raw data to H; the probability that the sample belongs to the class c is predicted in a soft label mode due to unsupervised data in the target domain; slw-LMMD considers the sample level weight of the target domain on the basis of LMMD, and the association degree and the similarity degree of each sample in the target domain and the source domain sample are different, so that different weights are given to each sample in the target domain during weight calculation, and the inter-domain alignment is facilitated; each sample is then dependent on the weight ω c Belong to each class, so that the unbiased estimate of slw-LMMD is:
wherein
And
the ith instance of the source domain and the jth instance of the target domain are respectively output through the feature extractor,
and
respectively represent
And
the weight that belongs to the category c is,
for the output of the target domain samples passing through the domain discriminator, since the source domain samples have a domain label of 1,
the output range is 0 to 1,
the larger the value, the greater the similarity degree between the target domain sample and the source domain sample, and the higher the weight should be when aligning between domains; y is ic Is a vector y i For the c element, the source domain data uses the real one-hot label to calculate the weight, and the target domain data uses the softmax output of the label classifier to calculate the weight; slw-LMMD considers sample level weight, calculates instance mapping value on each category, calculates distance on each subdomain respectively, and then calculates expected value on the category; the term for this part of the loss function is:
(5-4) due to the existence of the gradient inversion layer between the feature extractor and the domain discriminator, the training process is simplified, the aim of resisting training can be achieved in one cycle, and the total optimization target can be written as a schematic expression:
min J＝J F,Y +αJ F -λJ F,D
where α is the weight hyperparameter characterizing the distance metric slw-LMMD and λ is actually a parameter of the gradient inversion layer, which actually belongs to the part of the domain arbiter due to the presence of GRLThe network optimization goal is still min J F,D 。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210684731.7A CN115099270A (en) | 2022-06-16 | 2022-06-16 | Bearing fault diagnosis method under variable load based on sub-domain adaptive countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210684731.7A CN115099270A (en) | 2022-06-16 | 2022-06-16 | Bearing fault diagnosis method under variable load based on sub-domain adaptive countermeasure network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115099270A true CN115099270A (en) | 2022-09-23 |
Family
ID=83290284
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210684731.7A Pending CN115099270A (en) | 2022-06-16 | 2022-06-16 | Bearing fault diagnosis method under variable load based on sub-domain adaptive countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115099270A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115935187A (en) * | 2022-12-12 | 2023-04-07 | 中国地质大学（武汉） | Mechanical fault diagnosis method under variable working conditions based on nuclear sensitivity alignment network |
| CN115964661A (en) * | 2023-01-04 | 2023-04-14 | 东北林业大学 | Rotary machine fault diagnosis method and system based on domain impedance network |
| CN116304905A (en) * | 2023-02-03 | 2023-06-23 | 中国人民解放军92942部队 | Permanent magnet synchronous motor demagnetizing fault diagnosis method under multi-load working condition |
| CN116956048A (en) * | 2023-09-19 | 2023-10-27 | 北京航空航天大学 | Industrial equipment fault diagnosis method and device based on cross-domain generalized label |
| CN117871101A (en) * | 2024-03-11 | 2024-04-12 | 四川大学 | Multi-classifier fine-granularity bearing fault diagnosis field self-adaptive method and system |
| CN117961976A (en) * | 2024-03-29 | 2024-05-03 | 湖南大学 | Assembly robot online detection method and device based on generation diffusion migration |
-
2022
- 2022-06-16 CN CN202210684731.7A patent/CN115099270A/en active Pending
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115935187A (en) * | 2022-12-12 | 2023-04-07 | 中国地质大学（武汉） | Mechanical fault diagnosis method under variable working conditions based on nuclear sensitivity alignment network |
| CN115935187B (en) * | 2022-12-12 | 2023-08-22 | 中国地质大学（武汉） | Nuclear sensitivity alignment network-based mechanical fault diagnosis method under variable working conditions |
| CN115964661A (en) * | 2023-01-04 | 2023-04-14 | 东北林业大学 | Rotary machine fault diagnosis method and system based on domain impedance network |
| CN115964661B (en) * | 2023-01-04 | 2023-09-08 | 东北林业大学 | Rotary machine fault diagnosis method and system based on domain-by-domain network |
| CN116304905A (en) * | 2023-02-03 | 2023-06-23 | 中国人民解放军92942部队 | Permanent magnet synchronous motor demagnetizing fault diagnosis method under multi-load working condition |
| CN116304905B (en) * | 2023-02-03 | 2024-01-23 | 中国人民解放军92942部队 | Permanent magnet synchronous motor demagnetizing fault diagnosis method under multi-load working condition |
| CN116956048A (en) * | 2023-09-19 | 2023-10-27 | 北京航空航天大学 | Industrial equipment fault diagnosis method and device based on cross-domain generalized label |
| CN116956048B (en) * | 2023-09-19 | 2023-12-15 | 北京航空航天大学 | Industrial equipment fault diagnosis method and device based on cross-domain generalized label |
| CN117871101A (en) * | 2024-03-11 | 2024-04-12 | 四川大学 | Multi-classifier fine-granularity bearing fault diagnosis field self-adaptive method and system |
| CN117871101B (en) * | 2024-03-11 | 2024-05-10 | 四川大学 | Multi-classifier fine-granularity bearing fault diagnosis field self-adaptive method and system |
| CN117961976A (en) * | 2024-03-29 | 2024-05-03 | 湖南大学 | Assembly robot online detection method and device based on generation diffusion migration |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115099270A (en) | Bearing fault diagnosis method under variable load based on sub-domain adaptive countermeasure network | |
| Zhang et al. | A fault diagnosis method for wind turbines gearbox based on adaptive loss weighted meta-ResNet under noisy labels | |
| Xiao et al. | Domain adaptive motor fault diagnosis using deep transfer learning | |
| Zhang et al. | Web-supervised network with softly update-drop training for fine-grained visual classification | |
| Collier et al. | Correlated input-dependent label noise in large-scale image classification | |
| CN112308147B (en) | Rotary machinery fault diagnosis method based on multi-source domain anchor adapter integrated migration | |
| CN113567130A (en) | Bearing fault diagnosis method based on multiple working conditions of equipment | |
| CN115563536A (en) | Rolling bearing fault diagnosis method based on subdomain self-adaptation | |
| CN114429150A (en) | Rolling bearing fault diagnosis method and system under variable working conditions based on improved depth subdomain adaptive network | |
| CN113723007A (en) | Mechanical equipment residual life prediction method based on DRSN and sparrow search optimization BilSTM | |
| CN116028876A (en) | Rolling bearing fault diagnosis method based on transfer learning | |
| CN116049719A (en) | Fault detection and estimation method, system and medium based on migration type generation model | |
| CN116894187A (en) | Gear box fault diagnosis method based on deep migration learning | |
| CN115688040A (en) | Mechanical equipment fault diagnosis method, device, equipment and readable storage medium | |
| CN114925809B (en) | LSTM-based printer bearing fault diagnosis method and device | |
| Zhang et al. | Deep transfer learning with self-attention for industry sensor fusion tasks | |
| Estévez et al. | Cytological image analysis with a genetic fuzzy finite state machine | |
| CN115935187B (en) | Nuclear sensitivity alignment network-based mechanical fault diagnosis method under variable working conditions | |
| CN115049852B (en) | Bearing fault diagnosis method and device, storage medium and electronic equipment | |
| CN116595313A (en) | Fault diagnosis method and system based on multi-scale residual error attention network | |
| Al Iqbal | Eclectic rule extraction from Neural Networks using aggregated Decision Trees | |
| Feuer et al. | A meta-analysis of distributionally-robust models | |
| Ren et al. | Meta-Learning With Distributional Similarity Preference for Few-Shot Fault Diagnosis Under Varying Working Conditions | |
| Ahn et al. | Fine tuning Pre trained Models for Robustness Under Noisy Labels | |
| Chen et al. | Gearbox fault diagnosis classification with empirical mode decomposition based on improved long short-term memory |
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 |