Pressurized water reactor fuel assembly damage position prediction method based on machine learning

Abstract

The invention discloses a pressurized water reactor fuel assembly damage position prediction method based on machine learning, and the method comprises the steps: building and training a machine learning model through a decision tree algorithm based on time evolution data of damaged primary loop nuclide concentration obtained by reactor core physical analysis software and evolution simulation calculation software; according to the method, a preprocessed multi-dimensional nuclide concentration stable value feature vector is taken as input, a prediction result of a damage position is taken as output, model overfitting is effectively controlled through pre-pruning and post-pruning strategies, generalization ability is ensured, the damage probability of each position can be provided, and prediction of the damage position of the pressurized water reactor fuel assembly is realized.

Description

Technical Field

[0001] This invention relates to the field of pressurized water reactor core physics calculation technology, specifically to a method for predicting the location of pressurized water reactor fuel assembly failure based on machine learning. Background Technology

[0002] The integrity of the fuel cladding is a critical requirement for the nuclear safety of pressurized water reactors. Damage to the fuel cladding can trigger radioactive releases, forcing reactor shutdown and causing significant losses. Currently, damage location primarily relies on offline detection after shutdown, a time-consuming process that severely impacts operations. Therefore, developing online damage location technology is essential.

[0003] The core of traditional online diagnostic methods is to utilize 134 Cs / 137 The Cs activity ratio is used to locate the burnup region of damaged fuel. However, this method is based on the assumption that "a single damage point is located in a uniform burnup region". When there are multiple damage points with different burnup in the core, the nuclide signals in the coolant are in a mixed state. The single mapping relationship of "ratio-burnup" relied upon by traditional methods fails, making it impossible to distinguish multiple sources. The positioning accuracy can only be limited to the "region level" and cannot achieve "component-level" positioning.

[0004] In recent years, intelligent methods such as machine learning have demonstrated speed and accuracy advantages in damage parameter inversion, and multi-nucleoside "fingerprint" features have provided new evidence for refined diagnosis. However, current intelligent diagnostic research mainly focuses on damage degree determination, and its output is not spatial coordinates, failing to solve the core problem of "spatial localization of damage points." Furthermore, existing methods often have a one-sided approach to feature utilization, failing to fully integrate multi-dimensional nuclide spectral information, thus limiting the practical application capabilities of the models.

[0005] In summary, existing technologies are insufficient for complex operating conditions and are limited by their underlying principles. Therefore, there is an urgent need to develop a new technology that can comprehensively utilize multidimensional nuclide big data, intelligently analyze mixed signals, and achieve component-level positioning. Summary of the Invention

[0006] To overcome the problems existing in the prior art, the present invention aims to provide a machine learning-based method for predicting the location of pressurized water reactor fuel assemblies failure: based on the time evolution data of primary loop nuclide concentration after failure obtained by core physics analysis software and evolution simulation software, the method takes a preprocessed multidimensional nuclide concentration stability value feature vector as input and the predicted failure location as output. Through pre-pruning and post-pruning strategies, the method effectively controls model overfitting, ensures generalization ability, and provides the failure probability at each location, thereby achieving the prediction of the failure location of pressurized water reactor fuel assemblies.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A machine learning-based method for predicting the location of fuel assembly failure in a pressurized water reactor includes the following steps: Step 1: Building the Machine Learning Model Dataset Core physics analysis software was used to calculate the nuclide concentration and burn-up depth of each fuel assembly at burn-up point. Combined with pressurized water reactor operating parameters, including primary coolant mass and nuclide concentration or activity, purge or discharge mass flow rate, boron concentration, boron decay rate, and nuclide data, a model was created to calculate the evolution of primary coolant concentration over time after each hypothetical fuel assembly fails at a predetermined height. The nuclide data included burn-up chain, decay constant, decontamination factor, and escape rate coefficient. The specific model is as follows: An empirical parameter, the escape rate coefficient, is introduced to describe the overall release efficiency of fission products from fuel to coolant. The primary loop nuclide generation pathways are considered: nuclide release from the reactor core and nucleus decay. Removal pathways include: nuclide decay removal, removal due to boron dilution, and purification by the desalination unit. The following equilibrium equation is established: The core equilibrium equation in a primary loop: The daughter nucleus equilibrium equation in a primary loop: in, N i Nuclide in the primary coolant i Concentration, number of atoms / g; F This represents the percentage of fuel rod cladding damage. R i nuclide i Escape rate coefficient, s -1 ; N Fi Nuclide in fuel i The amount of accumulated atoms; M The mass of the primary coolant is expressed in grams. Q For purification or discharge of mass flow rate, g / s; l i nuclide i decay constant, s -1 ; l j nuclide j decay constant, s -1 ; f ij nuclide j decay into nuclides i Branch ratio; DF i nuclide i Detergent agents; B 0 represents the initial boron concentration, in ppm; βThe boron concentration reduction rate is expressed in ppm / s. D i nuclide i The dilution factor, s -1 The formula is as follows: The differential equations are numerically discretized and solved to obtain the evolution data of primary loop nuclide concentration over time after each hypothetical fuel assembly is damaged at a preset height, and a machine learning model dataset is constructed.

[0008] Step 2: Dataset Preprocessing The dataset simulates the entire process of the evolution of the concentration of multiple nuclides in the primary coolant over time when any fuel assembly fails at a preset height during reactor operation. The dataset includes fuel assemblies and nodes along the axial direction of each assembly, which are determined according to the needs of core calculation and analysis. Each height node of each assembly corresponds to an independent sub-dataset. Each sub-dataset consists of the monitored key nuclides and time series data from the moment of failure, with a given step size and number of steps. A feature aggregation preprocessing strategy is adopted. According to the physical law that, under constant operating conditions, after the release and purification of nuclides from damaged fuel in the primary loop reach a dynamic equilibrium, the concentration of each nuclide will tend to a saturated stable value. The stable value or average value of each nuclide concentration sequence in the later stage of the simulation is extracted. This set of stable values ​​constitutes a "nuclide concentration fingerprint" feature vector that represents the "damage spatial location, i.e., component number and axial height", which is used as the input of the machine learning model. Step 3: Building the Machine Learning Model Framework Construct a classification model that can learn the complex mapping relationship between "nuclide concentration fingerprint" and "spatial location of damage"; 1) Core Algorithm Selection: Decision Tree Decision trees are chosen as the base classifier. A decision tree is a tree-structured model that classifies samples by judging the feature values ​​layer by layer. 2) Feature selection criteria: Information gain When constructing each node of the decision tree, it is necessary to select the feature, i.e., nuclide, that can best distinguish different damage locations; information gain is used as the feature selection criterion, with the goal of maximizing the purity of the dataset after partitioning, i.e., samples within the same branch should belong to the same damage location as much as possible. When constructing a decision tree, the algorithm calculates the information gain of all available features and selects the feature with the largest information gain as the splitting criterion for the current node, thereby gradually narrowing down the possible range of the damaged location in the most efficient way. 3) Overfitting control strategy To prevent the model from overfitting to noise in the training data and causing a decline in performance on new data, the following pruning strategy is adopted: Pre-pruning: During the construction of the decision tree, stop conditions are applied in advance, that is, the maximum depth and minimum number of samples are limited, to prevent the decision tree from growing too deep and too complex; Post-pruning: First, allow the decision tree to grow fully, which may lead to overfitting. Then, examine the non-leaf nodes from the bottom up. If replacing the non-leaf nodes with leaf nodes can improve the overall accuracy of the model on the validation set, then prune. This method usually results in a decision tree with stronger generalization ability. Step 4: Model Training, Validation, and Optimization The preprocessed "nuclide concentration fingerprint-damage spatial location" dataset was randomly divided into a training set and an independent validation set according to a set ratio. A decision tree model was fitted using the training set data. During training, key hyperparameters are systematically tuned using grid search or random search combined with cross-validation to optimize model performance. After the model is trained, its performance is evaluated using an independent validation set, with the core metric being the Top-K classification accuracy of the damaged spatial location.

[0009] Preferably, the decision tree structure includes the following: Root node: Contains all training data and represents the beginning of the problem; Internal nodes, or decision nodes, correspond to the test conditions for a certain nuclide concentration characteristic, and divide the data into different branches based on the test results. Leaf nodes, also known as terminal nodes, represent the final classification result, i.e., a specific damaged fuel assembly and its height location. The path from the root node to any leaf node forms a classification rule, which can predict the location of damage in new nuclide observation data.

[0010] Preferably, the key hyperparameters mentioned in step four include: max_depth: The maximum depth of the tree, which controls the model complexity; min_samples_split: The minimum number of samples required to further split an internal node; min_samples_leaf: The minimum number of samples required for a leaf node; criterion: Feature selection criteria.

[0011] Compared with the prior art, the beneficial effects of the present invention are: 1. Compared with the limitations of traditional leak detection technology which is restricted by the shutdown state, the method of this invention uses a trained machine learning model to predict the location of pressurized water reactor fuel assembly damage, which can realize online location of damaged fuel assembly.

[0012] 2. Compared to existing detection technologies that can only predict the batch of damaged components, the method of this invention uses core physics analysis software and evolution simulation software to obtain data on the evolution of primary loop nuclide concentration over time after damage and trains a machine learning model, which can predict the location of damaged components.

[0013] 3. The dataset construction method and decision tree model framework of the present invention can improve positioning accuracy or adapt to more complex damage scenarios by incorporating more types of nuclide data or power plant operating parameters as features, or by replacing / integrating more complex machine learning models (such as random forests and gradient boosting trees). Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0015] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. This description is intended to enable those skilled in the art to understand and implement the present invention, and is not intended to limit the scope of protection of the present invention.

[0016] like Figure 1 As shown, the core of the machine learning-based pressurized water reactor fuel assembly failure location prediction method provided by this invention lies in constructing, training, validating, and optimizing a failure location prediction model (i.e., a machine learning model) based on a preprocessed training set to predict the location of the failed fuel assembly. The specific steps are as follows: Step 1: Using core physics analysis software, such as the NECP-Bamboo program from the NECP team at Xi'an Jiaotong University, calculate the nuclide concentration and burn-down depth of each fuel assembly at burn-up point. Combined with pressurized water reactor operating parameters, including primary coolant mass and nuclide concentration or activity, purge or discharge mass flow rate, boron concentration, boron decay rate, and nuclide data, model and calculate the evolution of primary coolant concentration over time after each hypothetical fuel assembly fails at a predetermined height. The nuclide data includes burn-down chain, decay constant, decontamination factor, and escape rate coefficient. The specific model is as follows: An empirical parameter, the escape rate coefficient, is introduced to describe the overall release efficiency of fission products from fuel to coolant. The primary loop nuclide generation pathways are considered: nuclide release from the reactor core and nucleus decay. Removal pathways include: nuclide decay removal, removal due to boron dilution, and purification by the desalination unit. The following equilibrium equation is established: The core equilibrium equation in a primary loop: The daughter nucleus equilibrium equation in a primary loop: in, N iNuclide in the primary coolant i Concentration, number of atoms / g; F This represents the percentage of fuel rod cladding damage. R i nuclide i Escape rate coefficient, s -1 ; N Fi Nuclide in fuel i The amount of accumulated atoms; M The mass of the primary coolant is expressed in grams. Q For purification or discharge of mass flow rate, g / s; l i nuclide i decay constant, s -1 ; l j nuclide j decay constant, s -1 ; f ij nuclide j decay into nuclides i Branch ratio; DF i nuclide i Detergent agents; B 0 represents the initial boron concentration, in ppm; β The boron concentration reduction rate is expressed in ppm / s. D i nuclide i The dilution factor, s -1 The formula is as follows: The differential equations are numerically discretized and solved to obtain the evolution data of primary loop nuclide concentration over time after each hypothetical fuel assembly is damaged at a preset height, and a machine learning model dataset is constructed.

[0017] Step 2: Dataset preprocessing.

[0018] The dataset simulates the entire process of the evolution of multiple nuclides concentrations in the primary coolant over time when any fuel assembly fails at a preset height during reactor operation. The total number of fuel assemblies included in the dataset and the nodes along the axial direction of each assembly can be determined according to the needs of core computational analysis. Each height node of each assembly corresponds to an independent subset of the dataset. Each subset contains the key nuclides being monitored (such as...). 134 Cs、 137 Cs、 131 I, 133 Xe et al.) and time series data from the moment of failure, with a given step size and number of steps.

[0019] A feature aggregation preprocessing strategy is adopted. Based on the physical law that, under constant operating conditions, after the release and purification of nuclides from damaged fuel in the primary loop reach a dynamic equilibrium, the concentration of each nuclide will tend to a saturated stable value, the stable value or average value of each nuclide concentration sequence is extracted in the later stages of the simulation (such as the last few time steps). This set of stable values ​​constitutes a "nuclide concentration fingerprint" feature vector representing the "spatial location of damage (component number and axial height)," which serves as the input to the machine learning model.

[0020] Step 3: Building the Machine Learning Model Framework The core of the method of this invention is to construct a classification model that can learn the complex mapping relationship between "nuclide concentration fingerprint" and "damage spatial location (component number, axial height)".

[0021] 1) Core Algorithm Selection: Decision Tree A decision tree is chosen as the base classifier. A decision tree is a tree-structured model that classifies samples by making decisions layer by layer based on feature values. Its structure includes: Root node: Contains all training data and represents the beginning of the problem.

[0022] Internal nodes (decision nodes): correspond to the concentration characteristics of a certain nuclide (e.g., 134 The test conditions for Cs concentration (such as "whether it is greater than a certain threshold") are used to divide the data into different branches based on the test results.

[0023] Leaf node (terminal node): Represents the final classification result, i.e., a specific damaged fuel assembly and its height location.

[0024] The path from the root node to any leaf node forms a classification rule, which can predict the location of damage in new nuclide observation data.

[0025] 2) Feature selection criterion: Information Gain When constructing each node of the decision tree, it is necessary to select the feature (nucleus) that best distinguishes different damage locations. The method of this invention uses information gain as the feature selection criterion, and its goal is to maximize the purity of the dataset after partitioning, that is, to ensure that samples within the same branch belong to the same damage location as much as possible.

[0026] When constructing a decision tree, the algorithm calculates the information gain of all available features and selects the feature with the largest information gain as the splitting criterion for the current node, thereby gradually narrowing down the possible range of the damaged location in the most efficient way.

[0027] 3) Overfitting control strategy To prevent the model from overfitting to noise in the training data and causing a decline in performance on new data, the following pruning strategy is adopted: Pre-pruning: During the construction of the decision tree, apply the above stopping conditions in advance (such as limiting the maximum depth and the minimum number of samples) to prevent the decision tree from growing too deep or too complex.

[0028] Post-pruning: First, allow the decision tree to grow fully (potentially leading to overfitting), then examine the non-leaf nodes from the bottom up. If replacing a non-leaf node with a leaf node (using its most common category as the prediction result) improves the model's overall accuracy on the validation set, then pruning is performed. This method typically yields decision trees with stronger generalization capabilities.

[0029] Step 4: Model Training, Validation, and Optimization The preprocessed "nuclide concentration fingerprint-damage spatial location" dataset was randomly divided into a training set and an independent validation set in an approximately 8:2 ratio. A decision tree model was fitted using the training set data.

[0030] During training, key hyperparameters are systematically tuned to optimize model performance using either grid search or random search combined with cross-validation. These key hyperparameters include: max_depth: The maximum depth of the tree, which controls the complexity of the model.

[0031] min_samples_split: The minimum number of samples required to further split an internal node.

[0032] min_samples_leaf: The minimum number of samples required for a leaf node.

[0033] criterion: Feature selection criteria (such as "gini" or "entropy").

[0034] After the model is trained, its performance is evaluated using an independent validation set. The core metric is the "Top-K classification accuracy of the damage location (component + height)". For example, Top-1 accuracy represents the proportion of the location predicted by the model with the highest probability that is the actual damage location; Top-3 accuracy represents the proportion of the actual damage location that appears in the top three candidate locations predicted by the model.

Claims

1. A method for predicting the location of pressurized water reactor fuel assembly failure based on machine learning, characterized in that, Includes the following steps: Step 1: Building the Machine Learning Model Dataset Core physics analysis software was used to calculate the nuclide concentration and burn-up depth of each fuel assembly at burn-up point. Combined with the operating parameters of the pressurized water reactor, including the primary coolant mass and nuclide concentration or activity, purge or discharge mass flow rate, boron concentration, boron decay rate and nuclide data, modeling and calculating the evolution of primary coolant concentration over time after each hypothetical fuel assembly breaks at a preset height, and constructing a machine learning model dataset; the nuclide data includes burn-up chain, decay constant, decontamination factor, escape rate coefficient; Step 2: Dataset Preprocessing The dataset simulates the entire process of the evolution of the concentration of multiple nuclides in the primary coolant over time when any fuel assembly is damaged at a preset height during reactor operation. The dataset contains fuel assemblies and nodes along the axial direction of each assembly, which are determined according to the needs of core computational analysis. Each height node of each assembly box corresponds to an independent sub-dataset. Each sub-dataset consists of the monitored key nuclides and time series data from the moment of failure, with a given step size and number of steps. A feature aggregation preprocessing strategy is adopted. According to the physical law that, under constant operating conditions, after the release and purification of nuclides from damaged fuel in the primary loop reach a dynamic equilibrium, the concentration of each nuclide will tend to a saturated stable value. The stable value or average value of each nuclide concentration sequence in the later stage of the simulation is extracted. This set of stable values ​​constitutes a "nuclide concentration fingerprint" feature vector that represents the "damage spatial location, i.e., component number and axial height", which is used as the input of the machine learning model. Step 3: Building the Machine Learning Model Framework Construct a classification model that can learn the complex mapping relationship between "nuclide concentration fingerprint" and "spatial location of damage"; 1) Core Algorithm Selection: Decision Tree Decision trees are chosen as the base classifier. A decision tree is a tree-structured model that classifies samples by judging the feature values ​​layer by layer. 2) Feature selection criteria: Information gain When constructing each node of the decision tree, it is necessary to select the feature, i.e., nuclide, that can best distinguish different damage locations; information gain is used as the feature selection criterion, with the goal of maximizing the purity of the dataset after partitioning, i.e., samples within the same branch should belong to the same damage location as much as possible. When constructing a decision tree, the algorithm calculates the information gain of all available features and selects the feature with the largest information gain as the splitting criterion for the current node, thereby gradually narrowing down the possible range of the damaged location in the most efficient way. 3) Overfitting control strategy To prevent the model from overfitting to noise in the training data and causing a decline in performance on new data, the following pruning strategy is adopted: Pre-pruning: During the construction of the decision tree, stop conditions are applied in advance, that is, the maximum depth and minimum number of samples are limited, to prevent the decision tree from growing too deep and too complex; Post-pruning: First, allow the decision tree to grow fully, which may lead to overfitting. Then, examine the non-leaf nodes from the bottom up. If replacing the non-leaf nodes with leaf nodes can improve the overall accuracy of the model on the validation set, then prune. This method usually results in a decision tree with stronger generalization ability. Step 4: Model Training, Validation, and Optimization The preprocessed "nuclide concentration fingerprint-damage spatial location" dataset was randomly divided into a training set and an independent validation set according to a set ratio. A decision tree model was fitted using the training set data. During training, key hyperparameters are systematically tuned using grid search or random search combined with cross-validation to optimize model performance. After the model is trained, its performance is evaluated using an independent validation set, with the core metric being the Top-K classification accuracy of the damaged spatial location.

2. The method for predicting the location of pressurized water reactor fuel assembly failure based on machine learning according to claim 1, characterized in that, The method for modeling and calculating the evolution of primary loop nuclide concentrations over time after each hypothetical fuel assembly fails at a preset height is as follows: An empirical parameter, the escape rate coefficient, is introduced to describe the overall release efficiency of fission products from fuel to coolant. The primary loop nuclide generation pathways are considered: nuclide release from the reactor core and nucleus decay. Removal pathways include: nuclide decay removal, removal due to boron dilution, and purification by the desalination unit. The following equilibrium equation is established: The core equilibrium equation in a primary loop: The daughter nucleus equilibrium equation in a primary loop: in, N i Nuclide in the primary coolant i Concentration, number of atoms / g; F This represents the percentage of fuel rod cladding damage. R i nuclide i Escape rate coefficient, s -1 ; N Fi Nuclide in fuel i The amount of accumulated atoms; M The mass of the primary coolant is expressed in grams. Q For purification or discharge of mass flow rate, g / s; λ i nuclide i decay constant, s -1 ; λ j nuclide j decay constant, s -1 ; f ij nuclide j decay into nuclides i Branch ratio; DF i nuclide i Decontamination-removing agents; B 0 represents the initial boron concentration, in ppm; β The boron concentration reduction rate is expressed in ppm / s. D i nuclide i The dilution factor, s -1 The formula is as follows: The differential equations are numerically discretized and solved to obtain the evolution data of primary loop nuclide concentration over time after each hypothetical fuel assembly is damaged at a preset height, and a machine learning model dataset is constructed.

3. The method for predicting the location of pressurized water reactor fuel assembly failure based on machine learning according to claim 1, characterized in that, The decision tree structure includes the following: Root node: Contains all training data and represents the beginning of the problem; Internal nodes, or decision nodes, correspond to the test conditions for a certain nuclide concentration characteristic, and divide the data into different branches based on the test results. Leaf nodes, also known as terminal nodes, represent the final classification result, i.e., a specific damaged fuel assembly and its height location. The path from the root node to any leaf node forms a classification rule, which can predict the location of damage in new nuclide observation data.

4. The method for predicting the location of pressurized water reactor fuel assembly failure based on machine learning according to claim 1, characterized in that, The key hyperparameters mentioned in step four include: max_depth: The maximum depth of the tree, which controls the model complexity; min_samples_split: The minimum number of samples required to further split an internal node; min_samples_leaf: The minimum number of samples required for a leaf node; criterion: Feature selection criteria.

Images