Nuclear power plant composite fault diagnosis method based on data augmentation and residual error capsule network

Abstract

The invention provides a nuclear power plant composite fault diagnosis method based on data augmentation and a residual error capsule network, and the method comprises the following steps: S1, collecting characteristic parameters in a historical operation process of a nuclear power plant, carrying out the marking of working condition information, and storing the working condition information in a fault database; s2, initializing the characteristic parameters in the database by adopting a data augmentation method based on Gaussian noise expansion; s3, constructing a composite fault diagnosis model based on a residual capsule network, and realizing mapping from input features to fault types; and S4, combining the trained fault diagnosis model with a main control platform of the nuclear power plant, carrying out on-line testing, and giving a diagnosis result. According to the method, the sample feature domain is widened by adopting the data augmentation method based on Gaussian noise expansion, so that the fault diagnosis model can be better generalized to various operation conditions, and the phenomena of misjudgment, missing judgment and the like are reduced.

Description

Technical Field

[0001] The present invention relates to the field of power plant fault diagnosis, and in particular to a nuclear power plant composite fault diagnosis method and system based on data augmentation and residual capsule network. Background Art

[0002] At present, fault diagnosis models based on artificial intelligence algorithms have been applied in many fields, and algorithms have been applied in nuclear power plants, such as fault classification based on convolutional neural networks and state parameter prediction based on long short-term memory networks. However, most nuclear power plant fault diagnosis methods are based on offline training models to diagnose known single faults. In the real-time operation of nuclear power plants, there may be multiple compound faults caused by a single fault, or unknown operating states caused by operating condition changes. If compound faults and unknown faults cannot be detected, false alarms and missed alarms will occur, seriously affecting the accuracy of the nuclear power plant fault diagnosis model.

[0003] In the Chinese patent document with publication number CN115436057A, a comparative capsule network method for intelligent diagnosis of wheel bearings under high noise is disclosed, which includes: collecting vibration signals of train wheel bearings; establishing a comparative capsule network to filter and screen characteristic signals under high noise to extract more representative information; transmitting the collected bearing signals to the network model, which can identify single faults such as inner ring fault and outer ring fault of bearings, composite faults of outer ring + inner ring, composite faults of inner ring + rolling element, etc. Although this patent document can perform composite fault diagnosis on train wheel bearings, it cannot be applied to more complex system-level fault diagnosis, and cannot be applied to unknown states caused by changes in working conditions. Summary of the invention

[0004] In view of the defects in the prior art, the object of the present invention is to provide a nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network.

[0005] According to the present invention, a nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network includes the following steps:

[0006] Step S1: Collect characteristic parameters during the historical operation of the nuclear power plant, mark the operating condition information, and store them in the fault database;

[0007] Step S2: Initialize the feature parameters in the database using a data augmentation method based on Gaussian noise expansion;

[0008] Step S3: construct a composite fault diagnosis model based on the residual capsule network to achieve mapping from input features to fault types;

[0009] Step S4: Combine the trained fault diagnosis model with the main control platform of the nuclear power plant, conduct online testing, and provide diagnosis results.

[0010] Preferably, the characteristic parameters include: all analog quantity characteristics under the operating state of the nuclear power plant, and different data samples are divided using a sliding window method; each data sample contains the time series information of all analog quantities, reflecting the operating characteristics under the corresponding working conditions.

[0011] Preferably, in step S2, the signal-to-noise ratio SNR is used to measure the intensity of the added Gaussian noise, and its calculation formula is as follows:

[0012]

[0013] Where P s is the power of the signal, P n is the power of the noise, in dB;

[0014] The step S2 further includes: performing data normalization processing on all features after initialization, and recording a normalization information table of each parameter; the normalization processing includes:

[0015] Each feature is linearly transformed to the range of [0,1], and the calculation formula is as follows:

[0016]

[0017] In the formula, x is the initial eigenvalue, x * is the normalized eigenvalue, x max is the maximum value of the eigenvalue, x min is the minimum value of this eigenvalue.

[0018] Preferably, the composite fault diagnosis model uses Resnet18 as a feature extractor and a capsule neural network as a composite fault classifier, wherein Resnet18 includes 1 convolutional layer for input and 8 residual connection blocks, and the last fully connected layer is replaced by a capsule neural network layer.

[0019] Preferably, the composite fault diagnosis model uses the marginal classification loss function as the cost function for training, and its calculation formula is:

[0020]

[0021] Where J represents the marginal loss function value, L c represents the marginal loss function value of the c-th sample, C is the number of all known fault types, T c Represents the category indicator function, that is, when T c =1, indicating that the current sample has a type c fault, T c=0 means that no type c fault occurs, m + With m - They represent the upper and lower bounds of the reliability confidence, λ is the weight of the non-fault category loss, Indicates the confidence that the model predicts that the sample belongs to the cth class.

[0022] Preferably, the training method of the composite fault diagnosis model includes:

[0023] The feature parameters are divided into training data set, validation data set, and test data set in a ratio of 8:1:1;

[0024] The training data set uses label information to train the composite fault diagnosis model, and uses the verification data set to verify in each training step, and saves the model with the highest accuracy in real time. After all training steps are completed, the saved model is used to verify the diagnosis effect on the test data set.

[0025] Preferably, the online test in step S4 includes: acquiring online operation status parameters collected by the on-site DCS system in real time, preprocessing all the parameters through a normalized information table, and then transferring them to the composite fault diagnosis model for testing.

[0026] Preferably, the diagnosis result includes:

[0027] - Check that the nuclear power plant is currently in normal condition and the units are operating normally, so no action is taken;

[0028] -Detect whether the nuclear power plant is currently in a known single fault or known compound fault, give the current fault type, assist the operator, and record the fault data to the database;

[0029] -Detect that the nuclear power plant is currently in an unknown fault, prompt the operator to check the operating status of the unit, record the unknown fault data in the cache area, and update it to the existing composite fault diagnosis model in an incremental learning manner after being marked by experts, so as to realize the continuous updating of the model.

[0030] A nuclear power plant composite fault diagnosis system based on data augmentation and residual capsule network provided by the present invention includes the following modules:

[0031] Module M1: collects characteristic parameters of the nuclear power plant during historical operation, annotates the operating information, and stores it in the fault database;

[0032] Module M2: Initialize the feature parameters in the database using a data augmentation method based on Gaussian noise expansion;

[0033] Module M3: Construct a composite fault diagnosis model based on residual capsule network to achieve mapping from input features to fault types;

[0034] Module M4: Combine the trained fault diagnosis model with the main control platform of the nuclear power plant, conduct online testing, and provide diagnostic results.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention broadens the sample feature domain by adopting a data augmentation method based on Gaussian noise expansion, so that the fault diagnosis model can be better generalized to a variety of operating conditions, thereby reducing the phenomenon of misjudgment and missed judgment;

[0037] 2. The present invention uses a fault classifier based on a residual capsule network to simultaneously detect single faults, compound faults, and unknown faults in a nuclear power plant, providing operators with more detailed and distinguishable diagnostic results. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0039] Figure 1 It is a fault diagnosis flow chart disclosed in the present invention;

[0040] Figure 2 This is a schematic diagram of data augmentation in the present invention;

[0041] Figure 3 It is a schematic diagram of the dynamic routing algorithm flow in the present invention;

[0042] Figure 4 Schematic diagram of the capsule neural network structure in the present invention;

[0043] Figure 5 It is a schematic diagram of the structure of the fault diagnosis model in the present invention;

[0044] Figure 6 This is a diagram of the diagnosis results of the fault diagnosis model in the present invention; DETAILED DESCRIPTION

[0045] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0046] The present invention discloses a nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network. Figure 1 As shown, the following steps are included:

[0047] Step S1: Collect characteristic parameters during the historical operation of the nuclear power plant, where the analog quantity characteristics are divided into a series of characteristic matrices using a sliding window method with a fixed window width, and the operating condition information is labeled and stored in the fault database in the form of labeled samples.

[0048] Among them, step S1 uses all analog quantity characteristics under the operation state of the nuclear power plant, and uses the sliding window method to divide different data samples, so that a large number of samples can be obtained in a single operation of the unit. At the same time, each sample contains the time series information of all analog quantities, which can fully reflect the operation characteristics under the corresponding working conditions, and is more conducive to the fault diagnosis model to analyze and learn the fault mechanism and judge the operation state.

[0049] Step S2: Initialize the parameter features in the database using a data augmentation method based on Gaussian noise expansion, then perform data normalization on all features, record the normalization information table of each parameter, and divide the training data set, validation data set, and test data set into a ratio of 8:1:1. Figure 2 shown.

[0050] The actual measurement of nuclear power plant operating parameters may be affected by various types and degrees of noise, and the noise response may be aggravated under complex fault conditions. The introduction of stationary additive Gaussian noise can enhance the robustness of the model and avoid noise interference. At the same time, the sample feature domain after using the data augmentation method is wider, which can enhance the adaptability of the model to unknown operating states and reduce the possibility of false alarms or misreporting in the online detection stage.

[0051] Specifically, the signal-to-noise ratio (SNR) is used to measure the intensity of the added Gaussian noise, and its calculation formula is as follows:

[0052]

[0053] Where P s is the power of the signal, P n is the noise power in dB.

[0054] Specifically, data normalization is performed for each feature parameter separately, and all features are linearly transformed to the range of [0,1] to speed up the training and convergence of the fault diagnosis model. The calculation formula is as follows:

[0055]

[0056] In the formula, x is the initial eigenvalue, x * is the normalized eigenvalue, x max is the maximum value of the eigenvalue, x min is the minimum value of this eigenvalue.

[0057] The training data set uses label information to train the fault diagnosis model, and uses the validation set data for verification in each training step. The model with the highest accuracy in the validation set is saved in real time. After all training steps are completed, the saved model is used to test the diagnostic effect on the test set. This data set division method can effectively avoid underfitting and overfitting problems.

[0058] Step S3: Construct a composite fault diagnosis model based on residual capsule network, in which Resnet18 is used as the feature extractor and capsule network is used as the composite fault classifier, where Resnet18 contains 1 convolutional layer for input and 8 residual connection blocks, and the last fully connected layer is replaced by a capsule neural network layer to achieve mapping from input features to fault types.

[0059] The present invention uses Resnet18 as a feature extractor, has a specific network structure, and is suitable for input samples of different sizes. Therefore, the fault diagnosis model is more universal.

[0060] The present invention uses capsule neural network to replace the fully connected neural network of Resnet18 as the fault classification layer, and uses vectors as neurons in the network for calculation, which is no longer in the traditional scalar form. Therefore, the capsule network contains more feature information; the capsule network updates the capsule layer parameters through a unique dynamic routing mechanism, increases the coupling coefficient between neuron nodes, and uses the idea of ​​vector inner product to make the network model output as close to the label information as possible; the output vector form is:

[0061] y=(φ0,φ1,φ2,…,φ n ),

[0062] Where each scalar φ i Corresponding to whether the i-th single fault occurs, the model can directly output the decoupling result of the composite fault. When the output vector is a full 0 vector, or the output vector form is unreasonable, if the model detects normal operation and fault operation at the same time, it means that the nuclear power plant is in an unknown operating condition at this time. Therefore, the fault diagnosis model has the ability of open set identification and detection. Figure 3 and Figure 4 shown.

[0063] The present invention also trains the network model by designing a marginal classification loss function as a cost function. Different from the cross entropy loss function, the marginal loss function measures the similarity of different categories based on the Euclidean distance, effectively expanding the difference between classes and reducing the degree of intra-class dispersion. The calculation formula is as follows:

[0064]

[0065] Where J represents the marginal loss function value, L crepresents the marginal loss function value of the c-th sample, C is the number of all known fault types, T c Represents the category indicator function, that is, when T c =1, indicating that the current sample has a type c fault, T c =0 means that no type c fault occurs, m + With m - They represent the upper and lower bounds of the reliability confidence, and λ is the weight of the non-fault category loss. This loss function can promote the output prediction vector of the fault diagnosis model to have a real fault handling confidence of no less than m. + , and the reliability of the non-existent fault handling is no higher than m - , Indicates the confidence that the model predicts that the sample belongs to the cth class.

[0066] Step S4: Combine the fault diagnosis model trained offline with the main control platform of the nuclear power plant to obtain the online operation status parameters collected by the on-site DCS system in real time. After all parameters are pre-processed in the normalized information table, they are transferred to the fault diagnosis model and tested online to give the diagnosis results to assist the operator in operation. Figure 5 and Figure 6 shown.

[0067] According to the fault diagnosis model structure provided by the present invention, there are three types of outputs in the online diagnosis process:

[0068] 1. The nuclear power plant is currently in normal condition and the units are operating normally, so no action is taken;

[0069] 2. Detect that the nuclear power plant is currently in a known single fault or known compound fault, give the current fault type, assist the operator, and record the fault data in the database;

[0070] 3. Detect that the nuclear power plant is currently in an unknown fault, prompt the operator to check the operating status of the unit, record the unknown fault data in the cache area, and update it to the existing fault diagnosis model in an incremental learning manner after being marked by experts, so as to realize the continuous updating of the model.

[0071] The present invention also provides a nuclear power plant composite fault diagnosis system based on data augmentation and residual capsule networks. The nuclear power plant composite fault diagnosis system based on data augmentation and residual capsule networks can be realized by executing the process steps of the nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule networks, that is, technical personnel in this field can understand the nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule networks as a preferred implementation of the nuclear power plant composite fault diagnosis system based on data augmentation and residual capsule networks.

[0072] The present invention discloses a nuclear power plant composite fault diagnosis system based on data augmentation and residual capsule network, comprising the following modules:

[0073] Module M1: collects characteristic parameters of the nuclear power plant during historical operation, annotates the operating information, and stores it in the fault database;

[0074] Module M2: Initialize the feature parameters in the database using a data augmentation method based on Gaussian noise expansion;

[0075] Module M3: Construct a composite fault diagnosis model based on residual capsule network to achieve mapping from input features to fault types;

[0076] Module M4: Combine the trained fault diagnosis model with the main control platform of the nuclear power plant, conduct online testing, and provide diagnostic results.

[0077] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0078] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network, characterized in that: The following steps are involved: Step S1: Collect characteristic parameters during the historical operation of the nuclear power plant, mark the operating condition information, and store them in the fault database; Step S2: Initialize the feature parameters in the database using a data augmentation method based on Gaussian noise expansion; Step S3: construct a composite fault diagnosis model based on the residual capsule network to achieve mapping from input features to fault types; Step S4: Combine the trained fault diagnosis model with the main control platform of the nuclear power plant, conduct online testing, and provide diagnosis results.

2. The nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network according to claim 1 is characterized in that: The characteristic parameters include: all analog quantity characteristics under the operating state of the nuclear power plant, and different data samples are divided using a sliding window method; each data sample contains the time series information of all analog quantities, reflecting the operating characteristics under the corresponding working conditions.

3. The nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network according to claim 1 is characterized in that: In step S2, the signal-to-noise ratio (SNR) is used to measure the intensity of the added Gaussian noise, and its calculation formula is as follows: Where P s is the power of the signal, P n is the power of the noise, in dB; The step S2 further includes: performing data normalization processing on all features after initialization, and recording a normalization information table of each parameter; the normalization processing includes: Each feature is linearly transformed to the range of [0,1], and the calculation formula is as follows: In the formula, x is the initial eigenvalue, x * is the normalized eigenvalue, x max is the maximum value of the eigenvalue, x min is the minimum value of this eigenvalue.

4. The nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network according to claim 1 is characterized in that: The composite fault diagnosis model adopts Resnet18 as a feature extractor and a capsule neural network as a composite fault classifier, wherein Resnet18 includes 1 convolutional layer for input and 8 residual connection blocks, and the last fully connected layer is replaced by a capsule neural network layer.

5. The nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network according to claim 1, characterized in that: The composite fault diagnosis model uses the marginal classification loss function as the cost function for training, and its calculation formula is: Where J represents the marginal loss function value, L c represents the marginal loss function value of the c-th sample, C is the number of all known fault types, T c Represents the category indicator function, that is, when T c =1, indicating that the current sample has a type c fault, T c =0 means that no type c fault occurs, m + With m - They represent the upper and lower bounds of the reliability confidence, λ is the weight of the non-fault category loss, Indicates the confidence that the model predicts that the sample belongs to the cth class.

6. The nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network according to claim 1, characterized in that: The training method of the composite fault diagnosis model includes: The feature parameters are divided into training data set, validation data set, and test data set in a ratio of 8:1:1; The training data set uses label information to train the composite fault diagnosis model, and uses the verification data set to verify in each training step, and saves the model with the highest accuracy in real time. After all training steps are completed, the saved model is used to verify the diagnosis effect on the test data set.

7. The nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network according to claim 1 is characterized in that: The online test in step S4 includes: acquiring the online operation status parameters collected by the on-site DCS system in real time, preprocessing all the parameters through a normalized information table, and then transferring them to the composite fault diagnosis model for testing.

8. The nuclear power plant composite fault diagnosis method based on data augmentation and residual capsule network according to claim 1, characterized in that: The diagnostic results include: - Check that the nuclear power plant is currently in normal condition and the units are operating normally, so no action is taken; -Detect whether the nuclear power plant is currently in a known single fault or known compound fault, give the current fault type, assist the operator, and record the fault data to the database; -Detect that the nuclear power plant is currently in an unknown fault, prompt the operator to check the operating status of the unit, record the unknown fault data in the cache area, and update it to the existing composite fault diagnosis model in an incremental learning manner after being marked by experts, so as to realize the continuous updating of the model.

9. A nuclear power plant composite fault diagnosis system based on data augmentation and residual capsule network, characterized in that: Includes the following modules: Module M1: collects characteristic parameters of the nuclear power plant during historical operation, annotates the operating information, and stores it in the fault database; Module M2: Initialize the feature parameters in the database using a data augmentation method based on Gaussian noise expansion; Module M3: Construct a composite fault diagnosis model based on residual capsule network to achieve mapping from input features to fault types; Module M4: Combine the trained fault diagnosis model with the main control platform of the nuclear power plant, conduct online testing, and provide diagnostic results.

Images