Lead-bismuth cooled microreactor 210 A Fast Inversion Method and System for Multi-Task Source Items

Abstract

The application relates to the field of nuclear reactor radiation safety, and particularly discloses a lead-bismuth cooled micro reactor 210 A Po multi-source term rapid inversion method and system are provided.The method comprises the following steps: obtaining target leakage parameters of a target reactor under a target leakage working condition; establishing a corresponding 210 Po transport diffusion model for the target leakage working condition; inputting the target leakage parameters into the 210 Po transport diffusion model; collecting a 210 Po three-dimensional dynamic concentration distribution; extracting 15-dimensional enhanced space-time features; constructing a data set and dividing the data set into a training set, a verification set and a test set; constructing a physical information neural network model; training and optimizing the physical information neural network model by using the training set and the verification set; inputting the test set into the trained and optimized physical information neural network model; and outputting a 210 Po leakage inversion result.The method can realize rapid inversion of a radioactive source term in a lead-bismuth reactor break leakage accident within a short monitoring window by using a parameter-level weak constraint physical information neural network.

Description

Technical Field

[0001] This invention relates to the field of nuclear reactor radiation safety technology, and particularly to a lead-bismuth cooled microreactor. 210 A fast inversion method and system for multi-task source items in Po. Background Technology

[0002] Lead-bismuth cooled microreactors, with their high energy density and inherent safety advantages, are a core reactor type for power supply in remote areas and deep-sea nuclear energy equipment. However, the lead-bismuth coolant is activated by neutrons to generate... 210 Polonium (Po) possesses strong alpha radioactivity and high volatility. After a leak occurs through a breach in the reactor core, it spreads as hydrogen polonium, making it a core safety risk source for lead-bismuth reactors. According to the IAEA SSR-2 / 1 nuclear safety specifications, after an accident, source term inversion must be performed rapidly and accurately to determine the leak location, release intensity, and delineate the radiation evacuation zone. Currently... 210 Po source term inversion techniques are mainly divided into two categories: traditional physical model inversion techniques and pure data-driven deep learning inversion techniques.

[0003] Traditional physical model inversion techniques, including Kalman filtering and Bayesian inverse modeling, rely on iterative calculations of a complete atmospheric / crater diffusion model. For example, one source term assimilation inversion scheme based on ensemble Kalman filtering corrects leakage source parameters by iteratively solving a crater diffusion numerical model. This scheme has complete physical logic and strong interpretability, but its core relies on multiple rounds of high-precision numerical simulation iterative calculations. A single complete inversion can take several minutes, which cannot meet the needs of short-term emergency inversion for accidents, and lacks the implementation processes related to feature engineering and multi-task deep learning inference.

[0004] Pure data-driven deep learning inversion techniques combine deep feedforward networks, TCN networks, and CNN networks, directly using time-series data from concentration monitoring to train the model, resulting in fast inference speed. For example, a single-task inversion scheme for multi-nucleoside source items based on a TCN-2DCNN fusion clustering model integrates TCN and 2D-CNN to extract time-series image features, achieving only a single regression of multi-nucleoside release amounts, with low prediction error and good noise resistance. Another example is the static source distribution inversion scheme based on SDICNN convolutional networks, which designs SDICNN convolutional networks to reconstruct two-dimensional static source distribution images by sampling radiation parameters and extracts spatial distribution features through convolution. However, these techniques have three major drawbacks: they only utilize local data from the original measurement points, losing the global spatial structure of the concentration field; they lack physical constraints, and the predicted values ​​under turbulent disturbances do not conform to the physical laws of diffusion; and they only support a single task (classification / regression), unable to simultaneously output accident type and leakage height. Summary of the Invention

[0005] Based on this, the present invention addresses the problems of insufficient response speed and limited inversion accuracy in the inversion of radioactive source terms in lead-bismuth reactor rupture leakage accidents, and provides a lead-bismuth cooled microreactor. 210 The Po multi-task source term rapid inversion method and system can achieve high-precision rapid inversion of leakage source terms under limited monitoring point conditions, and can provide a reference for lead-bismuth reactor accident diagnosis and emergency response.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a lead-bismuth cooled microreactor. 210 The Po multi-task source item fast inversion method includes the following steps: S1. Obtain the target leakage parameters under the target reactor leakage condition; S2. Establish corresponding conditions for the target leakage scenario. 210 The Po transport-diffusion model inputs the target leakage parameters to... 210 Po transport and diffusion model, data collection 210 The three-dimensional dynamic concentration distribution of Po was obtained. 210 Time series data of Po concentration; S3. Based on the above 210 15-dimensional enhanced spatiotemporal features were extracted from time-series data of Po concentration, and then a dataset was constructed and randomly divided into training set, validation set and test set. S4. Construct a physical information neural network model that integrates physical constraints and multi-task learning mechanisms, and use the training set and validation set to train and optimize the physical information neural network model; S5. Input the test set into the trained and optimized physical information neural network model, and output... 210 Po leakage inversion results.

[0007] As a further improvement to the above-described solution of the present invention, in step S1, the target leakage condition includes: The leakage at the bottom center of the main vessel is caused by localized, penetrating failure of the bottom end cap due to the coupled effects of long-term thermal cycling and liquid metal corrosion. The breach is located at the center of the bottom of the main vessel, and the breach height is... ; The leakage condition on the lower sidewall of the main vessel is caused by low-cycle fatigue cracking in the welded area or supporting structure of the lower part of the main vessel under thermal fatigue and stress concentration. The rupture point is located on the lower sidewall of the main vessel, and the height of the rupture point is... satisfy: , This represents the initial liquid level of the lead-bismuth coolant in the main container; The leakage condition on the upper sidewall of the main vessel is caused by the propagation of thermal fatigue cracks in the upper cylindrical region of the main vessel under the effects of thermal stratification and thermal transients. The rupture point is located on the upper sidewall of the main vessel, and the height of the rupture point is... satisfy: ; The leakage condition on the upper sidewall of the main container is caused by thermo-mechanical coupling cracking in the upper part of the main container near the free surface of the coolant, under the coupled effect of severe temperature fluctuations at the gas-liquid interface and structural vibration. The rupture point is located on the upper sidewall of the main container, at a height of [missing information]. satisfy: .

[0008] As a further improvement to the above-described solution of the present invention, in step S1, the target leakage parameters include: (1) Leakage velocity :

[0009] In the formula: Represents the driving pressure difference. , Represents the pressure of the covered gas. This represents the density of lead-bismuth coolant (considering temperature variations), such as in the temperature range of 300°C to 600°C. ρ Between 10300 and 10700 kg / m 3 between; h 0 represents the initial liquid level of lead-bismuth coolant in the main container; h b This represents the height of the leak rupture location from the bottom of the main container; Represents the frictional resistance coefficient along the path. , Re Represents the Reynolds number, , The dynamic viscosity of lead-bismuth coolant; Represents the inlet local drag coefficient. =0.5; d This represents the diameter of the leak point rupture. L This represents the length of the rupture channel at the leakage point; (2) Mass flow rate :

[0010] In the formula, The cross-sectional area of ​​the breach representing the leak point; (3) 210 Po evaporation release boundary area :

[0011] In the formula, represent t The cumulative leakage volume of lead-bismuth coolant at any given time. , This represents the thickness of the lead-bismuth coolant spread.

[0012] As a further improvement to the above-described solution of the present invention, in step S2, the... 210 The Po transport and diffusion model was constructed using multiphysics simulation software, employing asymmetric ventilation conditions. Based on flow field calculations, it was described using the rare matter transport equation. 210 The diffusion behavior of Po; a mass reduction term based on the effective lifetime τ is introduced into the rarefied mass transport equation. R :

[0013] In the formula: C represent 210 Po concentration; τ represents... 210 The effective lifetime of Po in air, τ=150s.

[0014] As a further improvement to the above-described solution of the present invention, in step S2, by... 210 The Po transport and diffusion model includes N discrete monitoring points with three-dimensional coordinates, and a monitoring sensor is installed at each monitoring point. i The coordinates of the monitoring sensors are ,1≤ i ≤N; After the target leakage condition occurs, each of the monitoring sensors collects data at sampling intervals of 0.1 to 1 second within a predetermined time period. 210 Po concentration was obtained from data collected by various monitoring sensors. 210 Po concentration time series data .

[0015] This invention makes fuller use of short-time monitoring time series samples. 120s time series measurement data can be used to construct multi-dimensional physical feature samples in batches. It can extract effective supervision information from multiple dimensions such as spatiotemporal gradient, cloud distribution, and diffusion transport, rather than just directly inputting the original concentration time series.

[0016] As a further improvement to the above-described solution of the present invention, in step S3, the 15-dimensional enhanced spatiotemporal features include: The spatiotemporal field has strong dimensional characteristics, including peak activity. Cumulative field energy and local high intensity ratio ;in: represent t Real-time collection 210 Po concentration; , Representing the i Data collected by monitoring sensors 210 Po concentration For the preset concentration threshold, Representative of collection 210 Po concentration exceeds preset concentration threshold The number of monitoring sensors; Spatial distribution dimensional characteristics, including concentration-weighted spatial centroid coordinates. Centroid offset distance L c = Global spatial discreteness and diffusion anisotropy ratio ;in: x c 、y c 、z c These represent the X, Y, and Z coordinates of the centroid of the concentration-weighted space, respectively, in meters. Representing the i Data collected by monitoring sensors 210 Po concentration, L c This represents the offset distance of the centroid of the concentration-weighted space relative to the center of the main container on the horizontal plane. , , These represent the concentration-weighted variances of the concentration field in the X, Y, and Z directions, respectively. , , ; Physical transport dimension features, including concentration-time gradient Spatial Laplace operator Normalized relative curvature K= and global average spatial gradient ;in, Represents a certain monitoring point 210 Changes in Po concentration over time; C Represents the spatial gradient, indicating the same time interval. 210 Changes in Po concentration with varying monitoring point locations; , , For the first i Concentration gradient components collected by a monitoring sensor; Deformation structural dimensional characteristics, including concentration distribution skewness kurtosis and structural information entropy ;in,r i Representing the i Spatial distance from each monitoring point to the centroid of the concentration-weighted space ; p i Representative monitoring point 210 The probability distribution of Po concentration after normalization .

[0017] This invention introduces physical enhancement features such as peak activity, diffusion anisotropy ratio, spatial Laplacian operator, and concentration distribution skewness into the input features, which are closely aligned with the crater. 210 The true transport mechanism of Po convection diffusion and sedimentation attenuation is preserved. 210 Physical mechanism of Po diffusion.

[0018] As a further improvement to the above-mentioned scheme of the present invention, step S3 also includes: using the SMOTE synthetic minority class oversampling algorithm to generate synthetic samples for target leakage conditions with a small number of samples, eliminating class distribution imbalance; and using Z-score standardization to eliminate the difference in the 15-dimensional enhanced spatiotemporal feature dimensions. The SMOTE synthetic minority class oversampling algorithm is used to weighted sample to balance the differences in the number of high-order and low-order samples, assigning higher training weights to boundary, high-risk, and difficult-to-distinguish samples, thus avoiding low recognition accuracy for a small number of boundary condition samples.

[0019] As a further improvement to the above-mentioned scheme of the present invention, in step S4, the physical information neural network model includes an input layer, a shared feature extraction layer, a leakage condition classification branch, and a leakage height regression branch; the shared feature extraction layer is used to extract high-order nonlinear spatiotemporal information from the 15-dimensional enhanced spatiotemporal features of the input; the leakage condition classification branch is connected after the shared feature extraction layer and includes two fully connected layers and a classification output layer, used to identify the output leakage condition type; the leakage height regression branch is connected after the shared feature extraction layer and includes two fully connected layers and a linear output layer, used to output the height of the leakage rupture location from the bottom of the main container.

[0020] As a further improvement to the above-described solution of the present invention, during the training process in step S4: consider 210 Migration behavior of Po in an asymmetric ventilation flow field 210 The concentration field of Po satisfies the following convection-diffusion control equation:

[0021] In the formula: u Represents the velocity vector of the flow field; D Represents the equivalent diffusion coefficient; R(C) Represents quality reduction items; Represents the source of leakage. , per unit area of ​​lead-bismuth surface 210 Po evaporation release flux. ; The Dirac function represents the source term being located at the bottom center of the main container; This is an indicator function, representing that the release source exists only in the area covered by the liquid pool; This represents the area covered by the liquid pool at the bottom of the crater; Based on the height of the leak rupture location from the bottom of the main container predicted by the physical information neural network model, the driving pressure difference and leakage velocity are first calculated. 210 Po evaporation release boundary area, combined with 210 The evaporation flux of Po was then used to derive the theoretical equivalent source strength corresponding to the predicted height of the leak rupture location from the bottom of the main container. ;

[0022] In the formula, per unit area of ​​lead-bismuth surface 210 Po evaporation release flux. The predicted location of the leak rupture is at time t, corresponding to the distance from the bottom of the main container. 210 Po evaporation release boundary area; Based on the peak activity and cumulative field energy characteristics, the measured equivalent source strength, used to characterize the actual observation results of the monitoring data, is then calculated. ;

[0023] in, , The inherent linear scaling factor of the crater flow field; with theoretical equivalent source strength Equivalent source strength as measured The difference constructs the physical residual - When the model prediction results satisfy the physical transport laws, the physical residuals Approaching zero; for Take the mean square error as the physical residual loss term. ; Construct the total loss function:

[0024] In the formula, L class Cross-entropy loss for representative working conditions classification, , y i This is a label indicating a genuine leak of operating conditions. The probability of the leakage condition category predicted by the model;L reg This represents the high regression mean square error loss due to leakage. , h i This represents the actual location of the leak / rupture point from the bottom of the main container. The height of the predicted leak rupture location from the bottom of the main container; L phys Represents the physical residual loss term; The cross-entropy loss weight coefficients are used for classifying operating conditions. The weighting coefficients for the mean squared error loss during leakage height regression are... This is the weighting coefficient for the physical residual loss term.

[0025] The Physical Information Neural Network (PINN) constructed by this method is a parameter-level weak constraint variant specifically for source term inversion. Unlike the classic forward PINN, which uses a concentration field as the output point-by-point strong constraint mode, the physical constraints of this method directly act on the leakage source parameters in the inversion output. It can ensure that the inversion results conform to the physical laws of convection and diffusion without the need for the network to output a full-space concentration field, thus meeting the rapid inversion needs of accident emergency scenarios.

[0026] The present invention also provides a lead-bismuth cooled micro reactor. 210 The Po multi-task source term fast inversion system is used to realize lead-bismuth cooled microreactors as described above. 210 The Po multi-task source item fast inversion method includes: The acquisition module is used to acquire the target leakage condition of the target reactor and the target leakage parameters under the target leakage condition; The simulation module is used to establish a corresponding simulation model for the target leakage condition. 210 The Po transport-diffusion model inputs the target leakage parameters to... 210 Po transport and diffusion model, data collection 210 The three-dimensional dynamic concentration distribution of Po was obtained. 210 Time series data of Po concentration; The feature extraction and dataset construction module is used to extract features based on the data. 210 15-dimensional enhanced spatiotemporal features were extracted from time-series data of Po concentration, and then a dataset was constructed. The dataset was randomly divided into training set, validation set and test set. The model building and training module is used to build a physical information neural network model that integrates physical constraints and multi-task learning mechanisms, and to train and optimize the physical information neural network model using the training set and validation set. The inversion module is used to input the test set into the trained and optimized physical information neural network model and output... 210 Po leakage inversion results.

[0027] Compared with the prior art, the present invention has the following beneficial effects: This invention enables rapid inversion of radioactive source terms in lead-bismuth reactor rupture and leakage accidents within a short monitoring window (120s), achieving accident scene identification and leakage height estimation. While maintaining high inversion accuracy, it also ensures physical consistency, significantly improving response time to sudden accidents and saving valuable time for emergency decision-making. High-precision source term inversion can be achieved with only a limited number of monitoring points, reducing reliance on dense sensor deployment. It has strong engineering practicality and promotional value, providing reliable technical support for lead-bismuth microreactor accident diagnosis, consequence assessment, and emergency response.

[0028] This invention effectively characterizes the spatiotemporal distribution of concentration fields under complex flow conditions by constructing a 15-dimensional enhanced spatiotemporal feature system that integrates spatial centroid, anisotropy, concentration gradient, and structural information entropy, providing a high-quality input representation for the model. Simultaneously, by introducing a physical residual term based on the Fick diffusion equation as a constraint, the multi-task neural network achieves physical consistency in the joint prediction of leakage condition classification and leakage height regression, achieving a classification accuracy of 0.999 and a mean absolute error of only 0.086 m for leakage height inversion, with a coefficient of determination R0. 2 The accuracy reached 0.9854, demonstrating extremely high inversion accuracy. The 15-dimensional enhanced spatiotemporal features and physical residual loss term can independently and synergistically improve the model's predictive stability and robustness, especially under the condition of leakage at the bottom of the main container. The error sources under the condition of leakage at high level were also clearly revealed, providing direction for future improvements. Attached Figure Description

[0029] Figure 1 A lead-bismuth cooled microreactor provided in this embodiment of the invention 210 Flowchart of the Po multi-task source item fast inversion method; Figure 2 The flow field within the lead-bismuth reactor crater in this embodiment of the invention is... 210 Simulation cloud map of Po diffusion distribution; Figure 3 This is a diagram illustrating the physical information neural network model framework in an embodiment of the present invention. Figure 4 This is a graph showing the performance verification results of the physical information neural network model under different target leakage conditions in the embodiments of the present invention; Figure 5 This is a box plot showing the distribution of the absolute error of the physical information neural network model under different target leakage conditions in this embodiment of the invention. Detailed Implementation

[0030] To facilitate understanding of the present invention, a more comprehensive description will be given below with reference to specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0032] Reference Figure 1 This embodiment proposes a lead-bismuth cooled microreactor. 210 The Po multi-task source item fast inversion method includes the following five steps S1-S5.

[0033] S1. Obtain the target leakage parameters under the target reactor's target leakage condition.

[0034] This embodiment addresses a pool-type natural circulation lead-bismuth reactor. The leak scenario assumes a partial rupture in the main vessel due to thermal fatigue or manufacturing defects, causing lead-bismuth coolant to leak into the protective vessel. As the accident progresses, the protective vessel also ruptures, leading to the leakage of lead-bismuth coolant. 210 The lead-bismuth coolant in Po eventually enters the crater space. The main vessel failure is a thin-walled, sharp-edge breach, and the lead-bismuth coolant flows abruptly from the volumetric cavity into the narrow breach, which is a typical sharp-edge inlet flow.

[0035] To comprehensively assess the radioactive safety risks of a lead-bismuth reactor at different breach locations, this embodiment sets up four typical lead-bismuth coolant target leakage scenarios. The bottom of the main container is set as the height reference z=0, and the initial liquid level of the lead-bismuth coolant in the main container is [missing information]. In this embodiment, =5.16m. The four target leakage conditions are: leakage at the bottom center of the main vessel, leakage at the lower side wall of the main vessel, leakage at the upper side wall of the main vessel, and leakage at the upper side wall of the main vessel.

[0036] (1) The leakage condition at the bottom center of the main vessel, denoted as Case 2, is caused by the localized through-hole failure of the bottom end cap of the main vessel under the coupled action of long-term thermal cycling and liquid metal corrosion. The rupture point is located at the bottom center of the main vessel. Simulate the failure of the bottom end cap.

[0037] (2) Leakage condition of the lower sidewall of the main vessel, denoted as Case 1, is caused by low-cycle fatigue cracking of the weld area or supporting structure of the lower part of the main vessel under thermal fatigue and stress concentration. The height of the crack location is as follows: satisfy: The rupture point is located at a height of 0 to 1.032 meters from the bottom of the main vessel, simulating a leak in the weld seam at the bottom of the main vessel. In this embodiment, the rupture point for the leak in the lower sidewall of the main vessel is located at a height of 0.43 meters from the bottom of the main vessel. .

[0038] (3) The leakage condition in the upper sidewall of the main vessel, denoted as Case 4, is caused by the propagation of thermal fatigue cracks in the upper cylindrical region of the main vessel under the action of thermal stratification and thermal transients. The height of the fracture location is [not specified]. satisfy: The breach is positioned at a height of 3.096–4.128 m from the bottom of the main container, simulating damage to the upper structure of the main container. In this embodiment, the breach in the upper sidewall leakage condition of the main container is positioned at a height of 3.9 m from the bottom of the main container. .

[0039] (4) Leakage on the upper side wall of the main container, denoted as Case 3, is caused by thermal-mechanical coupling cracking in the upper part of the main container near the free liquid surface of the coolant under the coupling effect of severe temperature fluctuations at the gas-liquid interface and structural vibration. The height of the crack location is as follows: satisfy: The rupture point is located at a height of 4.128–5.16 m from the bottom of the main container, close to the liquid surface, simulating a high-level crack near the liquid surface. In this embodiment, the rupture point for the leakage condition on the upper side wall of the main container is located at a height of 4.8 m from the bottom of the main container. .

[0040] Each target leakage condition is individually coded and used as a classification task supervision label. The leakage condition at the bottom center of the main container is coded as [1,0,0,0], the leakage condition at the lower side wall of the main container is coded as [0,1,0,0], the leakage condition at the upper side wall of the main container is coded as [0,0,1,0], and the leakage condition at the upper side wall of the main container is coded as [0,0,0,1].

[0041] The four typical target leakage conditions set in this embodiment are not randomly selected, but are engineering abstractions based on the typical failure modes that may occur in the main container of the lead-bismuth reactor under the coupled effects of thermal stress, liquid metal corrosion and structural fatigue. They cover the main axial leakage areas that may occur in the main container of the lead-bismuth reactor and can reflect the migration law of radioactivity in the asymmetric ventilation flow field under different height conditions more completely.

[0042] The target leakage parameters for each target leakage condition include: leakage velocity, mass flow rate, and evaporation release boundary area. To provide physically consistent source term inputs for numerical simulation, this embodiment quantitatively calculates the target leakage parameters for different target leakage conditions based on a short-pipe outflow model.

[0043] The leakage driving force mainly originates from the pressure of the covering gas. P gas The sum of the static pressure generated by the lead-bismuth liquid column above the rupture point drives the pressure difference. ΔP The initial leakage rate is determined by the following formula:

[0044] In the formula: Represents the driving pressure difference; Represents the pressure of the covered gas; This represents the density of lead-bismuth coolant (considering temperature variations), such as in the temperature range of 300°C to 600°C. ρ Between 10300 and 10700 kg / m 3 between; h 0 represents the initial liquid level of lead-bismuth coolant in the main container; h b This represents the height of the leak rupture location from the bottom of the main container.

[0045] Leakage flow rate The calculation formula is:

[0046] In the formula: Represents the frictional resistance coefficient along the friction path. , Re Represents the Reynolds number, , Represents the dynamic viscosity of lead-bismuth coolant; Represents the inlet local drag coefficient. =0.5; d This represents the diameter of the leak point rupture. L This represents the length of the rupture channel at the leak point.

[0047] After determining the leakage velocity, the mass flow rate is further calculated. :

[0048] In the formula, This represents the cross-sectional area of ​​the breach.

[0049] Since the liquid level in the main container drops very little within 10 minutes, the leakage process can be considered as quasi-steady-state flow, with the cumulative leakage volume... It increases linearly with time. Assuming the leaked lead-bismuth coolant spreads uniformly at the bottom of the crater, the liquid pool changes over time. 210 Po evaporation release boundary area for:

[0050] In the formula, represent t The cumulative leakage volume of lead-bismuth coolant at any given time. , This represents the thickness of the lead-bismuth coolant layer. =5mm.

[0051] S2. Establish corresponding conditions for the target leakage scenario. 210 The Po transport-diffusion model inputs the target leakage parameters to... 210 Po transport and diffusion model, data collection 210 The three-dimensional dynamic concentration distribution of Po was obtained. 210 Time series data of Po concentration.

[0052] This embodiment uses COMSOL simulation software to construct corresponding simulations for each target leakage condition. 210 The Po transport and diffusion model employs asymmetric ventilation conditions, such as... Figure 2 As shown, Figure 2 'a' represents the streamline velocity field distribution inside the crater. Figure 2 b represents the velocity field distribution on the surface of the crater. Figure 2 c is the cross-section 210 The Po concentration field is superimposed with the streamline velocity field. Considering the transport characteristics under accident conditions, the model describes the transport characteristics using the rarefied substance transport equation, based on the flow field calculations. 210 The diffusion behavior of Po.

[0053] Considering 210 The deposition effect of Po in air introduces a mass reduction term based on the effective lifetime τ into the rarefied mass transport equation. R :

[0054] In the formula: C represent 210 Po concentration; τ represents... 210 The effective lifetime of Po in air, τ=150s.

[0055] Through 210 The Po transport and diffusion model includes N discrete monitoring points with three-dimensional coordinates, and a monitoring sensor is installed at each monitoring point. i The coordinates of the monitoring sensors are ,1≤ i ≤N. Input the target leakage parameters into the corresponding... 210 In the Po transport and diffusion model, after the target leakage condition occurs, within 120 seconds, each monitoring sensor collects data at sampling intervals of 0.1 to 1 second. 210 Po concentration was obtained from data collected by various monitoring sensors. 210 Po concentration time series data .

[0056] This embodiment deployed 13 monitoring sensors (i.e., N=13), and the coordinates of each sensor are shown in Table 1 below. A total of 6123 valid samples were generated for the four target leakage scenarios.

[0057] Table 1. Three-dimensional coordinates of the monitoring sensor

[0058] S3. Based on 210 We extracted 15-dimensional enhanced spatiotemporal features from the time series data of Po concentration, and then constructed a dataset, which was randomly divided into training set, validation set and test set.

[0059] To enhance the model's ability to perceive the spatial asymmetry arising from four target leakage conditions, this embodiment constructs 15-dimensional enhanced spatiotemporal features, covering four dimensions: spatiotemporal field strength, spatial distribution, physical transport, and deformation structure. The spatiotemporal field has strong dimensional characteristics, including peak activity. Cumulative field energy and local high intensity ratio ;in: represent t Real-time data collection 210 Po concentration; , Representing the i Data collected by monitoring sensors 210 Po concentration For the preset concentration threshold, Representative of collection 210 Po concentration exceeds preset concentration threshold The number of monitoring sensors; Spatial distribution dimensional characteristics, including concentration-weighted spatial centroid coordinates. Centroid offset distance L c = Global spatial discreteness and diffusion anisotropy ratio ;in: x c 、y c 、z c These represent the X, Y, and Z coordinates of the centroid of the concentration-weighted space, respectively, in meters. Representing the i Data collected by monitoring sensors 210 Po concentration, L cThis represents the offset distance of the centroid of the concentration-weighted space relative to the center of the main container on the horizontal plane. It is used to characterize the degree of offset of the leakage source term in the horizontal direction. The larger the value, the more the leakage location is offset to one side of the main container. , , These represent the concentration-weighted variances (second-order central moments) of the concentration field in the X, Y, and Z directions, respectively, used for quantization. 210 The degree of dispersion of Po concentration distribution in each direction, , , Global spatial discreteness The unit is square meters; Physical transport dimension features, including concentration time gradient Spatial Laplace operator Normalized relative curvature K= and global average spatial gradient ;in, Represents a certain monitoring point 210 Changes in Po concentration over time; C Represents the spatial gradient, indicating the same time interval. 210 Changes in Po concentration with varying monitoring point locations; , , For the first i The concentration gradient components are collected by several monitoring sensors. The global average spatial gradient feature quantifies the overall spatial variation of the concentration field. The larger the magnitude of the global average spatial gradient, the more drastic the spatial variation of the concentration, which usually corresponds to a high gradient region near the pool leak source. If the monitoring points are a uniform grid, the global average spatial gradient can be approximated by finite difference. In this embodiment, the above-mentioned concentration temporal gradient, spatial Laplacian operator, and other micro-classification features are all obtained by directly differentiating the three-dimensional dynamic concentration obtained in step S2 using automatic differentiation technology. Compared with the traditional finite difference method, the numerical accuracy is higher, which can further improve the physical accuracy of the input features.

[0060] Deformation structural dimensional characteristics, including concentration distribution skewness kurtosis and structural information entropy ;in, r i Representing the i Spatial distance from each monitoring point to the centroid of the concentration-weighted space ; p i Representative monitoring point 210The probability distribution of Po concentration after normalization .

[0061] For each target leakage scenario, a separate dataset is created, and then all samples in each dataset are divided into training set, validation set, and test set according to the proportions of 70%, 20%, and 10% respectively, in order to avoid information leakage between training data and test data.

[0062] To address the issue of uneven sample distribution across different operating conditions, this paper employs the SMOTE (Synthetic Minority Oversampling Technique) algorithm to generate synthetic samples for target leakage conditions with insufficient sample numbers, thereby eliminating class distribution imbalance and improving the model's stability and robustness in the classification boundary region. Z-score normalization is used to eliminate the dimensional differences of the 15-dimensional enhanced spatiotemporal features.

[0063] S4. Construct a physical information neural network model that integrates physical constraints and multi-task learning mechanisms, and use training and validation sets to train and optimize the physical information neural network model.

[0064] like Figure 3 As shown, the Physics-Informed Neural Network (PINN) model in this embodiment, which integrates physical constraints and multi-task learning mechanisms, consists of an input layer, a shared feature extraction layer, a condition classification branch, and a leakage height regression branch. Specifically: the shared feature extraction layer extracts high-order nonlinear spatiotemporal information from the 15-dimensional enhanced spatiotemporal features of the input, employing the SiLU function as the activation function and introducing layer normalization to improve the training stability of deep networks; the leakage condition classification branch, connected after the shared feature extraction layer, includes two fully connected layers and a classification output layer, used to identify the output leakage condition type; the leakage height regression branch, connected after the shared feature extraction layer, includes two fully connected layers and a linear output layer, used to output the height of the leakage breach location from the bottom of the main container.

[0065] The physical information neural network model was trained using a training set, and end-to-end optimization was performed using the AdamW optimizer and a total loss function. It took 15-dimensional enhanced spatiotemporal features as input and the leakage condition type and leak rupture location height as output, with a maximum training epoch count of 150. After each epoch, the model performance was evaluated using a validation set. Performance metrics included accuracy, precision, recall, and F1 score. The weights of the models with the highest F1 scores on the validation set were saved.

[0066] In this embodiment, during the model training process: consider210 Migration behavior of Po in an asymmetric ventilation flow field 210 The concentration field of Po satisfies the following convection-diffusion control equation:

[0067] In the formula: u Represents the velocity vector of the flow field; D Represents the equivalent diffusion coefficient; R(C) Represents quality reduction items; The term representing the leakage source indicates the amount of water released into the crater from the leakage point per unit time and unit volume. 210 Po radioactivity, in units of Bq / (m²) 3 .s), , per unit area of ​​lead-bismuth surface 210 Po evaporation release flux. ; The Dirac function represents the source term being located at the bottom center of the main container; This is an indicator function, representing that the release source exists only in the area covered by the liquid pool; This represents the area covered by the liquid pool at the bottom of the crater.

[0068] It should be noted that this embodiment sets up two independent physical enhancement stages, which are applied to the feature extraction stage and the model training stage respectively. Their objectives and implementation methods do not overlap. High-precision differential calculation in the feature extraction stage: The step of automatically calculating the partial derivative of the concentration field is applied to the feature extraction process in step S3. For differential input features such as concentration time gradient and spatial Laplacian operator, the three-dimensional dynamic concentration field obtained in step S2 is directly differentiated by automatic differential technology. Compared with the traditional finite difference method, the numerical accuracy is higher, which is used to improve the physical accuracy of the input features.

[0069] Weak constraints at the parameter level during the training phase: Based on the above convection-diffusion control equations, a physical residual loss term is constructed to directly constrain the leakage source parameters output by the network inversion. This ensures that the inversion results conform to the physical laws of convection and diffusion without requiring the network to output a full-space concentration field. The specific implementation is as follows: During model training, physical residuals are constructed based on the Fick diffusion equation. To improve the physical consistency of model prediction results, specifically: The height of the leak rupture location from the bottom of the main container, predicted by the physical information neural network model, is substituted into the driving pressure difference, leakage velocity, and mass flow rate from step S1. 210 The formula for calculating the boundary area of ​​Po evaporation release, combined with this step. 210 The formula for Po evaporation release flux is used to derive the theoretical equivalent source strength corresponding to the predicted leakage height. ;

[0070] In the formula, per unit area of ​​lead-bismuth surface 210 Po evaporation release flux. The predicted location of the leak rupture is at time t, corresponding to the distance from the bottom of the main container. 210 Po evaporation release boundary area.

[0071] Based on the positive correlation between source intensity and concentration in the convection-diffusion control equation, the measured equivalent source intensity corresponding to the monitoring data is calculated from the peak activity and cumulative field energy characteristics in the 15-dimensional enhanced spatiotemporal features. ;

[0072] in, , This is the inherent linear scaling factor of the flow field in the crater, which is a conventional flow field calibration parameter in this field.

[0073] With theoretical equivalent source strength Equivalent source strength as measured The difference constructs the physical residual - The physical residual loss term is obtained by taking the mean square error of the physical residual. When the model prediction results satisfy the laws of physical transport, the physical residual term... Approaching zero.

[0074] The physical residual loss term is directly derived from the leakage height output by the network. During training, the gradient can be backpropagated to the shared feature extraction layer and the two task branches, forcing the network to learn feature mappings that conform to the physical laws of diffusion, thus avoiding non-physical inversion results from fitting flow field noise.

[0075] By introducing physical constraints, the model can not only learn the statistical characteristics in the monitoring data, but also maintain physical consistency with the transport and diffusion laws during the prediction process, thereby effectively suppressing non-physical prediction oscillations caused by local noise and sparse measurement points.

[0076] Construct the total loss function:

[0077] In the formula, L class Cross-entropy loss for representative working conditions classification, , y i This is a label indicating a genuine leak of operating conditions. The probability of the leakage condition category predicted by the model; L reg This represents the high regression mean square error loss due to leakage. , h i This represents the actual distance between the leak location and the bottom of the main container. The height of the leak rupture location predicted by the model from the bottom of the main container; L phys Represents the physical residual loss term; The cross-entropy loss weighting coefficients are used for classifying operating conditions. The weighting coefficients for the mean squared error loss during leakage height regression are... This is the weighting coefficient for the physical residual loss term.

[0078] To improve the model's generalization ability to complex operating conditions, random perturbations and Gaussian noise are further introduced during the training phase to simulate the impact of actual sensor measurement errors and local flow field fluctuations on concentration monitoring results.

[0079] S5. Input the test set into the trained and optimized physical information neural network model, and output... 210 Po leakage inversion results.

[0080] The 15-dimensional enhanced spatiotemporal features of the test set are input into the trained physical information neural network model, and the output is... 210 The Po leakage inversion results were used to comprehensively evaluate the model classification performance using metrics such as accuracy, precision, recall, and F1 score. The results are shown in Table 2.

[0081] Table 2 Performance Indicators for Target Leakage Condition Identification and Classification

[0082] As shown in Table 2, the proposed physical information neural network model exhibits high recognition accuracy under all four typical leakage conditions, with an overall classification accuracy of 0.999. Due to the significant differences in leakage height among the four leakage conditions, and under the influence of an asymmetric ventilation flow field, its... 210 Po plumes exhibit significant differences in spatial centroid evolution, diffusion direction, and local concentration distribution structure, thus demonstrating strong physical separability across different operating conditions. Specifically, Case 1 and Case 2, representing low-level leakage conditions, show relatively more stable identification performance. Case 3, however, suffers from some unsteady fluctuations in its concentration field due to top exhaust disturbances and local backflow, resulting in slightly lower classification accuracy compared to other conditions. Overall, however, the physical information neural network model can still reliably extract the spatial structural features corresponding to different operating conditions.

[0083] Furthermore, the coefficient of determination for the continuous regression task of leakage height in this embodiment reaches 0.9854, indicating that the model can accurately learn the evolution of the concentration field under different height conditions. The overall false alarm rate (FAR) of the model remains at a low level, indicating that the proposed physical information neural network model still has good stability in identifying operating conditions under complex flow field conditions.

[0084] Figure 4 , Figure 5 The performance verification results and absolute error distribution box plots under four target leakage conditions are presented in a visually intuitive way. Figure 4 (a) shows the performance verification results under the target leakage condition Case 1; (b) shows the performance verification results under the target leakage condition Case 2; (c) shows the performance verification results under the target leakage condition Case 3; and (d) shows the performance verification results under the target leakage condition Case 4. It can be observed that the median errors of Case 1, Case 2, and Case 4 are all at a low level, and the enclosure exhibits high compactness. Although Case 3 shows a predicted outlier of approximately 0.5m, its overall central trend still accurately anchors the height range of the leakage source. This deep fusion inversion method based on data perception and physical constraints not only improves the accuracy of accident diagnosis but also demonstrates the effectiveness of 15-dimensional enhanced spatiotemporal features in extracting nonlinear concentration field structure information, providing a physically consistent prediction benchmark for rapid source term inversion in the early stages of an accident.

[0085] To reveal the contributions of feature engineering and physical constraints to the inversion results, an ablation experiment was designed in this embodiment. By employing the variable elimination method, three comparative schemes were constructed while maintaining consistency in the basic architecture such as network depth and activation function. The aim was to quantify the specific performance of 15-dimensional enhanced spatiotemporal features, physical residual loss term, and SMOTE synthetic minority class oversampling algorithm in source term inversion. The results are shown in Table 3.

[0086] Table 3 Ablation Experiment

[0087] Comparing Scheme A with the scheme in this embodiment, it can be found that when the model loses the support of the 15-dimensional enhanced spatiotemporal features and directly processes the original... 210When using Po concentration data, the regression accuracy experienced a sharp decline, with the MAE increasing dramatically from 0.0860 m to 0.4215 m. Physical analysis revealed that the diffusion process of radionuclides within the crater exhibits highly spatiotemporal coupling characteristics. The original monitoring data only provides scattered, localized concentration sampling points, making it difficult to characterize the global asymmetric evolution trend of the flow field. However, the normalized Laplace operator and anisotropy ratio in the 15-dimensional enhanced spatiotemporal features successfully transformed the discrete signal into a physical descriptor with global semantics. Experimental data demonstrate that this feature system plays a decisive role in mitigating the ill-posedness of the inversion problem and is a key cornerstone for achieving accurate positioning.

[0088] Scheme B validated the necessity of the physical residual loss term. While retaining the 15-dimensional enhanced spatiotemporal features and removing the physical residual loss term, the model's classification accuracy remained at 0.9714, but the MAE increased from 0.0860 m on the baseline model to 0.1507 m. Further observation of the error distribution revealed that the model lacking the physical residual loss term was prone to non-physical prediction jitter in samples such as Case 3, which are heavily influenced by ventilation flow fields. The physical residual loss term forces the model to follow Fick's diffusion law at the gradient level, embedding physical constraints within the black-box neural network. This mechanism can act as a numerical calculation anchor, effectively filtering out spurious features caused by monitoring noise and ensuring that the inversion results still conform to the physical evolution logic under extreme conditions.

[0089] Scheme C examines the impact of data balancing techniques on model performance. Simulation data of lead-bismuth pile leak accidents often suffer from uneven distribution of operating conditions. Comparing Scheme C with the scheme in this embodiment, it can be found that after introducing the SMOTE synthetic minority class oversampling algorithm for synthetic oversampling, the model classification accuracy increased from 0.9571 to 1.000, and the MAE decreased from 0.1875 m to 0.0860 m. This indicates that sample balancing techniques can significantly improve the discrimination of samples near the classification boundary and indirectly improve the stability of the regression task.

[0090] In summary, the 15-dimensional enhanced spatiotemporal features and physical residual loss term proposed in this invention are not simply superimposed, but rather deeply complementary at the levels of physical representation and numerical driving. The 15-dimensional enhanced spatiotemporal features are responsible for capturing the structural heterogeneity of the concentration field at the perception layer, while the physical residual loss term ensures that the prediction results conform to the fundamental laws of thermal hydraulics at the decision layer. This synergistic mechanism enables the model to demonstrate high comprehensive performance and engineering practical potential in the lead-bismuth microreactor source term inversion task.

[0091] This embodiment is based on the above lead-bismuth cooled microreactor. 210 Based on the Po multi-task source term fast inversion method, a lead-bismuth cooled microreactor is also provided. 210The Po multi-task source item fast inversion system is used to implement this system, which includes: The acquisition module is used to acquire target leakage parameters under target reactor leakage conditions. The simulation module is used to establish a corresponding simulation model for the target leakage condition. 210 The Po transport-diffusion model inputs the target leakage parameters to... 210 Po transport and diffusion model, data collection 210 The three-dimensional dynamic concentration distribution of Po was obtained. 210 Time series data of Po concentration; The feature extraction and dataset construction module is used to extract features based on the data. 210 15-dimensional enhanced spatiotemporal features were extracted from time-series data of Po concentration, and then a dataset was constructed and randomly divided into training set, validation set and test set. The model building and training module is used to build a physical information neural network model that integrates physical constraints and multi-task learning mechanisms, and to train and optimize the physical information neural network model using the training set and validation set. The inversion module is used to input the test set into the trained and optimized physical information neural network model and output... 210 Po leakage inversion results.

[0092] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A lead-bismuth cooled microreactor 210 The Po multi-task source item fast inversion method is characterized by... It includes the following steps: S1. Obtain the target leakage parameters under the target reactor leakage condition; S2. Establish corresponding conditions for the target leakage scenario. 210 The Po transport-diffusion model inputs the target leakage parameters to... 210 Po transport and diffusion model, data collection 210 The three-dimensional dynamic concentration distribution of Po was obtained. 210 Time series data of Po concentration; S3. Based on the above 210 A 15-dimensional enhanced spatiotemporal feature was extracted from the time-series data of Po concentration, and then a dataset was constructed and randomly divided into a training set, a validation set, and a test set. The 15-dimensional enhanced spatiotemporal feature includes peak activity, cumulative field energy, local high intensity ratio, concentration-weighted spatial centroid coordinates, centroid offset distance, global spatial dispersion, diffusion anisotropy ratio, concentration temporal gradient, spatial Laplacian operator, normalized relative curvature, global average spatial gradient, concentration distribution skewness and kurtosis, and structural information entropy. S4. Construct a physical information neural network model that integrates physical constraints and multi-task learning mechanisms, and train and optimize the physical information neural network model using the training set and validation set; the physical information neural network model includes an input layer, a shared feature extraction layer, a leakage condition classification branch, and a leakage height regression branch; the shared feature extraction layer is used to extract high-order nonlinear spatiotemporal information from the 15-dimensional enhanced spatiotemporal features of the input; the leakage condition classification branch is connected after the shared feature extraction layer, including two fully connected layers and a classification output layer, used to identify the output leakage condition type; the leakage height regression branch is connected after the shared feature extraction layer, including two fully connected layers and a linear output layer, used to output the height of the leakage breach location from the bottom of the main container; S5. Input the test set into the trained and optimized physical information neural network model, and output... 210 Po leakage inversion results.

2. The lead-bismuth cooled microreactor according to claim 1 210 The Po multi-task source item fast inversion method is characterized by... In step S1, the target leakage condition includes: In the case of leakage at the bottom center of the main vessel, the rupture point is located at the bottom center of the main vessel, and the rupture point height is... ; In the case of leakage at the lower sidewall of the main vessel, the rupture point is located at the lower sidewall of the main vessel, and the rupture point height is... satisfy: , This represents the initial liquid level of the lead-bismuth coolant in the main container; In the case of leakage on the upper sidewall of the main vessel, the rupture point is located on the upper sidewall of the main vessel, and the height of the rupture point is... satisfy: ; In the case of leakage on the upper sidewall of the main vessel, the rupture point is located on the upper sidewall of the main vessel, and the height of the rupture point is... satisfy: .

3. The lead-bismuth cooled microreactor according to claim 1 210 The Po multi-task source item fast inversion method is characterized by... In step S1, the target leakage parameters include: (1) Leakage velocity : In the formula: Represents the driving pressure difference. , Represents the pressure of the covered gas. Represents the density of lead-bismuth coolant; h 0 represents the initial liquid level of lead-bismuth coolant in the main container; h b This represents the height of the leak rupture location from the bottom of the main container; Represents the frictional resistance coefficient along the friction path. , Re Represents the Reynolds number, , The dynamic viscosity of lead-bismuth coolant; Represents the inlet local drag coefficient; d This represents the diameter of the leak point rupture. L This represents the length of the rupture channel at the leakage point; (2) Mass flow rate : In the formula, The cross-sectional area of ​​the breach representing the leak point; (3) 210 Po evaporation release boundary area : In the formula, represent t The cumulative leakage volume of lead-bismuth coolant at any given time. , This represents the thickness of the lead-bismuth coolant spread.

4. The lead-bismuth cooled microreactor according to claim 1 210 The Po multi-task source item fast inversion method is characterized by... In step S2, the 210 The Po transport and diffusion model was constructed using multiphysics simulation software, employing asymmetric ventilation conditions. Based on flow field calculations, it was described using the rare matter transport equation. 210 The diffusion behavior of Po; a mass reduction term based on the effective lifetime τ is introduced into the rarefied mass transport equation. R : In the formula: C represent 210 Po concentration; τ represents... 210 Po's effective lifespan in air.

5. The lead-bismuth cooled microreactor according to claim 1 210 The Po multi-task source item fast inversion method is characterized by... In step S2, by the 210 The Po transport and diffusion model includes N discrete monitoring points with three-dimensional coordinates, and a monitoring sensor is installed at each monitoring point. i The coordinates of the monitoring sensors are ,1≤ i ≤N; After the target leakage condition occurs, each of the monitoring sensors collects data at sampling intervals of 0.1 to 1 second within a predetermined time period. 210 Po concentration was obtained from data collected by various monitoring sensors. 210 Po concentration time series data .

6. The lead-bismuth cooled microreactor according to claim 1 210 The Po multi-task source item fast inversion method is characterized by... In step S3, the 15-dimensional enhanced spatiotemporal features include: Spatiotemporal field strength characteristics, including peak activity Cumulative field energy and local high intensity ratio ;in: represent t Real-time collection 210 Po concentration; , Representing the i Data collected by monitoring sensors 210 Po concentration For the preset concentration threshold, Representative of collection 210 Po concentration exceeds preset concentration threshold The number of monitoring sensors; Spatial distribution dimensional characteristics, including concentration-weighted spatial centroid coordinates. Centroid offset distance L c = Global spatial discreteness and diffusion anisotropy ratio ;in: x c 、y c 、z c These represent the X, Y, and Z coordinates of the centroid of the concentration-weighted space, respectively, in meters. Representing the i Data collected by monitoring sensors 210 Po concentration, L c This represents the offset distance of the centroid of the concentration-weighted space relative to the center of the main container on the horizontal plane. , , These represent the concentration-weighted variances of the concentration field in the X, Y, and Z directions, respectively. , , ; Physical transport dimension features, including concentration-time gradient Spatial Laplace operator Normalized relative curvature K= and global average spatial gradient ;in, Represents a certain monitoring point 210 Changes in Po concentration over time; C Represents the spatial gradient, indicating the same time interval. 210 Changes in Po concentration with varying monitoring point locations; , , For the first i Concentration gradient components collected by a monitoring sensor; Deformation structural dimensional characteristics, including concentration distribution skewness kurtosis and structural information entropy ;in, r i Representing the i Spatial distance from each monitoring point to the centroid of the concentration-weighted space ; p i This represents the normalized probability distribution of concentrations at monitoring points. .

7. The lead-bismuth cooled microreactor according to claim 1 210 The Po multi-task source item fast inversion method is characterized by... Step S3 also includes: using the SMOTE synthetic minority class oversampling algorithm to generate synthetic samples for target leakage conditions with a small number of samples to eliminate class distribution imbalance; and using Z-score normalization to eliminate the dimensional differences of 15-dimensional enhanced spatiotemporal features.

8. The lead-bismuth cooled microreactor according to claim 6 210 The Po multi-task source item fast inversion method is characterized by... During the training process in step S4: consider 210 Migration behavior of Po in an asymmetric ventilation flow field 210 The concentration field of Po satisfies the following convection-diffusion control equation: In the formula: u Represents the velocity vector of the flow field; D Represents the equivalent diffusion coefficient; R(C) Represents quality reduction items; Represents the source of leakage. , per unit area of ​​lead-bismuth surface 210 Po evaporation release flux. ; The Dirac function represents the source term being located at the bottom center of the main container; This is an indicator function, representing that the release source exists only in the area covered by the liquid pool; This represents the area covered by the liquid pool at the bottom of the crater; Based on the height of the leak rupture location from the bottom of the main container predicted by the physical information neural network model, the driving pressure difference and leakage velocity are first calculated. 210 Po evaporation release boundary area, combined with 210 The evaporation flux of Po was then used to derive the theoretical equivalent source strength corresponding to the predicted height of the leak rupture location from the bottom of the main container. ; In the formula, per unit area of ​​lead-bismuth surface 210 Po evaporation release flux. The predicted location of the leak breach is at time t, corresponding to the distance from the bottom of the main container. 210 Po evaporation release boundary area; Based on the peak activity and cumulative field energy characteristics, the measured equivalent source strength, used to characterize the actual observation results of the monitoring data, is then calculated. ; in, , The inherent linear scaling factor of the crater flow field; with theoretical equivalent source strength Equivalent source strength as measured The difference constructs the physical residual - When the model prediction results satisfy the physical transport laws, the physical residuals Approaching zero; for Take the mean square error as the physical residual loss term. ; Construct the total loss function: In the formula, L class Cross-entropy loss for representative working conditions classification, , y i This is a label indicating a genuine leak of operating conditions. The probability of the leakage condition category predicted by the model; L reg This represents the high regression mean square error loss due to leakage. , h i This represents the actual location of the leak / rupture point from the bottom of the main container. The height of the predicted leak rupture location from the bottom of the main container; L phys Represents the physical residual loss term; The cross-entropy loss weight coefficients are used for classifying operating conditions. The weighting coefficients for the mean squared error loss during leakage height regression are... This is the weighting coefficient for the physical residual loss term.

9. A lead-bismuth cooled microreactor 210 The Po multi-task source item fast inversion system is characterized by... It is used to realize the lead-bismuth cooled microreactor as described in any one of claims 1-8. 210 The Po multi-task source item fast inversion method includes: The acquisition module is used to acquire target leakage parameters under target reactor leakage conditions. The simulation module is used to establish a corresponding simulation model for the target leakage condition. 210 The Po transport-diffusion model inputs the target leakage parameters to... 210 Po transport and diffusion model, data collection 210 The three-dimensional dynamic concentration distribution of Po was obtained. 210 Time series data of Po concentration; The feature extraction and dataset construction module is used to extract features based on the data. 210 15-dimensional enhanced spatiotemporal features were extracted from time-series data of Po concentration, and then a dataset was constructed and randomly divided into training set, validation set and test set. The model building and training module is used to build a physical information neural network model that integrates physical constraints and multi-task learning mechanisms, and to train and optimize the physical information neural network model using the training set and validation set. The inversion module is used to input the test set into the trained and optimized physical information neural network model and output... 210 Po leakage inversion results.

Images