Nuclear power plant variable power fault diagnosis method and system based on adversarial feature gating

Abstract

The application provides a nuclear power plant variable power fault diagnosis method and system based on an adversarial feature gating. The method introduces a discrete adversarial feature gating mechanism, forces the model to only focus on key features that are physically consistent under different working conditions, further combines a consensus voting strategy of multiple rounds of independent operation, eliminates false related features caused by random disturbance or numerical noise, and finally constructs a fault diagnosis model that can maintain high precision under an unseen power level without target domain data through explicit decoupling of feature selection and model training process.

Description

Technical Field

[0001] This invention relates to the field of intelligent safety monitoring and fault diagnosis technology for nuclear power plants, and in particular to a method and system for diagnosing variable power faults in nuclear power plants based on adversarial feature gating. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning models, with their powerful nonlinear fitting and feature extraction capabilities, have been widely used in the field of fault diagnosis of complex industrial systems such as nuclear power plants. However, in actual operation scenarios, the operating conditions of nuclear power units need to be adjusted according to the power grid load demand, resulting in frequent changes in the system's operating state. This leads to a "distribution offset" problem between the data collected under certain fixed power level conditions and the data under actual operating conditions when the model is deployed.

[0003] This distribution shift causes a high-precision diagnostic model trained under specific operating conditions to experience a sharp decline in diagnostic accuracy when dealing with new operating conditions, especially those with different power levels, severely weakening the model's practicality and reliability. To alleviate this problem, domain-adaptive transfer learning methods have been widely studied. Patent CN119668235B (application number: 202411799329.9) discloses a fault diagnosis method for multi-power-level nuclear power plants based on deep transfer learning. By introducing loss functions such as maximum mean difference, it aligns the feature distributions of the source and target domains, attempting to achieve cross-operating-condition transfer of diagnostic knowledge. However, such methods that rely on target domain data have significant limitations in engineering applications: on the one hand, before the model aligns to the new operating condition, it must undergo a target domain data accumulation and feature alignment process, which inevitably introduces an "unreliable delay period" in diagnostic functionality, during which the system lacks effective safety monitoring capabilities; on the other hand, and more critically, existing transfer learning methods typically implicitly assume that the target domain (new operating condition) has training samples (regardless of whether they are labeled) containing various fault modes. However, given the stringent operating procedures and low probability of failures in nuclear power plants, obtaining target domain data covering all failure categories is difficult; typically, only normal operation data can be obtained. This "missing target domain failure samples" makes it difficult for traditional transfer learning methods to construct complete failure feature maps, thus limiting their effectiveness in real-world industrial scenarios. Summary of the Invention

[0004] The purpose of this invention is to address the problems in existing technologies by proposing a method and system for fault diagnosis in nuclear power plants under dynamic variable power conditions based on adversarial feature gating. This invention provides a method to improve the diagnostic performance of fault diagnosis models at untrained power levels by explicitly decoupling feature selection from the model training process and utilizing a consensus adversarial feature gating mechanism to screen domain-invariant key physical features, thereby achieving high-precision and high-stability fault identification in nuclear power plants under dynamic variable power operating conditions.

[0005] This invention is achieved through the following technical solution: This invention proposes a method for diagnosing variable power faults in nuclear power plants based on adversarial feature gating, the method comprising: Step 101: Determine key monitoring parameters; Based on the nuclear power plant's process flow and safety analysis report, determine a set of key process parameters for fault diagnosis; Step 102: Collect multi-condition operation data; use a nuclear power plant digital simulation platform or a real operation history database to collect sensor time-series data covering multiple discrete typical training domains; Step 103: Construction of multi-source domain joint training set; for each variable in the collected sensor time series data. The MinMax normalization method is used for linear transformation to scale the data to a uniform preset range. The time-series data after MinMax transformation is processed to construct structured training and testing samples. A sliding time window technique is used to segment the continuous time-series signal into a series of two-dimensional data segments, and each data segment is labeled with its corresponding running state label to form labeled sample tuples. ,in For the input feature matrix, For fault category labels, Assign power level labels; merge training samples from all selected discrete typical training domains to construct a multi-source domain joint training set; Step 104: Preliminary screening of feature subsets based on adversarial feature gating; constructing an adversarial network model containing a feature gating module, a fault classification module, and a domain classification module based on a multi-source domain joint training set; the feature gating module generates binary feature gating vectors through the Sigmoid function and a pass-through estimator; inputting samples from the multi-source domain joint training set into the adversarial network model, and adding a gradient inversion layer (GRL) between the feature gating module and the domain classification module; using the Adam optimization algorithm, by minimizing the fault classification cross-entropy loss while maximizing the domain classification loss, forcing the gating vectors to retain only fault-sensitive and domain-invariant features; Step 105: Feature subset consensus voting; the adversarial feature gating training process is executed independently. This involves generating a set of binary feature gating vectors through multiple independent runs. Subsequently, the final consensus gating vector is derived using a majority voting strategy. ; Step 106: Model training based on decoupling features; consensus-based gating vector training. The final fault classification model is constructed using a multi-source domain joint training dataset. ; Step 107: Online real-time fault identification; the final fault classification model after training is used for... and consensus gating vector The system is jointly deployed in the online monitoring system of the nuclear power plant; it continuously collects sensor monitoring data of the nuclear power plant under the current operating conditions in real time. The real-time collected data is preprocessed using the MinMax feature normalization method and sliding time window technology to generate format-aligned time series samples to be tested, and then multiplied by a consensus gating vector. Real-time feature gating is performed, and the model is then input to output the real-time fault category and confidence level.

[0006] Furthermore, in step 102, the discrete typical training domain is a pre-selected, representative finite number of power levels, with an interval of 5%FP to 10%FP between each power level.

[0007] Furthermore, in step 103, the specific formula for data scaling is: (1) in, and These are the minimum and maximum values ​​of the corresponding parameters in the training set.

[0008] Further, in step 104, given the inclusion For each training batch of samples, the fault classification module minimizes the cross-entropy classification loss function as shown in equation (8). Optimize: (8) in, Indicates the total number of fault categories; It is a sample Category The true label; It is the predicted probability output by the fault classification module.

[0009] Furthermore, in step 104, during the forward propagation of the adversarial network model, GRL is used as an identity transformation; while during the back propagation, GRL incorporates the domain classification loss. The inverted gradient is then passed to the learnable vector in the feature gating module. Learnable vectors Overall loss As shown in equation (10): (10) in, For balance coefficient, This is the cross-entropy classification loss function.

[0010] Furthermore, in step 105, the consensus gating vector The calculation process is shown in equation (11): (11) in, Indicates an indicator function; Indicates the first Binary feature gating vectors obtained from each independent run; This is the preset consensus voting threshold.

[0011] Furthermore, in step 106, the consensus gating vector is utilized. For the original input sample Perform feature gating, which involves forcibly masking the data of non-critical feature channels to zero using element-wise Hadamard products to generate gated samples. Subsequently, the gated samples Directly input into the final fault classification model In the process of backpropagation training, only the cross-entropy classification loss function is used. Optimize network parameters.

[0012] This invention also proposes a nuclear power plant variable power fault diagnosis system based on adversarial feature gating, the system comprising: Data acquisition module: Used to collect historical or simulation data covering multiple discrete typical power levels from the digital simulation platform or historical database of the nuclear power plant to construct a training domain; at the same time, this data acquisition module directly or indirectly communicates with the underlying sensor network of the nuclear power plant to obtain real-time monitoring data under the current operating conditions of the nuclear power plant. Sample construction module: This module is connected to the data acquisition module and is used to receive the aforementioned historical / simulation data or real-time data. The sample construction module has a built-in feature preprocessing algorithm and uses a unified MinMax normalization method and sliding time window technology to segment and scale continuous time-series signals. For historical / simulation data, it outputs time window samples with fault labels and power level domain labels. For real-time data, it outputs unlabeled real-time time window samples. Gated Vector Construction Module: This module is communicatively connected to the sample construction module. It integrates a feature-gated learning network, a fault classifier, a domain classifier, and a gradient inversion layer. The module receives labeled time-window samples for adversarial training and incorporates a consensus voting mechanism to control the adversarial network. Each iteration independently executes a majority voting strategy, ultimately constructing and outputting a stable consensus gating vector. ; Feature gating module: It is communicatively connected to the sample construction module and the gating vector construction module respectively; the feature gating module receives labeled time window samples or real-time samples from the sample construction module, and receives consensus gating vectors fixed by the gating vector construction module; the feature gating module performs feature gating operation based on element-wise Hadamard product to forcibly mask the data of non-critical feature channels in the input sample to zero, and outputs gated time window samples or gated real-time samples; Model training module: It is connected to the feature gating module. The model training module receives the gating time window samples and optimizes the model parameters only based on the cross-entropy classification loss function in a stable physical feature space. Finally, it trains and outputs a diagnostic model with high cross-domain generalization ability. Online diagnostic module: It is communicatively connected to the feature gating module and the model training module respectively; the online diagnostic module loads the trained diagnostic model and receives real-time samples after gating generated by the feature gating module in real time; the online diagnostic module performs forward inference calculation on the real-time samples and outputs the diagnostic results and corresponding confidence levels of the current nuclear power plant operating conditions in real time. Human-machine interaction and alarm module: It is communicatively connected to the nuclear power plant monitoring interface and the online diagnostic module. This human-machine interaction and alarm module is used to synchronously present the real-time data trend and online diagnostic results of the nuclear power plant on the visual interface. When the diagnostic results indicate that a fault or abnormal state has a confidence level exceeding a set threshold, the human-machine interaction and alarm module triggers an audible and visual alarm and system linkage to provide decision support for the nuclear power plant operator.

[0013] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the nuclear power plant variable power fault diagnosis method based on adversarial feature gating.

[0014] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the nuclear power plant variable power fault diagnosis method based on adversarial feature gating.

[0015] The beneficial effects of this invention are: Zero-sample cross-domain generalization capability without target domain data: This invention, through a consensus-based adversarial feature gating mechanism, can proactively screen key physical features from high-dimensional monitoring data that possess both domain invariance and high fault sensitivity. Compared to traditional transfer learning or domain adaptation methods that rely on target domain data for feature alignment, this invention enables the model to directly and with high accuracy generalize to unseen target power conditions with only a finite number of discrete power levels (source domain) for training. This avoids the unreliable delay period in the initial deployment of the model under new conditions, reducing the data threshold and implementation cost for engineering deployment.

[0016] Possessing high mechanism interpretability and strong noise robustness: This invention abandons traditional implicit feature weight allocation (such as soft attention mechanisms) and uses discrete binary feature gating vectors to hard mask the input features. This not only effectively filters noisy features but also ensures that the feature subset on which the model depends has high physical consistency, greatly enhancing the physical rationality of the diagnostic logic and the credibility of the black-box model.

[0017] In summary, this invention proposes a highly generalizable and engineering-friendly intelligent fault diagnosis scheme for nuclear power plants, effectively overcoming the data distribution offset problem of traditional data-driven methods under variable power operation conditions, and providing solid technical support for realizing highly reliable intelligent monitoring of nuclear power plants. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart of the nuclear power plant variable power fault diagnosis method based on adversarial feature gating as described in this invention.

[0020] Figure 2 This is a diagram of the architecture of the nuclear power plant variable power fault diagnosis system based on adversarial feature gating as described in this invention.

[0021] Figure 3 A schematic diagram of the construction of a fault diagnosis model for a nuclear power plant with variable power output. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This invention aims to overcome the shortcomings of existing technologies and provide a method and system for fault diagnosis of nuclear power plants based on Consensus Adversarial Feature Gating (C-AFG). This method introduces a discrete adversarial feature gating mechanism, forcing the model to focus only on key features that are physically consistent under different operating conditions. Furthermore, it combines a multi-round independent consensus voting strategy to eliminate spurious correlation features caused by random disturbances or numerical noise. Finally, by explicitly decoupling feature selection and model training processes, a fault diagnosis model that maintains high accuracy even at unseen power levels can be constructed without target domain data.

[0024] Specifically, in combination Figures 1-3 This invention proposes a method for diagnosing variable power faults in nuclear power plants based on adversarial feature gating, the method comprising: Step 101: Determine key monitoring parameters; Based on the nuclear power plant's process flow and safety analysis report, a set of key process parameters for fault diagnosis is determined. These key process parameters include at least the pressure, temperature, flow rate, and water level of the reactor coolant system, as well as relevant variables from key subsystems such as the steam generator, main feedwater system, and pressurizer. They also include status information such as nuclear power and valve opening degree, and this information is stored as fixed parameters in a database.

[0025] Step 102: Collect multi-condition operating data; Using a nuclear power plant digital simulation platform or a real operational history database, sensor time-series data covering multiple discrete typical training domains are collected. The discrete typical training domains are a pre-selected, representative, and finite number of power levels. Preferably, the interval between each power level is 5%FP to 10%FP, for example, 70%FP, 80%FP, 90%FP, and 100%FP. For each selected typical power level, dynamic response process data for normal operating conditions and various preset typical fault types (such as loss-of-coolant accident LOCA, steam generator heat transfer tube rupture SGTR, main steam pipeline rupture MSLB, single-loop loss of flow LOF, etc.) are simulated or collected.

[0026] Unlike traditional methods that require continuous and dense sampling to cover all diagnostic power levels, the core of this invention lies in recognizing that similar dynamic processes exist for the same type of fault at different power levels. Data collection is only required at a few representative power levels, and the model can be effectively generalized to any intermediate power level by selecting key physical parameters that simultaneously possess domain-invariant characteristics and diagnostic value for model training.

[0027] Step 103: Construction of a joint training set for multiple source domains; For each variable in the collected sensor time series data The MinMax normalization method is used for linear transformation to scale the data to a uniform preset range. The specific formula is shown in (1): (1) in, and These represent the minimum and maximum values ​​of the corresponding parameters in the training set. The time-series data after the MinMax transformation is processed to construct structured training and testing samples. A sliding time window technique is used, setting a fixed window length (e.g., 10 seconds) and step size (e.g., 1 second) to divide the continuous time-series signal into a series of two-dimensional data segments, each segment having the dimension of (number of variables × number of time steps). Each data segment is then labeled with its corresponding running state label, forming labeled sample tuples. ,in For the input feature matrix, For fault category labels, The power level is labeled. Training samples from all selected discrete typical training domains (e.g., 70%FP, 80%FP, 90%FP, 100%FP) are merged to construct a multi-source domain joint training set. The constructed sample set is then divided into a training set and a test set in an 8:2 ratio.

[0028] Step 104: Preliminary screening of feature subsets based on adversarial feature gating; Based on the constructed multi-source domain joint training set, feature selection is carried out for model construction. The model architecture mainly consists of three parts, including a feature gating module. Fault classification module Domain classification module Among them, the feature gating module The core of this approach lies in introducing a discrete binary feature gating vector, designed to effectively filter out domain-specific features that vary with operating conditions, while retaining key physical features highly sensitive to fault diagnosis. This gating vector is achieved by introducing a dimension... Learnable vectors Perform parameterization, where This represents the total dimension of the sensor input parameters. For the first... The retention probability of each feature The Sigmoid activation function shown in equation (2) is used for modeling: (2) Further applying the threshold operation shown in equation (3) generates the final binary gate vector. : (3) in, Let represent the indicator function. To address the non-differentiability of the indicator function during backpropagation, a pass-through estimator is used, allowing the gradient to be updated through a hard thresholding layer. Subsequently, the nth sample constructed by scaling and time sliding windowing. The sample is processed into a gated sample by the feature gating operation shown in equation (4). : (4) in, Represents the element-wise Hadamard product; Represents a column vector consisting entirely of 1s; Item This indicates that a spatial selection mask is broadcast along the time dimension, thereby aligning the gating vector with the dimension of the input sample. Through this operation, features corresponding to the zero-value gate are masked at all time steps, effectively excluding them from subsequent computation. Gated samples As a fault classification module Domain classification module Shared input.

[0029] Specifically, the fault classification module Domain classification module Both are built upon one-dimensional convolutional neural networks, and their network architecture mainly consists of two parts: a convolutional feature extraction layer and a fully connected classification layer. The convolutional feature extraction layer is used to extract high-dimensional temporal features, while the fully connected classification layer maps the flattened features to the sample's label space. (Fault classification module) The output is the probability distribution of the current sample belonging to each fault category, while the domain classification module... The output is the probability distribution of the current sample belonging to each power level. For the th power level in the module... The forward propagation calculation process of the one-dimensional convolutional layer is shown in Equation (5): (5) in, Indicates the first The input feature map of the layer; This represents the output feature map after convolution and nonlinear activation. Indicates the first Layer-learnable one-dimensional convolutional kernel weights; Indicates the bias term; The symbol represents a one-dimensional convolution operation; is the activation function. The high-dimensional feature maps extracted by multiple convolutional layers are first flattened into one-dimensional vectors, and then input into a fully connected layer for dimensionality mapping in the feature space. The calculation process of the fully connected layer is shown in equation (6): (6) in, The input vector for the fully connected layer; and These represent the weight matrix and bias vector of this layer, respectively; finally, the network outputs the Logit value. The Softmax function shown in equation (7) is used to convert the probability distribution into a class probability distribution. .

[0030] (7) in, and Each represents a category Logit value and predicted probability; This represents the total number of categories in the classification task. Specifically, in the fault classification module... middle, Represents the number of preset fault types (e.g., LOCA, SGTR, etc.); in the domain classification module middle, This represents the number of discrete power levels contained in the source domain (e.g., 70% FP, 100% FP, etc.). Given a range of... For each training batch of samples, the fault classification module minimizes the cross-entropy classification loss function as shown in equation (8). Optimize: (8) in, Indicates the total number of fault categories; It is a sample Category The actual label (using One-hot encoding); This is the predicted probability output by the fault classification module. Meanwhile, the domain classification module... Aimed at being based on the same gating samples Identify the power level to which a sample belongs. Its optimization objective is to minimize the domain classification loss function as shown in equation (9). : (9) in, This indicates the number of power levels contained in the source domain. It is a sample Belongs to the A true label for a power level This is the predicted probability output by the domain classification module. It is used to guide the feature gating module. Retaining only key physical features that simultaneously possess domain invariance and fault separability, this step introduces a gradient reversal layer (GRL) between the feature gating module and the domain classification module. During the model's forward propagation, the GRL acts as an identity transformation; while during backpropagation, the GRL incorporates the domain classification loss. Invert the gradient (multiply by a negative coefficient) The learnable vector is then passed to the feature gating module. Learnable vectors Overall loss As shown in equation (10): (10) in, This is the balance coefficient. Preferably, Taken as ; Step 105: Consensus voting on feature subsets; To improve the robustness of feature selection, this invention introduces a consensus mechanism into the adversarial feature gating framework. Specifically, the adversarial feature gating training process is executed independently. Next, preferably Set to 8. Generate a set of binary feature gating vectors through multiple independent runs. Subsequently, the final consensus gating vector is derived using a majority voting strategy. The calculation process is shown in equation (11): (11) in, Indicates an indicator function; Indicates the first The binary gated vector obtained from each independent run; This is a preset consensus voting threshold. Preferably, Set to 0.5, this means that a feature will only be retained in the final consensus gating vector if it has been selected in a majority of independent runs. This consensus voting process effectively filters out sporadic noise interference, thereby stabilizing the feature selection results.

[0031] Step 106: Model training based on decoupling features; Consensus-based gating vectors In addition to constructing a multi-source domain joint training dataset, this step uses a one-dimensional convolutional neural network architecture to build the final fault classification model. At this stage, the feature gating module and the domain classification module used for feature selection in the early stage... All were completely removed. Specifically, before inputting the training samples into the network, consensus gating vectors were used. For the original input sample Perform the feature gating operation as shown in Equation (4), that is, force the data of non-critical feature channels to be masked to zero by element-wise Hadamard product to generate gated samples. Subsequently, the gated samples Directly input into the final fault classification model In the backpropagation training process, only the cross-entropy classification loss function defined in equation (8) is used. Optimize the network parameters. Through this decoupled retraining method, the model can focus on learning fault modes within a stable physical feature space, thereby achieving better cross-domain generalization ability.

[0032] Step 107: Online real-time fault identification; The final fault classification model after training and the determined consensus gating vector This system is jointly deployed in the online monitoring system of nuclear power plants. It continuously collects sensor monitoring data from the nuclear power plant under its current operating conditions in real time. The real-time collected data is rigorously preprocessed using the MinMax feature normalization method and sliding time window technique to generate format-aligned time-series samples for testing. During diagnostic inference, a fixed consensus gating vector is first used... The feature gating operation shown in Equation (4) is performed on the test sample to filter out non-critical features in real time by element-wise Hadamard product, generating a gated real-time sample. Subsequently, the gated real-time sample is directly input into the trained fault classification model. In this process, the model calculates and outputs the fault category prediction results for the current operating state.

[0033] This invention also proposes a nuclear power plant variable power fault diagnosis system based on adversarial feature gating, the system comprising: Data acquisition module: Used to collect historical or simulation data covering multiple discrete typical power levels from the digital simulation platform or historical database of the nuclear power plant to construct a training domain; at the same time, this data acquisition module directly or indirectly communicates with the underlying sensor network of the nuclear power plant to obtain real-time monitoring data under the current operating conditions of the nuclear power plant. Sample construction module: This module is connected to the data acquisition module and is used to receive the aforementioned historical / simulation data or real-time data. The sample construction module has a built-in feature preprocessing algorithm and uses a unified MinMax normalization method and sliding time window technology to segment and scale continuous time-series signals. For historical / simulation data, it outputs time window samples with fault labels and power level domain labels. For real-time data, it outputs unlabeled real-time time window samples. Gated Vector Construction Module: This module is communicatively connected to the sample construction module. It integrates a feature-gated learning network, a fault classifier, a domain classifier, and a gradient inversion layer. The module receives labeled time-window samples for adversarial training and incorporates a consensus voting mechanism to control the adversarial network. Each iteration independently executes a majority voting strategy, ultimately constructing and outputting a stable consensus gating vector. ; Feature gating module: It is communicatively connected to the sample construction module and the gating vector construction module respectively; the feature gating module receives labeled time window samples or real-time samples from the sample construction module, and receives consensus gating vectors fixed by the gating vector construction module; the feature gating module performs feature gating operation based on element-wise Hadamard product to forcibly mask the data of non-critical feature channels in the input sample to zero, and outputs gated time window samples or gated real-time samples; Model training module: It is connected to the feature gating module. The model training module receives the gating time window samples and optimizes the model parameters only based on the cross-entropy classification loss function in a stable physical feature space. Finally, it trains and outputs a diagnostic model with high cross-domain generalization ability. Online diagnostic module: It is communicatively connected to the feature gating module and the model training module respectively; the online diagnostic module loads the trained diagnostic model and receives real-time samples after gating generated by the feature gating module in real time; the online diagnostic module performs forward inference calculation on the real-time samples and outputs the diagnostic results and corresponding confidence levels of the current nuclear power plant operating conditions in real time. Human-machine interaction and alarm module: It is communicatively connected to the nuclear power plant monitoring interface and the online diagnostic module. This human-machine interaction and alarm module is used to synchronously present the real-time data trend and online diagnostic results of the nuclear power plant on the visual interface. When the diagnostic results indicate that a fault or abnormal state has a confidence level exceeding a set threshold, the human-machine interaction and alarm module triggers an audible and visual alarm and system linkage to provide decision support for the nuclear power plant operator.

[0034] Example The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0035] Currently, traditional data-driven fault diagnosis methods are easily affected by domain feature shifts under the variable power operation conditions of nuclear power plants, resulting in a significant decrease in generalization performance under unseen operating conditions. Existing transfer learning-based techniques heavily rely on target domain data, making rapid deployment in engineering projects difficult. To address these issues, this invention provides a nuclear power plant variable power fault diagnosis method based on adversarial feature gating, such as... Figure 1 As shown, the method includes: Step 101: Determine key monitoring parameters For this embodiment of the invention, firstly, based on the process flow diagram and safety analysis report of a megawatt-class pressurized water reactor (PWR) nuclear power plant, 135 key process parameters for fault diagnosis are identified. These parameters cover the pressure, temperature, flow rate, and water level of the reactor coolant system, as well as the state variables of key subsystems such as the steam generator, main feedwater system, pressurizer, and safety injection system. This information is stored as fixed parameters in a database.

[0036] Step 102: Collect multi-condition operation data In this embodiment of the invention, a nuclear power plant digital simulation platform jointly built with RELAP5 and Simulink was used to collect sensor time-series data covering multiple discrete typical training domains. Typical power levels of 70%FP, 80%FP, 90%FP, and 100%FP were selected as source domains. At each selected power level, one normal operating condition and six preset typical fault types were simulated, including: Loss of Cooling Pipes (LOCAC), Loss of Hot Pipes (LOCAH), Loss of Flow in a Single Loop (LOF), Steam Generator Heat Transfer Tube Rupture (SGTR), Main Feedwater Pipe Rupture (FLB), and Main Steam Pipe Rupture (MSLB). The simulation duration for each scenario was set to 1000 seconds, with a sampling frequency of 1 Hz, to fully record its dynamic response process data.

[0037] Step 103: Construction of the multi-source domain joint training set In this embodiment of the invention, for the raw time-series data collected in step 102, the MinMax normalization method is used to linearly scale each variable to the [0,1] interval (using only the statistics of the four source domain data as the scaling benchmark). A sliding time window technique is used, with a window length of 10 seconds and a step size of 1 second, to segment the continuous signal into a series of two-dimensional data segments with a dimension of (135×10). Each data segment is labeled with its corresponding fault category label (7 categories in total) and power level label (4 categories in total). Samples from 70%FP, 80%FP, 90%FP, and 100%FP are merged into a multi-source domain joint training set, and then divided into a training set and a validation set in an 8:2 ratio.

[0038] Step 104: Preliminary screening of feature subsets based on adversarial feature gating In this embodiment of the invention, an adversarial network (based on a one-dimensional convolutional neural network architecture) is constructed, comprising a feature gating module, a fault classification module (based on a one-dimensional convolutional neural network architecture), and a domain classification module. The feature gating module contains 135 learnable parameters and generates binary feature gating vectors through a sigmoid function and a pass-through estimator. Joint training set samples are input into this network, and a gradient inversion layer is added between the feature gating module and the domain classification module. The Adam optimization algorithm (initial learning rate set to 0.001, batch size set to 1024) is used to minimize the fault classification cross-entropy loss while maximizing the domain classification loss, forcing the gating vectors to retain only fault-sensitive and domain-invariant features.

[0039] Step 105: Consensus Voting on Feature Subsets In this embodiment of the invention, the adversarial training process is executed independently 8 times under different random seeds. After obtaining 8 sets of binary gating vectors, a consensus voting threshold of 0.5 is set, that is, only features that are selected in more than half of the independent runs are retained, thus generating the final consensus gating vector. .

[0040] Step 106: Model Training Based on Decoupling Features In this embodiment of the invention, the feature gating module and the domain classification module in the above steps are removed. The consensus gating vector fixed in step 105 is then used. Perform an element-wise Hadamard product masking operation on the multi-source domain joint training set from step 103, setting the unselected feature channels to zero. Then, directly input the processed samples into a newly initialized 1D-CNN fault classification model, using only the fault classification loss for backpropagation training until the model converges, and save the model.

[0041] Step 107: Online Real-time Fault Identification In this embodiment of the invention, the trained diagnostic model and consensus gating vector are... Deployed in the monitoring system of the nuclear power plant. It continuously collects sensor data reflecting the current operation of the power plant. For the real-time collected data, it uses the same scaling and sliding windowing method to generate test samples, which are then multiplied by a consensus gating vector. Real-time feature gating is performed, and finally, the diagnostic model is input to output the real-time fault category and confidence level.

[0042] To clarify the technical solutions provided in this invention, a specific cross-domain diagnostic task is used as an example. Assume a nuclear power plant is currently operating at 86% FP (an intermediate power level not seen in the training set) and a single-loop loss of current (LOF) accident has occurred. The online monitoring system collects dynamic data from 135 sensors in real time. Due to the decrease in the main pump speed in loop A, the coolant flow rate in loop A will experience a significant physical decrease due to the LOF fault. Simultaneously, some parameters fluctuate slightly due to sensor noise floor or line noise. During inference, the system of this invention first applies a pre-defined consensus gating vector. Gating is applied to real-time samples. This gating vector successfully masks noise-affected parameters to zero while accurately preserving parameter channels with domain-invariant physical meaning. The gated samples are then input into a trained 1D-CNN classifier. By eliminating interference from operating conditions and noise, the model accurately identifies the dynamic pattern of the current sample as highly matching "single-loop current loss (LOF)," and immediately outputs a high-confidence diagnostic result, thus providing timely and reliable decision support for the operator.

[0043] The specific implementation methods provided in this invention achieve high-precision cross-domain diagnosis of nuclear power plant faults under unseen power levels by constructing a technical framework that combines adversarial gating mechanisms with consensus strategies. This method is not only simple and efficient in engineering deployment, but its mechanism of action is also interpretable, possessing high practical industrial application value.

[0044] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the nuclear power plant variable power fault diagnosis method based on adversarial feature gating.

[0045] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the nuclear power plant variable power fault diagnosis method based on adversarial feature gating.

[0046] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0047] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0048] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0049] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0050] The above provides a detailed description of the nuclear power plant variable power fault diagnosis method and system based on adversarial feature gating proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for fault diagnosis of variable power in nuclear power plants based on adversarial feature gating, characterized in that, The method includes: Step 101: Determine key monitoring parameters; Based on the nuclear power plant's process flow and safety analysis report, determine a set of key process parameters for fault diagnosis; Step 102: Collect multi-condition operation data; use a nuclear power plant digital simulation platform or a real operation history database to collect sensor time-series data covering multiple discrete typical training domains; Step 103: Construction of multi-source domain joint training set; for each variable in the collected sensor time series data. The MinMax normalization method is used for linear transformation to scale the data to a uniform preset range. The time-series data after MinMax transformation is processed to construct structured training and testing samples. A sliding time window technique is used to segment the continuous time-series signal into a series of two-dimensional data segments, and each data segment is labeled with its corresponding running state label to form labeled sample tuples. ,in For the input feature matrix, For fault category labels, Assign power level labels; merge training samples from all selected discrete typical training domains to construct a multi-source domain joint training set; Step 104: Preliminary screening of feature subsets based on adversarial feature gating; constructing an adversarial network model containing a feature gating module, a fault classification module, and a domain classification module based on a multi-source domain joint training set; the feature gating module generates binary feature gating vectors through the Sigmoid function and a pass-through estimator; inputting samples from the multi-source domain joint training set into the adversarial network model, and adding a gradient inversion layer (GRL) between the feature gating module and the domain classification module; using the Adam optimization algorithm, by minimizing the fault classification cross-entropy loss while maximizing the domain classification loss, forcing the gating vectors to retain only fault-sensitive and domain-invariant features; Step 105: Feature subset consensus voting; the adversarial feature gating training process is executed independently. This involves generating a set of binary feature gating vectors through multiple independent runs. Subsequently, the final consensus gating vector is derived using a majority voting strategy. ; Step 106: Model training based on decoupling features; consensus-based gating vector training. The final fault classification model is constructed using a multi-source domain joint training dataset. ; Step 107: Online real-time fault identification; the final fault classification model after training is used for... and consensus gating vector The system is jointly deployed in the online monitoring system of the nuclear power plant; it continuously collects sensor monitoring data of the nuclear power plant under the current operating conditions in real time. The real-time collected data is preprocessed using the MinMax feature normalization method and sliding time window technology to generate format-aligned time series samples to be tested, and then multiplied by a consensus gating vector. Real-time feature gating is performed, and the model is then input to output the real-time fault category and confidence level.

2. The method according to claim 1, characterized in that, In step 102, the discrete typical training domain is a pre-selected, representative finite number of power levels, with an interval of 5%FP to 10%FP between each power level.

3. The method according to claim 1, characterized in that, In step 103, the specific formula for data scaling is: (1) in, and These are the minimum and maximum values ​​of the corresponding parameters in the training set.

4. The method according to claim 1, characterized in that, In step 104, given the contents For each training batch of samples, the fault classification module minimizes the cross-entropy classification loss function as shown in equation (8). Optimize: (8) in, Indicates the total number of fault categories; It is a sample Category The true label; It is the predicted probability output by the fault classification module.

5. The method according to claim 1, characterized in that, In step 104, during the forward propagation of the adversarial network model, GRL is used as an identity transformation; while during the back propagation, GRL incorporates the domain classification loss. The inverted gradient is then passed to the learnable vector in the feature gating module. Learnable vectors Overall loss As shown in equation (10): (10) in, For balance coefficient, This is the cross-entropy classification loss function.

6. The method according to claim 1, characterized in that, In step 105, the consensus gating vector The calculation process is shown in equation (11): (11) in, Indicates an indicator function; Indicates the first Binary feature gating vectors obtained from each independent run; This is the preset consensus voting threshold.

7. The method according to claim 1, characterized in that, In step 106, consensus gating vectors are used. For the original input sample Perform feature gating, which involves forcibly masking the data of non-critical feature channels to zero using element-wise Hadamard products to generate gated samples. Subsequently, the gated samples Directly input into the final fault classification model In the process of backpropagation training, only the cross-entropy classification loss function is used. Optimize network parameters.

8. A nuclear power plant variable power fault diagnosis system based on adversarial feature gating, characterized in that, The system includes: Data acquisition module: Used to collect historical or simulation data covering multiple discrete typical power levels from the digital simulation platform or historical database of the nuclear power plant to construct a training domain; at the same time, this data acquisition module directly or indirectly communicates with the underlying sensor network of the nuclear power plant to obtain real-time monitoring data under the current operating conditions of the nuclear power plant. Sample construction module: This module is connected to the data acquisition module and is used to receive the aforementioned historical / simulation data or real-time data. The sample construction module has a built-in feature preprocessing algorithm and uses a unified MinMax normalization method and sliding time window technology to segment and scale continuous time-series signals. For historical / simulation data, it outputs time window samples with fault labels and power level domain labels. For real-time data, it outputs unlabeled real-time time window samples. Gated Vector Construction Module: This module is communicatively connected to the sample construction module. It integrates a feature-gated learning network, a fault classifier, a domain classifier, and a gradient inversion layer. The module receives labeled time-window samples for adversarial training and incorporates a consensus voting mechanism to control the adversarial network. Each iteration independently executes a majority voting strategy, ultimately constructing and outputting a stable consensus gating vector. ; Feature gating module: It is communicatively connected to the sample construction module and the gating vector construction module respectively; the feature gating module receives labeled time window samples or real-time samples from the sample construction module, and receives consensus gating vectors fixed by the gating vector construction module; the feature gating module performs feature gating operation based on element-wise Hadamard product to forcibly mask the data of non-critical feature channels in the input sample to zero, and outputs gated time window samples or gated real-time samples; Model training module: It is connected to the feature gating module. The model training module receives the gating time window samples and optimizes the model parameters only based on the cross-entropy classification loss function in a stable physical feature space. Finally, it trains and outputs a diagnostic model with high cross-domain generalization ability. Online diagnostic module: It is communicatively connected to the feature gating module and the model training module respectively; the online diagnostic module loads the trained diagnostic model and receives real-time samples after being gated by the feature gating module; the online diagnostic module performs forward inference calculation on the real-time samples and outputs the diagnostic results and corresponding confidence levels of the current nuclear power plant operating conditions in real time. Human-machine interaction and alarm module: It is communicatively connected to the nuclear power plant monitoring interface and the online diagnostic module. This human-machine interaction and alarm module is used to synchronously present the real-time data trend and online diagnostic results of the nuclear power plant on the visual interface. When the diagnostic results indicate that a fault or abnormal state has a confidence level exceeding a set threshold, the human-machine interaction and alarm module triggers an audible and visual alarm and system linkage to provide decision support for the nuclear power plant operator.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.

Images