Nuclear power steam turbine vibration fault diagnosis method, device, system and storage medium

Abstract

The present invention discloses a method, device, system, and storage medium for diagnosing vibration faults in nuclear-powered steam turbines, including: step S1, obtaining historical operating data of nuclear-powered steam turbines in a nuclear power plant; step S2, obtaining a privacy-protected federated deep domain adaptation-based fault diagnosis model based on the historical operating data; and step S3, inputting real-time nuclear power plant operating data into the privacy-protected federated deep domain adaptation-based fault diagnosis model to perform nuclear-powered steam turbine vibration fault diagnosis. The technical solution of the present invention enables rapid identification of fault categories while maintaining privacy, enabling maintenance personnel and systems to quickly respond to faults.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear power plant fault diagnosis technology, and in particular relates to a method and device, system, and storage medium for diagnosing vibration faults of nuclear power steam turbines. Background Art

[0002] With growing environmental awareness, sustainable manufacturing has become a crucial component across all industries. The safe and reliable operation and maintenance of nuclear steam turbines is crucial, requiring agile and accurate response and diagnosis of equipment fault signals. Due to safety requirements, nuclear power plants strictly isolate operational data, creating de facto data silos. Furthermore, insufficient fault sample data makes accurate fault diagnosis difficult. Establishing stable and reliable nuclear steam turbine vibration fault diagnosis across different nuclear power plants and nuclear equipment is a pressing issue. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a nuclear power steam turbine vibration fault diagnosis method and device, system, and storage medium, which can quickly identify the fault category while considering privacy protection, so that maintenance personnel and systems can respond quickly to the fault.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for diagnosing vibration faults of a nuclear power steam turbine, comprising:

[0006] Step S1, obtaining historical operating data of a nuclear power steam turbine in a nuclear power plant;

[0007] Step S2: obtaining a fault diagnosis model based on federated deep domain adaptation and considering privacy protection based on historical operating data of nuclear-powered steam turbines in a nuclear power plant;

[0008] Step S3: input the real-time nuclear power plant operation data into a fault diagnosis model based on federated deep domain adaptation with privacy protection in mind to perform nuclear power steam turbine vibration fault diagnosis.

[0009] Preferably, the nuclear power plant includes: a source nuclear power plant and a target nuclear power plant, wherein the source nuclear power plant includes multiple source nuclear-powered steam turbines and the target nuclear power plant includes multiple target nuclear-powered steam turbines; when the nuclear power plant initiates a request with a preset type of nuclear-powered steam turbine as the target, the central server will select the source nuclear-powered steam turbine from its subordinate nuclear power plants.

[0010] Preferably, step S2 includes:

[0011] Based on the historical operating data of nuclear power steam turbines in nuclear power plants, a BiLSTM-CNN-based feature extractor is used to extract deep temporal features from the fault diagnosis signals of nuclear power steam turbines, and the equipment characteristics of the source and target nuclear power steam turbines are obtained.

[0012] According to the equipment characteristics of the source and target nuclear-powered steam turbines, a fault diagnosis model is obtained through local knowledge transfer of domain adversarial training and an adaptive weight mechanism for dynamic federated learning.

[0013] The present invention also provides a nuclear power steam turbine vibration fault diagnosis device, comprising:

[0014] An acquisition module is used to obtain historical operating data of nuclear power steam turbines in nuclear power plants;

[0015] A training module is used to obtain a privacy-preserving federated deep domain adaptation-based fault diagnosis model based on historical operating data of nuclear power turbines in nuclear power plants;

[0016] The fault diagnosis module is used to input real-time nuclear power plant operation data into a privacy-protected federated deep domain adaptation-based fault diagnosis model to diagnose nuclear power turbine vibration faults.

[0017] Preferably, the nuclear power plant includes: a source nuclear power plant and a target nuclear power plant, wherein the source nuclear power plant includes multiple source nuclear-powered steam turbines and the target nuclear power plant includes multiple target nuclear-powered steam turbines; when the nuclear power plant initiates a request with a preset type of nuclear-powered steam turbine as the target, the central server will select the source nuclear-powered steam turbine from its subordinate nuclear power plants.

[0018] Preferably, the training device comprises:

[0019] An extraction unit is used to extract deep temporal features from the fault diagnosis signals of nuclear steam turbines in nuclear power plants using a BiLSTM-CNN-based feature extractor based on historical operating data of the nuclear steam turbines to obtain equipment features of the source and target nuclear steam turbines;

[0020] The training unit is used to obtain a fault diagnosis model based on the equipment characteristics of the source nuclear-powered steam turbine and the target nuclear-powered steam turbine through local knowledge transfer of domain adversarial training and an adaptive weight mechanism for dynamic federated learning.

[0021] An embodiment of the present invention also provides a nuclear power steam turbine vibration fault diagnosis system, comprising: a memory and a processor, wherein the memory stores a computer program run by the processor, and the computer program executes a nuclear power steam turbine vibration fault diagnosis method when run by the processor.

[0022] An embodiment of the present invention further provides a storage medium having a computer program stored thereon. The computer program executes a method for diagnosing vibration faults of a nuclear power steam turbine when the computer program is run.

[0023] This invention utilizes a privacy-preserving federated deep domain adaptation-based fault diagnosis model. Feature extraction is performed locally at both the source and target nuclear power plants, enabling secure feature sharing without compromising data privacy. Domain adversarial training is integrated into local model training to transfer vibration fault diagnosis knowledge. Furthermore, an adaptive weighting mechanism is designed to facilitate adaptive adjustment of model weights during federated aggregation. This technical solution enables rapid identification of fault categories while maintaining privacy, enabling maintenance personnel and systems to respond quickly to faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0025] Figure 1 This is a flow chart of a method for diagnosing vibration faults in a nuclear power steam turbine according to an embodiment of the present invention;

[0026] Figure 2 Schematic diagram of a method for diagnosing vibration faults in a nuclear power steam turbine according to an embodiment of the present invention;

[0027] Figure 3 Schematic diagram of the dynamic adversarial adaptive module DAAM. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] Example 1:

[0031] like Figure 1 、 2 As shown, an embodiment of the present invention provides a method for diagnosing vibration faults of a nuclear power steam turbine, comprising:

[0032] Step S1, obtaining historical operating data of a nuclear power steam turbine in a nuclear power plant;

[0033] Step S2: obtaining a fault diagnosis model based on federated deep domain adaptation and considering privacy protection based on historical operating data of nuclear-powered steam turbines in a nuclear power plant;

[0034] Step S3: input the real-time nuclear power plant operation data into a fault diagnosis model based on federated deep domain adaptation with privacy protection in mind to perform nuclear power steam turbine vibration fault diagnosis.

[0035] As one implementation of an embodiment of the present invention, a nuclear power plant includes a source nuclear power plant (SNPP) and a target nuclear power plant (TNPP). The source nuclear power plant includes multiple source nuclear power steam turbines (SSTs), and the target nuclear power plant includes multiple target nuclear power steam turbines (TSTs). When a nuclear power plant initiates a request targeting a preset type of nuclear power steam turbine, a central server selects a source nuclear power steam turbine from its subordinate nuclear power plants. The nuclear power steam turbines in each SNPP are similar devices operating under different operating conditions, and the datasets for these devices have the same label space and different marginal distributions.

[0036] As an implementation of the embodiment of the present invention, step S2 includes:

[0037] Step S21: Based on historical operating data of nuclear power steam turbines in a nuclear power plant, a feature extractor based on BiLSTM-CNN is used to extract deep temporal features from fault diagnosis signals of the nuclear power steam turbines to obtain equipment features of the source nuclear power steam turbine and the target nuclear power steam turbine;

[0038] Step S22: According to the equipment characteristics of the source nuclear-powered steam turbine and the target nuclear-powered steam turbine, a fault diagnosis model is obtained through local knowledge transfer of domain adversarial training and an adaptive weight mechanism for dynamic federated learning.

[0039] Furthermore, in step S21, the BiLSTM-CNN-based feature extractor consists of two BiLSTM layers, one LSTM layer, and two convolutional layers. Batch normalization (BN) is implemented after the convolutional layers to accelerate feature extraction convergence. Furthermore, a rectified linear unit (ReLU) activation function is designed after the BN layer. Together with the BiLSTM, it can alleviate the vanishing and exploding gradient problems.

[0040] For the SST data sample of the i-th SNPP Using a parameter θ f，i The feature extractor G f，i To obtain the features in the source domain Right now:

[0041]

[0042] Accordingly, the features in the target domain are obtained in TST. The only transmission between SNPP and TNPP is the vibration fault feature of nuclear power steam turbine.

[0043] Furthermore, in step S21, when a new device is put into use, it will not generate enough labeled data sets during its short-term operation, and will face the problem of high cost and long consumption of labeling these data. In order to bridge the gap between historical equipment operation data and new equipment operation data, a dynamic adversarial adaptation method is introduced to learn transferable features in nuclear power steam turbine vibration fault diagnosis, and a DAAM module (dynamic adversarial adaptive module) for FL (federated learning) is proposed, such as Figure 3 The DAAM module is based on the idea of ​​Domain Adversarial Network (DAN) and achieves local knowledge transfer by leveraging domain adversarial training.

[0044] Furthermore, each SNPP will use the local SSTs dataset to train the DAAM module. The DAAM module consists of a label classifier G y,i , global domain discriminator G d,i and local subdomain discriminator When performing local knowledge transfer, it is necessary to consider not only the differences in domain edge distributions between vibration fault categories (subdomains), but also the differences in conditional distributions. The training process of the SSTs dataset of the i-th SNPP is as follows:

[0045] In the label classifier, the source features extracted by the feature extractor are used and its corresponding tags As input, the label classifier is trained. The loss function of the label classifier is the cross entropy loss, which is defined as follows:

[0046]

[0047] Among them, G y,i and θ y,j Represent the label classifier and its parameters composed of the fully connected layer, express The corresponding label.

[0048] The loss function calculation formula of the global domain discriminator is as follows:

[0049]

[0050] Among them, G d,i and θ d,i are the global domain discriminator and its parameters, f i,j Depend on and Composition, p j Represents input The domain label (if f i,j From Then p j =0; if from Then p j =1).

[0051] In order to achieve more fine-grained domain adaptation in the vibration fault knowledge domain, a local subdomain discriminator is used to align the conditional distribution. For the vibration fault problem with n categories, the loss function of the local subdomain discriminator is defined as follows:

[0052]

[0053] in, and Represent the rth local domain discriminator and its parameters, is f i,j The corresponding predicted probability distribution.

[0054] In summary, the total loss function of the DAAM module consists of the label classifier loss, the global domain discriminator loss, and the local subdomain discriminator loss, and its expression is as follows:

[0055] Loss i =L y,i -β((1-ω i )L g,i +ω i L l,i )

[0056] Among them, β is a trade-off parameter used to adjust Loss i The global domain discriminator loss L g,i and the local subdomain discriminator loss L l,i The weight, ω i is a dynamic adversarial factor that is used to evaluate the importance between the marginal distribution and the conditional distribution.

[0057] According to the SSTs dataset of i SNPP TST dataset with TNPP The marginal distribution difference and conditional distribution difference between them, dynamic adversarial factor ω i During the iterative learning process, the weights of the global domain discriminator loss and the local subdomain discriminator loss in the domain loss will be automatically adjusted. The estimation formula is as follows:

[0058]

[0059] in, and Respectively and The rth category, and They are the A distances of the global domain discriminator and the local subdomain discriminator, respectively. The calculation formula is as follows:

[0060]

[0061] After the optimization target is determined, the total loss function of DAAM can be updated as:

[0062]

[0063] Among them, Θ i Represents all the parameters that DAAM needs to train, represented by θ y,i ,θ d,i , composition.

[0064] Furthermore, in step S22, in nuclear turbine vibration fault diagnosis, the datasets held by each SNPP differ significantly in quality and quantity. On the one hand, the performance of classifiers trained on the SST dataset varies; on the other hand, there are differences in marginal distributions and conditional distributions between the SST (source domain) and TST (target domain) datasets. The present invention employs a real-time adaptive weighting mechanism for dynamic federated learning. The central server comprehensively considers the classification performance and domain adaptability of the SNPPs and dynamically adjusts the weight of each SNPP in the global model.

[0065] Furthermore, in nuclear power plant vibration fault diagnosis, datasets from various SNPPs exhibit significant differences in quality and quantity. On the one hand, classifiers trained on SST datasets exhibit varying performance; on the other hand, there are differences in marginal and conditional distributions between SST (source domain) and TST (target domain) datasets. To address the problem studied in this paper, we designed an adaptive weighting mechanism for dynamic federated learning. The central server dynamically adjusts the weight of each SNPP in the global model, taking into account both the model's classification performance and the domain adaptability of the SNPPs.

[0066] In the process of local knowledge transfer, the DAAM loss of the local model is i Including the classifier loss L y,i , global domain discriminator loss L g,i and the local subdomain discriminator loss L l,i . Classifier loss L y,i It can measure the classification performance of the model, and the global domain discriminator loss L g,i and the local subdomain discriminator loss L l,iThe domain adaptation ability of the model can be evaluated. Therefore, the designed weight adaptation mechanism uses the local model loss Loss i Calculate the ratio of the local model to the global model.

[0067] Through dynamic federated learning, the central server extracts N from all SNPPs according to the extraction ratio Frac SC source clients participate in global training. g The global training, the local model loss of the i-th SNPP is Loss i , the similarity between the local model of the i-th SNPP and the TST target model is defined as:

[0068]

[0069] Therefore, the similarity ζ i The weights in the global model are expressed as:

[0070]

[0071] Among them, N SC is the number of SNPPs selected according to the extraction ratio Frac.

[0072] In the i g After global training, the model of each SNPP and similarity The data will be transmitted to the central server. The central server calculates the weight of each model in the global model based on the obtained similarity and aggregates the models through the adaptive weight mechanism. The calculation formula is as follows:

[0073]

[0074] in, Represents the local knowledge obtained after transfer

[0075] Example 2:

[0076] An embodiment of the present invention further provides a nuclear power steam turbine vibration fault diagnosis device, comprising:

[0077] An acquisition module is used to obtain historical operating data of nuclear power steam turbines in nuclear power plants;

[0078] A training module is used to obtain a privacy-preserving federated deep domain adaptation-based fault diagnosis model based on historical operating data of nuclear power turbines in nuclear power plants;

[0079] The fault diagnosis module is used to input real-time nuclear power plant operation data into a privacy-protected federated deep domain adaptation-based fault diagnosis model to diagnose nuclear power turbine vibration faults.

[0080] As an implementation method of an embodiment of the present invention, a nuclear power plant includes: a source nuclear power plant and a target nuclear power plant, wherein the source nuclear power plant includes multiple source nuclear-powered steam turbines, and the target nuclear power plant includes multiple target nuclear-powered steam turbines; when the nuclear power plant initiates a request with a preset type of nuclear-powered steam turbine as the target, the central server will select the source nuclear-powered steam turbine from its subordinate nuclear power plants.

[0081] As an implementation of an embodiment of the present invention, the training device includes:

[0082] An extraction unit is used to extract deep temporal features from the fault diagnosis signals of nuclear steam turbines in nuclear power plants using a BiLSTM-CNN-based feature extractor based on historical operating data of the nuclear steam turbines to obtain equipment features of the source and target nuclear steam turbines;

[0083] The training unit is used to obtain a fault diagnosis model based on the equipment characteristics of the source nuclear-powered steam turbine and the target nuclear-powered steam turbine through local knowledge transfer of domain adversarial training and an adaptive weight mechanism for dynamic federated learning.

[0084] Example 3:

[0085] An embodiment of the present invention also provides a nuclear power steam turbine vibration fault diagnosis system, comprising: a memory and a processor, wherein the memory stores a computer program run by the processor, and the computer program executes a nuclear power steam turbine vibration fault diagnosis method when run by the processor.

[0086] Example 4:

[0087] An embodiment of the present invention further provides a storage medium having a computer program stored thereon. The computer program executes a method for diagnosing vibration faults of a nuclear power steam turbine when the computer program is run.

[0088] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for diagnosing vibration faults of a nuclear power steam turbine, characterized in that: include: Step S1, obtaining historical operating data of a nuclear power steam turbine in a nuclear power plant; Step S2: obtaining a fault diagnosis model based on federated deep domain adaptation and considering privacy protection based on historical operating data of nuclear-powered steam turbines in a nuclear power plant; Step S3: inputting the real-time nuclear power plant operation data into a fault diagnosis model based on federated deep domain adaptation with privacy protection into the nuclear power steam turbine vibration fault diagnosis; The nuclear power plant includes: a source nuclear power plant SNPP and a target nuclear power plant TNPP, wherein the source nuclear power plant includes multiple source nuclear power steam turbines SST, and the target nuclear power plant includes multiple target nuclear power steam turbines TST; when the nuclear power plant initiates a request with a preset type of nuclear power steam turbine as the target, the central server will select the source nuclear power steam turbine from its subordinate nuclear power plants; Step S2 includes: Based on the historical operating data of nuclear power steam turbines in nuclear power plants, a BiLSTM-CNN-based feature extractor is used to extract deep temporal features from the fault diagnosis signals of nuclear power steam turbines, and the equipment characteristics of the source and target nuclear power steam turbines are obtained. Based on the equipment characteristics of the source and target nuclear power steam turbines, a fault diagnosis model is obtained through local knowledge transfer of domain adversarial training and an adaptive weight mechanism for dynamic federated learning. In the process of local knowledge transfer, the DAAM loss of the local model is i Including the classifier loss L y,i , global domain discriminator loss L g,i and the local subdomain discriminator loss L l,i ; Classifier loss L y,i Measure the classification performance of the model, the global domain discriminator loss L g,i and the local subdomain discriminator loss L l,i Evaluate the domain adaptability of the model; the adaptive weight mechanism uses local model loss Loss i Calculate the proportion of the local model in the global model; Through dynamic federated learning, the central server extracts N from all SNPPs according to the extraction ratio Frac SC source clients participate in global training; for the i-th g The global training, the local model loss of the i-th SNPP is Loss i , the similarity between the local model of the i-th SNPP and the TST target model is defined as: Therefore, the similarity ζ i The weights in the global model are expressed as: Among them, N SC is the number of SNPPs selected according to the extraction ratio Frac; In the i g After global training, the model of each SNPP and similarity The data will be transmitted to the central server; the central server calculates the weight of each model in the global model based on the obtained similarity and aggregates the models through the adaptive weight mechanism. The calculation formula is as follows: in, Represents the local knowledge obtained after transfer 2. A nuclear power steam turbine vibration fault diagnosis device for implementing the nuclear power steam turbine vibration fault diagnosis method according to claim 1, characterized in that: include: An acquisition module is used to obtain historical operating data of nuclear power steam turbines in nuclear power plants; A training module is used to obtain a privacy-preserving federated deep domain adaptation-based fault diagnosis model based on historical operating data of nuclear power turbines in nuclear power plants; A fault diagnosis module is used to input real-time nuclear power plant operation data into a privacy-preserving federated deep domain adaptation-based fault diagnosis model to perform nuclear power steam turbine vibration fault diagnosis; The nuclear power plant includes: a source nuclear power plant and a target nuclear power plant, wherein the source nuclear power plant includes multiple source nuclear power steam turbines, and the target nuclear power plant includes multiple target nuclear power steam turbines; when the nuclear power plant initiates a request with a preset type of nuclear power steam turbine as the target, the central server will select the source nuclear power steam turbine from its subordinate nuclear power plants; The training device includes: An extraction unit is used to extract deep temporal features from the fault diagnosis signals of nuclear steam turbines in nuclear power plants using a BiLSTM-CNN-based feature extractor based on historical operating data of the nuclear steam turbines to obtain equipment features of the source and target nuclear steam turbines; The training unit is used to obtain a fault diagnosis model based on the equipment characteristics of the source nuclear-powered steam turbine and the target nuclear-powered steam turbine through local knowledge transfer of domain adversarial training and an adaptive weight mechanism for dynamic federated learning.

3. A nuclear power steam turbine vibration fault diagnosis system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the nuclear power steam turbine vibration fault diagnosis method according to claim 1 is executed.

4. A storage medium, characterized in that The storage medium stores a computer program, which executes the nuclear power steam turbine vibration fault diagnosis method according to claim 1 when running.

Images