CN115795313A - Training method of nuclear main pump fault diagnosis model, fault diagnosis method and system - Google Patents
Training method of nuclear main pump fault diagnosis model, fault diagnosis method and system Download PDFInfo
- Publication number
- CN115795313A CN115795313A CN202310059996.2A CN202310059996A CN115795313A CN 115795313 A CN115795313 A CN 115795313A CN 202310059996 A CN202310059996 A CN 202310059996A CN 115795313 A CN115795313 A CN 115795313A
- Authority
- CN
- China
- Prior art keywords
- fault
- hyper
- fault diagnosis
- parameter
- model
- 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
- 238000003745 diagnosis Methods 0.000 title claims abstract description 95
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000012549 training Methods 0.000 title claims abstract description 49
- 230000006870 function Effects 0.000 claims description 32
- 238000007781 pre-processing Methods 0.000 claims description 18
- 239000002826 coolant Substances 0.000 claims description 17
- 238000005457 optimization Methods 0.000 claims description 17
- 238000009826 distribution Methods 0.000 claims description 16
- 239000000126 substance Substances 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 8
- 238000012795 verification Methods 0.000 claims description 8
- 238000010200 validation analysis Methods 0.000 claims description 5
- 101100533306 Mus musculus Setx gene Proteins 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 3
- 238000003062 neural network model Methods 0.000 claims description 3
- 230000011664 signaling Effects 0.000 claims description 3
- 230000000087 stabilizing effect Effects 0.000 claims description 3
- 230000002265 prevention Effects 0.000 claims 1
- 230000008569 process Effects 0.000 abstract description 2
- 238000005070 sampling Methods 0.000 description 11
- 238000005516 engineering process Methods 0.000 description 5
- 238000013508 migration Methods 0.000 description 3
- 230000005012 migration Effects 0.000 description 3
- 238000005299 abrasion Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 230000003449 preventive effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000001816 cooling Methods 0.000 description 1
- 238000005336 cracking Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Abstract
The invention relates to the field of nuclear main pump fault diagnosis, in particular to a training method, a fault diagnosis method and a system of a nuclear main pump fault diagnosis model. The invention provides a training method of a nuclear main pump fault diagnosis model, which comprises the steps of constructing a training sample by combining a source domain and a target domain, extracting input data characteristics through a characteristic extractor, judging a domain label of input data by a domain discriminator based on the input data characteristics, so that the characteristics extracted by the characteristic extractor are common characteristics of the target domain and the source domain by combining the dependence of the characteristics extracted by the domain discriminator on the domain and learning the training sample; the fault classifier can realize fault diagnosis of the target domain sample through learning of the source domain sample, and requirements on the target domain sample in a training process are greatly reduced. The method and the device ensure that the fault diagnosis model obtained when the target domain samples are insufficient can still realize good fault classification in the target domain, and improve the performance of the fault diagnosis model.
Description
Technical Field
The invention relates to the field of nuclear main pump fault diagnosis, in particular to a training method, a fault diagnosis method and a system of a nuclear main pump fault diagnosis model.
Background
The nuclear main pump is one of critical devices in the reactor. The function of the reactor is to ensure the stable circulation of the coolant in the primary loop, thereby cooling the reactor core. As the only rotating equipment in the nuclear island, the nuclear main pump has a complex integral structure, inevitably causes various faults after long-time operation, slightly causes abnormal shutdown, increases economic cost, seriously causes serious accidents and brings hidden dangers to social stability. With the rapid development of computer technology in recent years in industrial engineering, fault diagnosis technology has gradually shifted from traditional probabilistic inference model-based methods to data-driven methods that do not require model building. Researchers related to the field of home and abroad fault diagnosis research gradually begin to try, develop and explore the application value potential of the intelligent fault diagnosis technology in the scenes of aerospace, mechanical manufacturing, industrial engineering and the like by combining the deep learning technology, and obtain a plurality of research results. However, as time goes on, when the overall operating environment or working condition of the nuclear device changes, the originally trained model loses the diagnosis precision and the generalization performance for the fault data in the new environment. In addition, the task of labeling fault samples under new operating conditions is extremely challenging. Therefore, how to fully utilize the existing model and part of the marked samples to diagnose the unmarked samples is a key problem for further development of the nuclear main pump fault diagnosis technology.
Disclosure of Invention
In order to solve the defects that in the prior art, nuclear main pump labeled samples are difficult to obtain under new working conditions, and the generalization performance of a deep learning model trained on the existing working condition samples is poor under the new working conditions, the invention provides a training method of a nuclear main pump fault diagnosis model, and the fault diagnosis model obtained by training has good generalization performance.
The invention provides a training method of a nuclear main pump fault diagnosis model, which comprises the following steps:
s1, combining the selected hyper-parameters to construct a basic model consisting of a feature extractor, a domain discriminator and a fault classifier; the basic model is a neural network model, and the hyper-parameters are used for defining the structure of the basic model; the fault classifier is used for judging fault types based on the features extracted by the feature extractor; the output of the fault classifier is the probability that the input data belongs to various fault categories;
constructing training samples{(x,y)|x∈D s ∪D t ,y∈Y}，D s A set of source domain samples is represented,D t a set of samples representing the target domain is shown,xwhich represents the input data, is,yrepresenting input dataxThe corresponding category of the fault is,Yrepresenting a set of source domain fault labels and target domain fault labels; constructing validation samples{(x)|x∈D s ∪D t }The training samples and the verification samples are both known in source, namely, the training samples and the verification samples are known to be source domain samples or target domain samples;
s2, making the basic model learnk1Training samples are updated to update the weight parameters;k1is a set value;
s3, combiningk2First loss function of verification sample calculation base modelL d And a second loss functionL c ；k2Is a set value; first loss functionL d The fuzzy capability of the basic model to the domain is evaluated, and the smaller the first loss function is, the larger the probability that the feature extracted by the feature extractor is a common feature of the source domain sample and the target domain sample is; second loss functionL c The fault diagnosis capability of the basic model on the input data is evaluated, and the accuracy of the fault category output by the fault classifier is higher when the second loss function is smaller;
s4, judging whether the requirements are met|L d |≤L1And is provided with|L c |≤L2；L1AndL2for a set loss threshold(ii) a If not, returning to the step S2; if yes, stabilizing the weight parameters of the basic model, and extracting the feature extractor and the fault classifier to form a fault diagnosis model.
Preferably, in S1, the selecting of the hyper-parameter includes the following steps:
SA1, constructionn0Each hyper-parameter group consists of a group of hyper-parameters; combining each super parameter set to construct a basic model;
SA2, training the basic model by combining the training samples, and when the iteration number of the basic model reaches a set valuek3Computing a second loss function of the base modelL c Model loss corresponding to the hyperparameter set;
SA3, constructing a hyper-parameter sample(x’,f(x’))，x’A set of super-parameters is represented,f(x’)for the super parameter setx’The corresponding model loss; constructing a hyper-parametric sample pool, wherein the initial value of the hyper-parametric sample pool comprises SA1 settingn0A hyper-parameter sample corresponding to each hyper-parameter set;
SA4, making all hyper-parameter samples in the hyper-parameter sample pool accord with Gaussian distribution, and calculating lower signaling value of the Gaussian distributionU(x’)=µ(f(x’))-k×σ(f(x’))；µ(f(x’))Is the mean of the gaussian distribution and,σ(f(x’)is the variance of the gaussian distribution and,kto set confidence, 0.05≦k≦0.1；
SA5, lower confidence interval [ 2 ] on the Gaussian distributionU(x’)，µ(f(x’)]Selecting the minimum value as a target loss;
SA6, judging whether the target loss is less than all model losses in a hyper-parameter sample pool; if not, acquiring a hyper-parameter set corresponding to the target loss, adding the target loss and the hyper-parameter set corresponding to the target loss into a hyper-parameter sample pool as a hyper-parameter sample, and then returning to the step SA4; if so, enabling the hyper-parameter group corresponding to the target loss as an optimal hyper-parameter group; the optimal set of hyper-parameters is the hyper-parameters selected in S1.
Preferably, the first loss functionL d And a second loss functionL c Optimizing the target setting by combining the weight parameters:
the parameter optimization goal of the feature extractor is as follows: order toL d To a maximum value ofL c A minimum value is achieved;
the parameter optimization goal of the domain discriminator is as follows: order toL d A maximum value is achieved;
the parameter optimization target of the fault classifier is as follows: order toL c A maximum value is achieved.
Preferably, the first loss function is:
wherein, the first and the second end of the pipe are connected with each other,x i denotes the firstiInput data in a sample;d i is composed ofx i The true domain tag of (a); if it is notx i From the target domain, thend i =1; if it is notx i From the source domain, thend i =0；D(F(x i ))The expression domain discriminator judges the input datax i Probability from target domain, 0≦D(F(x i ))≦1；βA very small constant representing a preventive denominator of 0,βis a set value;Iindicating the number of samples.
Preferably, the second loss function is:
wherein the content of the first and second substances,Ps(x i )representing base models against input datax i The maximum probability value of the output fault category probabilities;x i is shown asiThe input data in one of the samples is,Iindicating the number of samples.
Preferably, in S2, the weight parameter updating method is as follows:
wherein the content of the first and second substances,k d representing a first loss functionL d The penalty factor of (2) is determined,k c representing the second loss functionL c The penalty factor of (2) is determined,k d andk c are all set values;αis a function of the learning rate of the underlying model,αis a set hyper-parameter;
to the left of the arrowθ G Weight parameter representing updated feature extractor, right of arrowθ G A weight parameter representing the feature extractor before updating; to the left of the arrowθ D Representing the updated weight parameter of the pre-discriminator, to the right of the arrowθ D Representing a weight parameter of the pre-arbiter before updating; to the left of the arrowθ C Weight parameter representing updated fault classifier, right of arrowθ C A weight parameter representing a fault classifier before updating;
ǝL c /ǝθ G representing the second loss functionL c Partial derivatives passed on the feature extractor;ǝL d /ǝθ G representing a first loss functionL d Partial derivatives passed on the feature extractor;ǝL c /ǝθ C representing the second loss functionL c Partial derivatives passed on the fault classifier;ǝL d /ǝθ D representing a first loss functionL d Partial derivatives passed on the domain arbiter.
Preferably, the basic model further comprises a data preprocessing module, and the data preprocessing module preprocesses the sampling data to obtain input data; order to input dataxThe corresponding sampled data is recorded asx0；
Wherein the content of the first and second substances,τis an adaptive threshold; soft () is the threshold softening function; the sampled data includes one or more of vibration signal data, current data, and thermal parameters of the coolant; the thermal parameters of the coolant include temperature data of the coolant, flow data of the coolant, and pressure data of the coolant.
The nuclear main pump fault diagnosis method provided by the invention realizes the application of the fault diagnosis model.
The invention provides a nuclear main pump fault diagnosis method, which comprises the steps of firstly obtaining a fault diagnosis model, wherein the fault diagnosis model is obtained by adopting a training method of the nuclear main pump fault diagnosis model; then acquiring input data of the nuclear main pump to be diagnosed, inputting the input data into a fault diagnosis model, and acquiring a fault category corresponding to the maximum probability output by the fault diagnosis model as a diagnosis result; the fault categories include normal stable operating conditions, bearing wear, misalignment of the rotor, impeller eye rub-on, shroud rub-on, and some or all of the rotor cracks.
The nuclear main pump fault diagnosis system provided by the invention provides a carrier for the nuclear main pump fault diagnosis method, and facilitates the popularization of the nuclear main pump fault diagnosis method.
The invention provides a fault diagnosis system of a nuclear main pump, which comprises a memory, wherein a fault diagnosis model and a computer program are stored in the memory, and the computer program is used for realizing the fault diagnosis method of the nuclear main pump when being executed.
The invention provides another nuclear main pump fault diagnosis system which comprises a memory and a processor, wherein a computer program is stored in the memory, the processor is connected with the memory, and the processor is used for executing the computer program to realize the nuclear main pump fault diagnosis method.
The invention has the advantages that:
(1) The invention provides a training method of a nuclear main pump fault diagnosis model, which comprises the steps of constructing a training sample by combining a source domain and a target domain, extracting input data characteristics through a characteristic extractor, judging a domain label of input data by a domain discriminator based on the input data characteristics, so that the characteristics extracted by the characteristic extractor are common characteristics of the target domain and the source domain by combining the dependence of the characteristics extracted by the domain discriminator fuzzy characteristic extractor on the domain and learning the training sample; the fault classifier can realize fault diagnosis of the target domain sample through learning of the source domain sample, and requirements on the target domain sample in a training process are greatly reduced. The method ensures that the fault diagnosis model obtained when the target domain samples are insufficient can still realize good fault classification in the target domain, and improves the performance of the fault diagnosis model.
(2) The method is based on the deep learning characteristic migration algorithm, migration generalization of the nuclear main pump fault diagnosis model under different operation conditions is realized, accurate positioning and diagnosis of the nuclear main pump fault are realized, and the training cost of the diagnosis model for redeployment under different operation conditions is reduced.
(3) In the invention, the weight parameter optimization target of the fault classifier and the weight parameter optimization target of the domain discriminator form a confrontation game relation, so that the characteristics selected by the characteristic extractor and available for migration have domain invariance and independence, wherein the domain invariance means that the finally extracted characteristics can confuse the domain discriminator and cannot judge whether the characteristics come from a source domain or a target domain, and a fault diagnosis model trained by sampling more source domain samples can be better applied to fault diagnosis of the target domain; the independence means that the extracted features can complete the loss target in the fault classifier, and an ideal classification diagnosis effect is achieved.
(4) According to the method, the first loss function and the second loss function are set in combination with the weight parameter optimization target, the loss targets in the common feature extraction process of different domains are unified through gradient inversion operation, and the performance of the fault diagnosis model is further improved.
(5) According to the invention, the self-adaptive threshold softening of the sampling data is realized through data preprocessing; the data preprocessing module can update the threshold value in the basic model training, reduces the risks of gradient disappearance and gradient explosion when the basic model parameters are updated, and further improves the reliability and the convergence efficiency of the model training.
(6) The method and the device combine vibration signal data, current data, thermal parameters of coolant and the like to construct sampling data and input data, provide more monitoring bases for fault diagnosis, and are more comprehensive in fault state characteristics through fault information reaction collected from different mode data compared with the method and the device for monitoring the faults of the nuclear main pump through single vibration data.
(7) The invention also provides a hyper-parameter optimization method, which is characterized in that initial hyper-parameter distribution is selected by a random method, and the next possible optimal value is selected by utilizing the performance of the hyper-parameters. The choice of each hyper-parameter depends on previous attempts. Second, the next set of hyper-parameters is selected based on the history and performance is evaluated until the best combination is found or the maximum number of trials is reached. And finally, acquiring a more ideal parameter selection result under fewer iteration times. In the invention, the fault diagnosis model is trained based on the optimized hyper-parameters, and the performance of the fault diagnosis model is further improved.
Drawings
FIG. 1 is a flowchart of a method for training a fault diagnosis model of a nuclear main pump;
FIG. 2 is a flowchart of a method for hyper-parameter optimization.
Detailed Description
In the embodiment, the data of the nuclear main pump is combined to diagnose the fault of the nuclear main pump, and the sampling data is preprocessed, the noise data is washed away, and the input data is obtained. In this embodiment, the sampling data includes one or more of vibration signal data, current data, and thermal parameters of the coolant; the thermal parameters of the coolant include temperature data of the coolant, flow data of the coolant, and pressure data of the coolant. During specific implementation, the sampling data can comprise vibration signal data of the nuclear main pump at a plurality of different positions, and the current data can be acquired from the key positions of the nuclear main pump, so that the comprehensiveness of the sampling data on the state representation of the nuclear main pump is improved, and the accuracy of fault diagnosis is improved.
Data preprocessing method
The data is preprocessedThe method is used for preprocessing the sampling data to obtain input data; inputting dataxCorresponding sampled data is recorded asx0；
Wherein the content of the first and second substances,τis an adaptive threshold; soft () is a threshold softening function.
Fault diagnosis model
In the present embodiment, the input data is subjected to the failure diagnosis by the failure diagnosis model. The fault diagnosis model comprises: a feature extractor and a fault classifier; the input of the characteristic extractor is input data; the characteristic extractor is used for extracting the characteristics of the input data; the output of the feature extractor is the input of a fault classifier, and the output of the fault classifier is the output of the fault diagnosis model;
the output of the fault classifier is the probability that the test target belongs to each fault category, and the sum of all the probabilities is 1. In this embodiment, the fault category includes normal stable operation conditions, bearing wear, misalignment of the rotor, collision and abrasion of the impeller opening ring, collision and abrasion of the shield sleeve, rotor cracking, and the like.
It should be noted that, in the present embodiment, the input of the feature extractor can be directly used as the input of the fault diagnosis model. The fault diagnosis model also can be set to further comprise a data preprocessing module, and the data preprocessing module is used for realizing the data preprocessing method; at this time, the input of the data preprocessing module is the input of the fault diagnosis model, and the output of the data preprocessing module is the output of the feature extractor.
Nuclear main pump fault diagnosis method
The method for diagnosing the fault of the nuclear main pump comprises the steps of firstly obtaining input data of the nuclear main pump to be diagnosed, inputting the input data into a feature extractor in a fault diagnosis model, and outputting probability of each fault classification by a fault classifier of the fault diagnosis model; and acquiring the fault category corresponding to the maximum probability as a diagnosis result.
Training method of nuclear main pump fault diagnosis model
Referring to fig. 1, the acquisition of the failure diagnosis model in the present embodiment includes the following steps S1 to S4.
S1, combining the selected hyper-parameters to construct a basic model consisting of a feature extractor, a domain discriminator and a fault classifier; the input of the feature extractor is input data of the basic model, and the output of the feature extractor is the feature of the input data; the input of the domain discriminator and the input of the fault classifier are both the output of the feature extractor; the output of the domain discriminator is the domain label of the input data, namely the input data comes from the source domain or the target domain; the output of the fault classifier is the probability that the input data belongs to various fault categories;
constructing training samples{(x,y)|x∈D s ∪D t ,y∈Y}，D s A set of source domain samples is represented,D t a set of samples representing the target domain is represented,xwhich represents the input data, is,yrepresenting input dataxThe corresponding category of the fault is a category of fault,Yrepresenting a set of source domain fault labels and target domain fault labels; constructing validation samples{(x)|x∈D s ∪D t }Both the training samples and the validation samples are known to be of known origin, i.e. whether they are known as source domain samples or target domain samples.
In specific implementation, the target domain sample is obtained by data preprocessing of the sampling data under the target working condition, and the source domain sample is obtained by data preprocessing of the sampling data under the source domain working condition. Aiming at the nuclear main pump, a target working condition and a source area working condition can be distinguished according to the rotating speed, namely the target working condition is the working condition of the nuclear main pump at the target rotating speed, and the source area working condition is the working condition of the nuclear main pump at the set rotating speed. The working condition data under the set rotating speed is easier to collect than the working condition data under the target rotating speed so as to provide more source domain samples and facilitate the training of the basic model.
S2, making the basic model studyk1Training samples are updated to update the weight parameters;k1is a set value.
In this step, the set first loss function is combinedL d And a second loss functionL c Updating a base modelFirst loss functionL d The fuzzy capability of the basic model to the domain is evaluated, and the smaller the first loss function is, the larger the probability that the feature extracted by the feature extractor is the common feature of the source domain sample and the target domain sample is; second loss functionL c The smaller the second loss function is, the higher the accuracy of the fault category representing the output of the fault classifier is.
The weight parameter updating method comprises the following steps:
wherein the content of the first and second substances,k d representing a first loss functionL d The penalty factor of (2) is determined,k c representing the second loss functionL c The penalty factor of (2) is calculated,k d andk c are all set values;αfor the learning rate of the underlying model,αis a set hyper-parameter;
to the left of the arrowθ G Weight parameter representing updated feature extractor, right of arrowθ G A weight parameter representing the feature extractor before updating; to the left of the arrowθ D Weight parameter representing updated pre-decider, right of arrowθ D Representing a weight parameter of the pre-arbiter before updating; to the left of the arrowθ C Weight parameter representing updated fault classifier, right of arrowθ C A weight parameter representing a fault classifier before updating;
ǝL c /ǝθ G representing the second loss functionL c Partial derivatives passed on the feature extractor;ǝL d /ǝθ G representing a first loss functionL d Partial derivatives passed on the feature extractor;ǝL c /ǝθ C representing the second loss functionL c Partial derivatives passed on the fault classifier;ǝL d /ǝθ D representing a first loss functionL d Partial derivatives passed on the domain arbiter.
Specifically, in the present embodiment, the first loss function is set in combination with the weight parameter optimization targetL d And a second loss functionL c 。
In this embodiment, the weight parameter optimization objective is:
θ G =arg{maxL d (θ D ,θ G ),minL c (θ G ,θ C )}
θ D =arg maxL d (θ D ,θ G )
θ C =arg maxL c (θc,θ G )
whereinθ G ，θ D And withθ C Respectively representing parameters of a feature extractor, a domain discriminator and a fault classifier;L d is a first loss function representing the loss of the domain discriminator;L c a second loss function representing the loss of the fault classifier;
θ G =arg{maxL d (θ D ,θ G ),minL c (θ G ,θ C ) And expressing that the parameter optimization target of the feature extractor is as follows: order toL d To a maximum value ofL c A minimum value is achieved;
θ D =arg maxL d (θ D ,θ G ) The parameter optimization objective for the representation domain arbiter is: order toL d A maximum value is achieved;
θ C =arg maxL c (θc,θ G ) The parameter optimization objective representing the fault classifier is: order toL c A maximum value is achieved.
Wherein the content of the first and second substances,x i is shown asiInput data in a sample;d i is composed ofx i The true domain tag of (a); if it is usedx i From the target domain, thend i =1; if it is usedx i From the source domain, thend i =0；D(F(x i ))The representation domain discriminator judges the input datax i Probability from target domain, 0≦D(F(x i ))≦1；βA very small constant representing a preventive denominator of 0,βis a set value, and can be specifically in the interval [10 ] -6 ,10 -5 ]Taking the value;Iindicating the number of samples. Specifically, in the step S2,Inumber of training samples learned for the base model.
Wherein the content of the first and second substances,Ps(x i )representing base models for input datax i The maximum probability value of the output fault category probabilities;
the fault classifier comprises a full connection layer and a softmax function; the output of the full connection layer is the weighted value of the input data corresponding to different fault categories; the input of the softmax function is the output of the full connection layer, and the output of the softmax function is the probability value of the maximum possible fault classification corresponding to the input data;
namely:
Ps(x i )representing input datax i A probability value of the corresponding maximum possible fault category,eis a natural number, and is provided with a plurality of groups,z i input data for output of full connection layerx i A corresponding maximum weight value;z ij input data for output of full connection layerx i Corresponds to the firstjThe weight value of each of the fault categories,z ij ∈Y，Yrepresenting a set of source domain fault labels and target domain fault labelsYAnd | is the total number of the source domain fault labels and the target domain fault labels.
S3, combiningk2First loss function of verification sample calculation base modelL d And a second loss functionL c ；k2Is a set value.
In this step:
wherein, the first and the second end of the pipe are connected with each other,x i is shown asiInput data in a validation sample;d i is composed ofx i The true domain tag of (a); if the sample is verifiedx i From the target domain, thend i =1; if the verification sample is from the source domain, thend i =0；D(F(x i ))The representation domain discriminator judges the input datax i From the target domainProbability of 0≦D(F(x i ))≦1；
Wherein, the first and the second end of the pipe are connected with each other,Ps(x i )representing base models for input datax i And outputting the maximum probability value of the fault category probabilities.
S4, judging whether the requirements are met|L d |≤L1And is|L c |≤L2；L1AndL2is a set loss threshold; if not, returning to the step S2; if yes, stabilizing the weight parameters of the basic model, and extracting the feature extractor and the fault classifier to form a fault diagnosis model.
In this embodiment, the basic model is a neural network model, and the hyper-parameters are used to define the structure of the basic model, such as convolution size, convolution kernel size, learning rate, convergence rate, and the like. The weight parameter is the weight of each network node in the basic model.
It should be noted that, in the present embodiment, the input of the basic model is consistent with the input of the fault diagnosis model, and if the fault diagnosis model includes the data preprocessing module, the basic model should also include the data preprocessing module; when the basic model comprises a data preprocessing module, the threshold is self-adaptedτUpdates may be made during base model training.
Optimization training method of fault diagnosis model
Referring to fig. 2, in the present embodiment, the following steps SA1 to SA6 are combined to first optimize the hyper-parameters and then train the fault diagnosis model.
SA1, constructionn0Each hyper-parameter set is composed of a group of hyper-parameters, each hyper-parameter set corresponds to a basic model, and each basic model comprises a feature extractor, a domain discriminator and a fault classifier;
SA2, training the basic model by combining the training samples, and when the iteration times of the basic model reach a set valuek3Computing a second loss function of the base modelL c Model loss corresponding to the hyperparameter set;
SA3, constructing a hyper-parameter sample(x’,f(x’))，x’A set of super-parameters is represented,f(x’)for the super parameter setx’The corresponding model loss; constructing a hyper-parametric sample pool, wherein the initial value of the hyper-parametric sample pool comprises SA1 settingn0A hyper-parameter sample corresponding to each hyper-parameter group;
SA4, making all hyper-parameter samples in the hyper-parameter sample pool accord with Gaussian distribution, and calculating the lower signaling value of the Gaussian distributionU(x’)=µ(f(x’))-k×σ(f(x’))；µ(f(x’))Is the mean of the gaussian distribution and,σ(f(x’)is the variance of the gaussian distribution and,kto set confidence, 0.05≦k≦0.1；
SA5, lower confidence interval [ 2 ] on the Gaussian distributionU(x’)，µ(f(x’)]Selecting the minimum value as a target loss;
SA6, judging whether the target loss is less than all model losses in the hyper-parameter sample pool; if not, acquiring a hyper-parameter group corresponding to the target loss, adding the target loss and the hyper-parameter group corresponding to the target loss into a hyper-parameter sample pool as a hyper-parameter sample, and then returning to the step SA4; if so, enabling the hyper-parameter group corresponding to the target loss as an optimal hyper-parameter group; and then, executing the steps S1-S4 by combining the optimal super parameter group to obtain a fault diagnosis model.
The invention is not to be considered as limited to the specific embodiments shown and described, but is to be understood to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.
Claims (10)
1. A training method for a nuclear main pump fault diagnosis model is characterized by comprising the following steps:
s1, combining the selected hyper-parameters to construct a basic model composed of a feature extractor, a domain discriminator and a fault classifier; the basic model is a neural network model, and the hyper-parameters are used for defining the structure of the basic model; the fault classifier is used for judging fault types based on the features extracted by the feature extractor; the output of the fault classifier is the probability that the input data belongs to various fault categories;
constructing training samples{(x,y)|x∈D s ∪D t ,y∈Y}，D s A set of source domain samples is represented,D t a set of samples representing the target domain is shown,xwhich represents the input data, is,yrepresenting input dataxThe corresponding category of the fault is,Yrepresenting a set of source domain fault labels and target domain fault labels; constructing validation samples{(x)|x∈D s ∪D t }The training samples and the verification samples are both known to be from a source, namely, whether the training samples and the verification samples are known to be source domain samples or target domain samples;
s2, making the basic model learnk1Training samples are updated to update the weight parameters;k1is a set value;
s3, combiningk2First loss function of verification sample calculation base modelL d And a second loss functionL c ；k2Is a set value; first loss functionL d The fuzzy capability of the basic model to the domain is evaluated, and the smaller the first loss function is, the larger the probability that the feature extracted by the feature extractor is the common feature of the source domain sample and the target domain sample is; second loss functionL c The fault diagnosis capability of the basic model on the input data is evaluated, and the accuracy of the fault category output by the fault classifier is higher when the second loss function is smaller;
s4, judging whether the requirements are met|L d |≤L1And is|L c |≤L2；L1AndL2is a set loss threshold; if not, returning to the step S2; if yes, stabilizing the weight parameters of the basic model, and extracting the feature extractor and the fault classifier to form a fault diagnosis model.
2. The method for training the nuclear main pump fault diagnosis model according to claim 1, wherein in S1, the selection of the hyper-parameter comprises the following steps:
SA1, constructionn0Each hyper-parameter group consists of a group of hyper-parameters; combining each super parameter set to construct a basic model;
SA2, training the basic model by combining the training samples, and when the iteration number of the basic model reaches a set valuek3Computing a second loss function of the base modelL c Model loss corresponding to the super parameter group;
SA3, constructing a hyper-parameter sample(x’,f(x’))，x’A set of super-parameters is represented,f(x’)for the super parameter setx’The corresponding model loss; constructing a hyper-parametric sample pool, wherein the initial value of the hyper-parametric sample pool comprises SA1 settingn0A hyper-parameter sample corresponding to each hyper-parameter set;
SA4, making all hyper-parameter samples in the hyper-parameter sample pool accord with Gaussian distribution, and calculating lower signaling value of the Gaussian distributionU (x’)=µ(f(x’))-k×σ(f(x’))；µ(f(x’))Is the mean of the gaussian distribution and,σ(f(x’)is the variance of the gaussian distribution and,kto set confidence, 0.05≦k≦0.1；
SA5, lower confidence interval [ 2 ] on the Gaussian distributionU(x’)，µ(f(x’)],Selecting the minimum value as a target loss;
SA6, judging whether the target loss is less than all model losses in the hyper-parameter sample pool; if not, acquiring a hyper-parameter group corresponding to the target loss, adding the target loss and the hyper-parameter group corresponding to the target loss into a hyper-parameter sample pool as a hyper-parameter sample, and then returning to the step SA4; if so, enabling the hyperparameter group corresponding to the target loss to serve as an optimal hyperparameter group; the optimal set of hyper-parameters is the hyper-parameters selected in S1.
3. The method of claim 1, wherein the first loss function is a function of a first lossL d And a second loss functionL c Combined with weight parameter optimizationSetting a target:
the parameter optimization goal of the feature extractor is as follows: order toL d To achieve a maximum value, orderL c A minimum value is achieved;
the parameter optimization goal of the domain discriminator is as follows: order toL d A maximum value is achieved;
the parameter optimization target of the fault classifier is as follows: order toL c A maximum value is achieved.
4. The method for training a fault diagnosis model of a nuclear main pump according to claim 3, wherein the first loss function is:
wherein the content of the first and second substances,x i is shown asiInput data in a sample;d i is composed ofx i The true domain tag of (a); if it is notx i From the target domain, thend i =1; if it is notx i From the source domain, thend i =0；D(F(x i ))The expression domain discriminator judges the input datax i Probability from target domain, 0≦D(F(x i ))≦1；βA very small constant representing a prevention denominator of 0,βis a set value;Iindicating the number of samples.
5. The method for training a fault diagnosis model of a main nuclear pump according to claim 3, wherein the second loss function is:
wherein the content of the first and second substances,Ps(x i )representing base models against input datax i Of the output fault class probabilitiesA maximum probability value;x i denotes the firstiThe input data in one of the samples is,Iindicating the number of samples.
6. The method for training the nuclear main pump fault diagnosis model according to claim 1, wherein in S2, the weight parameter updating mode is as follows:
wherein, the first and the second end of the pipe are connected with each other,k d representing a first loss functionL d The penalty factor of (2) is determined,k c representing the second loss functionL c The penalty factor of (2) is determined,k d andk c are all set values;αa hyper-parameter of a set learning rate as a basic model;
to the left of the arrowθ G Weight parameter representing updated feature extractor, right of arrowθ G A weight parameter representing the feature extractor before updating; to the left of the arrowθ D Representing the updated weight parameter of the pre-discriminator, to the right of the arrowθ D Representing a weight parameter of the pre-arbiter before updating; to the left of the arrowθ C Weight parameter representing updated fault classifier, right of arrowθ C A weight parameter representing a fault classifier before updating;
ǝL c /ǝθ G representing the second loss functionL c Partial derivatives passed on the feature extractor;ǝL d /ǝθ G representing a first loss functionL d Partial derivatives passed on the feature extractor;ǝL c /ǝθ C representing the second loss functionL c On-failure classifierPartial derivatives of the up-transfer;ǝL d /ǝθ D representing a first loss functionL d Partial derivatives passed on the domain arbiter.
7. The training method of the fault diagnosis model of the nuclear main pump according to claim 1, wherein the basic model further comprises a data preprocessing module, the data preprocessing module preprocesses the sampled data to obtain input data; order to input dataxThe corresponding sampled data is recorded asx0；
Wherein the content of the first and second substances,τis an adaptive threshold; soft () is the threshold softening function; the sampled data includes one or more of vibration signal data, current data, and thermal parameters of the coolant; the thermal parameters of the coolant include temperature data of the coolant, flow data of the coolant, and pressure data of the coolant.
8. A fault diagnosis method for a nuclear main pump is characterized in that a fault diagnosis model is obtained firstly, and the fault diagnosis model is obtained by adopting the training method for the fault diagnosis model of the nuclear main pump according to any one of claims 1 to 7; then acquiring input data of the nuclear main pump to be diagnosed, inputting the input data into a fault diagnosis model, and acquiring a fault category corresponding to the maximum probability output by the fault diagnosis model as a diagnosis result; the fault categories include normal stable operating conditions, bearing wear, misalignment of the rotor, impeller eye rub-on, shroud rub-on, and some or all of the rotor cracks.
9. A nuclear main pump fault diagnosis system comprising a memory in which a fault diagnosis model and a computer program are stored, the computer program being adapted to implement the nuclear main pump fault diagnosis method according to claim 8 when executed.
10. A system for diagnosing a fault of a main pump of a nuclear reactor, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is connected with the memory and is used for executing the computer program to implement the method for diagnosing the fault of the main pump of the nuclear reactor according to claim 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310059996.2A CN115795313A (en) | 2023-01-16 | 2023-01-16 | Training method of nuclear main pump fault diagnosis model, fault diagnosis method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310059996.2A CN115795313A (en) | 2023-01-16 | 2023-01-16 | Training method of nuclear main pump fault diagnosis model, fault diagnosis method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115795313A true CN115795313A (en) | 2023-03-14 |
Family
ID=85429758
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310059996.2A Pending CN115795313A (en) | 2023-01-16 | 2023-01-16 | Training method of nuclear main pump fault diagnosis model, fault diagnosis method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115795313A (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110837850A (en) * | 2019-10-23 | 2020-02-25 | 浙江大学 | Unsupervised domain adaptation method based on counterstudy loss function |
| CN111652121A (en) * | 2020-06-01 | 2020-09-11 | 腾讯科技（深圳）有限公司 | Training method of expression migration model, and expression migration method and device |
| CN112200114A (en) * | 2020-10-20 | 2021-01-08 | 上海明略人工智能(集团)有限公司 | Fault diagnosis model training method and device, electronic equipment and storage medium |
| CN114048568A (en) * | 2021-11-17 | 2022-02-15 | 大连理工大学 | Rotating machine fault diagnosis method based on multi-source migration fusion contraction framework |
| CN114491823A (en) * | 2022-03-28 | 2022-05-13 | 西南交通大学 | Train bearing fault diagnosis method based on improved generation countermeasure network |
| CN115345255A (en) * | 2022-08-31 | 2022-11-15 | 北京理工大学 | Fault diagnosis method, control device, terminal and storage medium |
| CN115600150A (en) * | 2022-09-26 | 2023-01-13 | 郑州大学（Cn） | Multi-mode gearbox fault diagnosis method based on deep migration learning |
-
2023
- 2023-01-16 CN CN202310059996.2A patent/CN115795313A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110837850A (en) * | 2019-10-23 | 2020-02-25 | 浙江大学 | Unsupervised domain adaptation method based on counterstudy loss function |
| CN111652121A (en) * | 2020-06-01 | 2020-09-11 | 腾讯科技（深圳）有限公司 | Training method of expression migration model, and expression migration method and device |
| CN112200114A (en) * | 2020-10-20 | 2021-01-08 | 上海明略人工智能(集团)有限公司 | Fault diagnosis model training method and device, electronic equipment and storage medium |
| CN114048568A (en) * | 2021-11-17 | 2022-02-15 | 大连理工大学 | Rotating machine fault diagnosis method based on multi-source migration fusion contraction framework |
| CN114491823A (en) * | 2022-03-28 | 2022-05-13 | 西南交通大学 | Train bearing fault diagnosis method based on improved generation countermeasure network |
| CN115345255A (en) * | 2022-08-31 | 2022-11-15 | 北京理工大学 | Fault diagnosis method, control device, terminal and storage medium |
| CN115600150A (en) * | 2022-09-26 | 2023-01-13 | 郑州大学（Cn） | Multi-mode gearbox fault diagnosis method based on deep migration learning |
Non-Patent Citations (1)
| Title |
|---|
| 姚源涛: "基于深度知识迁移的新型核系统诊断预测与容错技术研究" * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111914873A (en) | Two-stage cloud server unsupervised anomaly prediction method | |
| CN111340238A (en) | Fault diagnosis method, device, equipment and storage medium of industrial system | |
| CN113312447B (en) | Semi-supervised log anomaly detection method based on probability label estimation | |
| CN113988126B (en) | Rolling bearing fault diagnosis method based on few-label data feature migration | |
| Bao et al. | A deep transfer learning network for structural condition identification with limited real-world training data | |
| CN113159126B (en) | Industrial Internet of things equipment fault diagnosis method based on universal domain adaptation | |
| CN112199890B (en) | System-level fault diagnosis method for integrated nuclear power device | |
| CN111611867B (en) | Rolling bearing intelligent fault diagnosis method based on multi-classification fuzzy correlation vector machine | |
| CN114048688A (en) | Method for predicting service life of bearing of wind power generator | |
| CN116028876A (en) | Rolling bearing fault diagnosis method based on transfer learning | |
| CN115600150A (en) | Multi-mode gearbox fault diagnosis method based on deep migration learning | |
| CN115017513A (en) | Intelligent contract vulnerability detection method based on artificial intelligence | |
| CN116628592A (en) | Dynamic equipment fault diagnosis method based on improved generation type countering network | |
| CN115345255A (en) | Fault diagnosis method, control device, terminal and storage medium | |
| CN116304820A (en) | Bearing fault type prediction method and system based on multi-source domain transfer learning | |
| Li et al. | Source-free domain adaptation framework for fault diagnosis of rotation machinery under data privacy | |
| Xu et al. | TLS-WGAN-GP: A generative adversarial network model for data-driven fault root cause location | |
| CN115791174B (en) | Rolling bearing abnormality diagnosis method, system, electronic equipment and storage medium | |
| CN117388709A (en) | Energy storage battery fault diagnosis method based on whale optimization nuclear extreme learning machine | |
| CN115795313A (en) | Training method of nuclear main pump fault diagnosis model, fault diagnosis method and system | |
| CN112731890A (en) | Power plant equipment fault detection method and device | |
| CN116167438A (en) | Transformer fault diagnosis method based on improved quantum genetic algorithm optimized SVM | |
| CN113076217B (en) | Disk fault prediction method based on domestic platform | |
| CN114356778A (en) | Deep reinforcement learning software testing method based on coverage guide fuzzy test | |
| CN113723592A (en) | Fault diagnosis method based on wind power gear box monitoring system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20230314 |
|
| RJ01 | Rejection of invention patent application after publication |