Nuclear accident dynamic safety operation time assessment method and system based on data fusion

Abstract

The invention discloses a nuclear accident dynamic safe operation time assessment method and system based on data fusion, relates to the technical field of nuclear radiation protection, and solves the technical problems that the safe operation time is not accurate enough and dynamic dose assessment is difficult in nuclear accident emergency response. According to the technical scheme, nuclear accident prediction data and site real-time monitoring data are fused, a data assimilation fusion algorithm and an intelligent analysis model are combined, the radiation dose and environmental conditions of an accident site are comprehensively evaluated, and safety operation time and risk levels of different areas are dynamically calculated and updated; by considering the building shielding effect and supporting the federated learning framework, the cooperative evaluation of multiple monitoring points is realized, so that the operation time evaluation precision and decision-making efficiency in the nuclear accident emergency response are greatly improved. Finally, operation suggestions can be fed back and updated in real time, and the real-time performance and accuracy of emergency command decisions are ensured.

Description

Technical Field

[0001] This application relates to the field of nuclear radiation protection technology, and in particular to a method and system for assessing dynamic safe working time in nuclear accidents based on data fusion. Background Technology

[0002] Following a nuclear accident, rapidly assessing the impact of safe working hours and radiation dose on personnel health in the affected area is a critical task in nuclear accident emergency response. Current nuclear accident consequence assessment and decision support systems primarily rely on static physical model chains, generating radiation diffusion predictions and dose distributions based on parameters such as initial accident source intensity and meteorological conditions. However, because radiation diffusion and dose levels at the accident site are significantly influenced by the dynamic environment, the real-time adaptability and accuracy of existing fixed-parameter model chains are limited, making it difficult to provide emergency command with real-time updated information on working hours and dose assessments.

[0003] The shortcomings of existing nuclear accident consequence assessment and decision support systems are: (1) Insufficient real-time performance and dynamic adaptability: Radiation levels and weather conditions at nuclear accident sites can change at any time, and traditional fixed-parameter model chains struggle to adapt to dynamic feedback from real-time data, resulting in poor timeliness of operational time assessments. In environments with rapidly changing wind speed and direction, radiation diffusion can fluctuate, rendering fixed predictions inapplicable. Furthermore, the system cannot adjust operational times in real time when building shielding effectiveness or environmental conditions change, causing operational recommendations to lag behind.

[0004] (2) Single-factor judgment and work safety analysis: Existing nuclear accident consequence assessment systems often rely on single dose standards or thresholds to determine work time, failing to adequately consider complex factors such as building shielding effectiveness and dynamic changes in the accident source. Analysis methods based on single dose thresholds are ill-suited to reflecting the actual risk levels in different areas. For instance, in areas with buildings or shielding facilities, radiation doses are significantly shielded; if the system does not consider shielding factors, work time assessments will be inaccurate, impacting actual personnel safety.

[0005] (3) Insufficient ability to effectively integrate monitoring data and forecast results: While some existing nuclear accident consequence assessment systems receive on-site monitoring data, they lack the ability to deeply integrate monitoring data with predictions using intelligent algorithms. They often process the data through simple overlay, lacking data assimilation and intelligent analysis mechanisms. For example, if monitoring data indicates a rapid increase in radiation dose in a region, and the system fails to quickly integrate the monitoring data to correct initial predictions, it may lead to delays in operational time assessments and prevent timely adjustments to operational recommendations.

[0006] (4) Lack of dynamic partitioning adjustment mechanism supported by intelligent algorithms: In nuclear accident emergency response, recommendations for protective zones and work schedules often need to be dynamically adjusted based on real-time environmental conditions. However, traditional nuclear accident consequence assessment systems largely rely on static zoning schemes and lack automated zoning adjustments. For example, as the accident progresses or on-site radiation levels change, the system fails to automatically adjust the safe work schedules for each zone based on new data, affecting the scientific validity of the zoning and the timeliness of emergency command. The lack of intelligent algorithm support also makes it difficult for the system to provide timely protection recommendations in rapidly changing accident scenarios.

[0007] (5) Lack of intelligent decision support: Traditional nuclear accident consequence assessment systems primarily rely on static analysis and simple weighting methods, failing to fully utilize intelligent algorithms (such as deep learning or machine learning) for intelligent analysis and real-time adjustment of multi-source data. Given the complex and ever-changing data sources at emergency response sites, intelligent algorithms offer significant advantages in emergency scenarios, but they have not been fully developed in existing systems. This results in insufficient accuracy in operational time assessment and dosage control, making it difficult to adapt to dynamically changing emergency needs.

[0008] In summary, traditional nuclear accident consequence assessment systems have significant limitations in areas such as real-time dynamic data fusion, intelligent analysis of operation time calculation, and comprehensive multi-factor assessment, making it difficult to meet the needs of modern nuclear accident emergency response for precise safe operation time and dynamic dose assessment. Summary of the Invention

[0009] This application provides a data fusion-based method and system for assessing the dynamic safe working time of nuclear accidents. Its technical purpose is to provide accurate assessment results for the safe working time and radiation dose in different areas of a nuclear accident site, and to provide scientific support for protection zoning and dose control in nuclear accident emergency response.

[0010] The above-mentioned technical objective of this application is achieved through the following technical solution: A method for assessing dynamic safe working time in nuclear accidents based on data fusion includes: Data acquisition phase: Acquire nuclear accident consequence assessment prediction data and other data within the nuclear accident assessment area, and acquire monitoring data at the nuclear accident site; wherein, the prediction data includes continuously changing radiation diffusion trends and dose distribution maps, the monitoring data includes historical and current radiation dose monitoring data and meteorological observation data; the other data includes population distribution at different times and shielding factors of various buildings; Data assimilation and fusion stage: Assimilation and fusion calculations are performed on the predicted data and monitoring data to generate continuously changing first fused predicted data; Safe operation time assessment phase: Based on the first fusion prediction data and other data, the preliminary safe operation time of different zones with dynamic changes is analyzed and calculated; Feedback and dynamic adjustment phase: Abnormal data in the current monitoring data is captured and corrected. Then, the corrected abnormal data is fused with the first fusion prediction data to generate the second fusion prediction data. The preliminary safe operation time is updated in real time based on the second fusion prediction data to obtain the dynamically updated safe operation time. Risk level classification phase: Based on the second fusion prediction data and dynamically updated safe operation time, the operation zones and risk levels are classified.

[0011] Preferably, the assimilation and fusion calculation of the predicted data and real-time data includes: The spatiotemporal field corresponding to the predicted data Observational constraint field constructed from sparse monitoring data In the spatial domain Deformation matching is performed internally, and then the non-equilibrium extension of the Benamou–Brenier dynamic formula is solved to obtain the deformation trajectory with minimum "transport work". , and final state , making the spacetime field Under the constraint of non-conservation of mass, it is continuously deformed into the observation constraint field. The uniform nearest neighbors are represented as: ; ; in, ; Represents the deformation velocity field; , indicating spatial location; This represents the dose field corresponding to the predicted data; This represents the field obtained by interpolating monitoring data, used to represent the consistency target of observations; This represents the density or field value during deformation; This represents non-conservative source and sink terms to characterize sedimentation and decay effects; For the weighting factor; The constraint function representing consistency with observations. , Indicates the first Location of each monitoring point For quality weights; , representing the consistency tolerance threshold; let the model field after deformation alignment be . ; with model field For low-fidelity references, monitoring data is high-fidelity information; for residuals... Modeling was performed, and the Co-Kriging method was used to analyze the residuals. Recursively incorporate heteroscedastic noise, as follows: ; in, , This indicates the high-fidelity field corresponding to the monitoring data; Representing a Gaussian process , Represents the kernel function; This represents the heteroscedasticity term related to location and time. Indicates the cross-fidelity scale coefficient; according to Obtain the residual posterior mean With variance Thus, the statistical correction field is obtained, expressed as: ; in, This represents the high-fidelity field of the monitoring data; ; Residual The posterior mean of the Gaussian process, and Residual The posterior variance of the Gaussian process; This indicates the correction of the low-fidelity model field, and ; Statistical correction field For the data term, a physical residual constraint is introduced from the advection-diffusion-decay equation, expressed as: ; in, Indicates the field to be merged; This represents the partial derivative with respect to time; Represents the advection velocity field; Indicates the anisotropic diffusion coefficient; Indicates radioactive decay; Indicates the source term; The joint loss is constructed based on the physical residual constraints and is expressed as follows: ; in, Represents the weighting coefficient of the physical residual term; Indicates the weights of the boundary constraint terms; This represents a measure of the deviation between the boundary conditions and the initial conditions. The neural network is trained using a joint loss function to obtain a fusion solution that satisfies the consistency between the physical equations and the observations. , That is, the first fused prediction data; among which, To achieve the optimal network output field with the best parameters, To achieve a minimum while simultaneously satisfying both data consistency and physical consistency; In scenarios involving sensor anomalies or monitoring distribution drift, the Wasserstein sphere is used. For an uncertain set, perform sub-Bruker optimization to suppress the impact of extreme monitoring on the first fused prediction data, expressed as: ; Represent the hypothesis space of feasible functions; This represents the distribution of candidate data, taken from an empirical distribution. Centered on, with radius Wasserstein's ball; Relative to the distribution The expectation operator; Representation and observation constraint field The loss due to mismatch; An empirical distribution composed of historical observation samples; Used to control the robust radius against distribution drift; Preferably, the analysis and calculation of the preliminary safe operation time for dynamically changing different partitions based on the first fused prediction data and other data includes: Dose thresholds are determined according to radiation protection standards; The environmental dose rate at each location is predicted by combining an atmospheric diffusion model with current meteorological observation data. By introducing a building shielding factor to correct for the environmental dose rate, the effective dose rate is obtained, expressed as: ; in, Indicates the effective dose rate. Indicates the environmental dose rate. Indicates the shielding attenuation coefficient; The initial safe operating time, derived from the effective dose rate, is expressed as follows: ; in, Indicates the initial safe operation time. This indicates the dose limit for this mission. This indicates the cumulative dosage.

[0012] Preferably, ;in, Represents a linear decay function. This indicates the thickness of the shielding material.

[0013] Preferably, the process of generating the second fused prediction data includes: The time series corresponding to multi-point monitoring are stacked into a time data matrix according to time sliding windows. For time data matrix Robust principal component decomposition is performed to separate the stationary background. With sparsity anomaly Then we get the denoised background field. and anomaly mask , is represented as: ; in, Represents the nuclear norm. express Norm; This represents the sparse regularization weight; The dose rate sequence corresponding to the radiation dose monitoring data at each monitoring point was obtained using the online Bayesian change point detection method. The current runtime is calculated posteriorly to locate mutation points, represented as: ; in, This indicates the runtime, i.e., from the last mutation point to the current point. The length of the segment; This represents the posterior probability of the run length given all current observations. The prior transition probability represents the length of the operation; This indicates the predicted likelihood based on historical samples within the current segment; This represents the posterior length of the previous run; when Exceeding the threshold If a point is reached, that point is marked as a mutation point, and the data corresponding to that point is considered abnormal data. Construct a sensor map with monitoring stations as nodes and meteorological correlation, geographical proximity, and wind direction coupling as edges. For the background field Perform graph attention propagation to obtain spatially consistent dose field estimates. , is represented as: ; ; ; in, This represents the set of nodes, i.e., the set of monitoring points; Represents the set of edges; Represents attention weights and can be used for anomaly masks. Punishment is imposed to suppress the effects of contaminated sensors; Represents a node The updated vector representation; This represents the transpose of the attention parameter vector; Represents a node The set of neighboring nodes any node in ; Indicates a node original features The result of applying a linear transformation, Represents the learnable weight matrix; Represents a nonlinear activation function; Representing neighboring nodes Features after linear transformation; This indicates that the node embedding is mapped to a location. The field value; For dose field estimation Threshold exceeding probability field Calculations are performed to cross the threshold probability field. The probabilistic boundaries that form high-risk zones, short-operation zones, and safe zones are represented as: ; in, Indicates the partition threshold; Indicates the significance tolerance, with a confidence level of 1. ; Based on the location of abrupt change points, an LSTM network is used to learn and provide feedback on the temporal changes in radiation dose monitoring data. This allows for the automatic identification and correction of abrupt change points and abnormal data, resulting in corrected radiation dose monitoring data, represented as follows: ; ; in, This represents the hidden state of the LSTM network. This represents the input data at the current time step. This represents historical radiation dose monitoring data. This indicates the radiation dose monitoring data that has been corrected and fed back through the LSTM network; The weight matrix represents the recursion. Represents the bias vector; Corrected radiation dose monitoring data The probabilistic boundaries of each partition and the first fusion prediction data are fused to generate the second fusion prediction data.

[0014] Preferably, the step of updating the preliminary safe operation time in real time based on the second fusion prediction data to obtain a dynamically updated safe operation time includes: In the prediction time domain The above minimizes the multi-objective cost "time-dose-congestion," and applies probabilistic and tail risk constraints to obtain the dynamically updated safe operating time, expressed as: ; ; in, This indicates the dose increment along the recommended course of action. ; Indicates the cost of time. This indicates a penalty for localized congestion. , , All represent weighting coefficients; Represents the discrete time step; Indicates the current time; This represents a probability operator that constrains the probability of an event occurring to not exceed a given tolerance. This indicates that "single-step dose exceeds the threshold". The probability of "is no greater than" , Indicates the dose threshold; Represents the set of feasible strategies; This indicates the risk tolerance under opportunity constraints; Indicates tail risk constraint The upper limit; Represents the optimization variable; This represents the tail risk constraint; the probability constraint samples weather and source term uncertainties using the scenario method and provides a feasibility confidence guarantee.

[0015] Preferably, the step of dividing the work areas and risk levels based on the second fusion prediction data and dynamically updated safe work time includes: The risk level of each zone is calculated and expressed as follows: ; in, Indicates the area The risk level, Indicates the effective dose rate. Indicates the area Dynamically updated safe operating time This represents a risk assessment function that combines effective dose rate and safe operating time. The minimum dose propagation path to the target area is calculated using the fast travel method and the Eikonal approximation equation, and is expressed as follows: ; in, This represents the gradient of the shortest path to the target region. This represents the minimum dose propagation index reaching the target area. Represents the local optical length weight. Indicates in the target set Boundary conditions on, i.e. ; By minimizing dose exposure in the minimum dose propagation path using the HJ reachability model and considering obstacle avoidance, the optimal path from the starting point to the target area is obtained, expressed as: ; in, Indicates at time At that time, from the starting point To the target area The minimum dose path value function; Represents the velocity field along the path; This represents the dose at each point in the path; This represents the transpose of the gradient of the value function; The dose field is calibrated to the upper quantile using distribution-independent conservative uncertainty calibration, resulting in conservative upper quantile dose values, expressed as: ; in, This represents the conservative upper quantile dose value. Representing a path Dosage increments, This indicates the set dose limit. Indicates risk tolerance; Based on the risk levels of each zone, the work zones are divided according to the optimal path and the upper quantile dose value. Then, unstable small-scale structures in the work zone boundaries are removed by topological persistence techniques, and transient structures caused by noise at the work zone boundaries are screened out at multiple scales using persistent cohomology methods. The final zone boundaries of the work zones are obtained, represented as: ; in, This indicates the final partition boundary of the job partition. Indicates the persistence threshold; This represents the set of topological features obtained through persistent cohomology and that persists throughout the multi-scale filtering process. This indicates that the time to survival in a persistent barcode exceeds [a certain value]. The features are preserved, while the rest are treated as noise and removed.

[0016] A data fusion-based dynamic safety operation time assessment system for nuclear accidents, the system being used in the assessment method, the system comprising: The data acquisition module acquires nuclear accident consequence assessment prediction data and other data within the nuclear accident assessment area, and acquires monitoring data at the nuclear accident site; wherein, the prediction data includes continuously changing radiation diffusion trends and dose distribution maps, the monitoring data includes historical and current radiation dose monitoring data and meteorological observation data; the other data includes population distribution at different times and shielding factors of various buildings; The data assimilation and fusion module performs assimilation and fusion calculations on the predicted data and the monitoring data to generate continuously changing first fused predicted data. The safe operation time assessment module analyzes and calculates the preliminary safe operation time for different dynamically changing zones based on the first fused prediction data and other data. The feedback and dynamic adjustment module captures and corrects abnormal data in the current monitoring data, then merges the corrected abnormal data with the first fusion prediction data to generate the second fusion prediction data, and updates the preliminary safe operation time in real time based on the second fusion prediction data to obtain the dynamically updated safe operation time. The risk level classification module classifies work zones and risk levels based on the second fusion prediction data and dynamically updated safe operation time.

[0017] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements a data fusion-based dynamic safety operation time assessment method for nuclear accidents.

[0018] A computer storage medium storing a computer program that, when executed by a processor, implements a data fusion-based dynamic safety operation time assessment method for nuclear accidents.

[0019] The beneficial effects of this application are as follows: (1) Significantly improves real-time performance and accuracy By employing data assimilation and fusion technology, predicted data is integrated with on-site monitoring data in real time. The model can dynamically update the operation time assessment and dose distribution, making the operation time prediction more closely reflect the actual situation at the accident site. Compared with traditional static prediction, this application can dynamically adjust the results in complex accident environments, providing accurate and real-time operation time assessments for emergency response.

[0020] (2) Multifactor comprehensive analysis This application, in its assessment of operational time, not only considers real-time radiation dose but also comprehensively analyzes various factors such as weather changes, building shielding effectiveness, and the intensity of accident sources, thereby ensuring the comprehensiveness and scientific rigor of the assessment results. Multi-factor analysis enhances adaptability to different scenarios, making the operational recommendations more practical.

[0021] (3) Intelligent protection zoning and real-time adjustment This application utilizes intelligent algorithms to dynamically partition different areas based on risk levels and safe operating times. After generating initial partitions, the model automatically adjusts the partitioning results based on monitoring data, ensuring that the partitioning recommendations for different areas remain scientific and accurate at all times. This intelligent partitioning adjustment significantly enhances the scientific rigor and decision support capabilities in accident emergency response.

[0022] (4) Dynamic feedback mechanism and security enhancement The model continuously adjusts its operational time assessments and protection zone recommendations through a dynamic feedback mechanism based on monitoring data, ensuring its high adaptability to dynamic changes in accidents. By dynamically adjusting protection recommendations based on real-time monitoring data, the model provides scientific decision support to emergency command centers, significantly improving public safety and emergency response efficiency.

[0023] In summary, the intelligent dynamic safety operation time assessment method and system based on the fusion of nuclear accident prediction data and monitoring data described in this application achieves accurate assessment of operation time and protection zones at accident sites through dynamic fusion of multi-source data and intelligent zoning judgment, and has great technical advantages and broad application prospects. Attached Figure Description

[0024] Figure 1 This is a flowchart of the nuclear accident dynamic safe operation time assessment method based on data fusion in the embodiments of this application; Figure 2 This is a schematic diagram comparing the radiation curve attenuation under different shielding conditions in the embodiments of this application; Figure 3 The shielding attenuation coefficient in the embodiments of this application An estimation diagram; Figure 4 This is a schematic diagram of the federated closed loop of data assimilation and fusion in the embodiments of this application; Figure 5 This is a schematic diagram illustrating the abnormal prediction of dose timing by the LSTM network in the embodiments of this application; Figure 6 This is a schematic diagram of the training loss of the LSTM network in the embodiments of this application; Figure 7(a) is a schematic diagram of the predicted data for the diffusion of radioactive contaminant concentrations in the accident, and (b) is a schematic diagram of the second fusion predicted data for the diffusion of radioactive contaminant concentrations in the accident. Figure 8 (a) is a schematic diagram of the estimated safe working time after the accident is released a hours later, and (b) is a schematic diagram of the estimated safe working time after the accident is released b hours later. Detailed Implementation

[0025] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, the data fusion-based dynamic safe operation time assessment method for nuclear accidents described in this application includes: 100 Data Acquisition Phase: Acquire nuclear accident consequence assessment prediction data and other data within the nuclear accident assessment area, and acquire monitoring data at the nuclear accident site; wherein, the prediction data includes continuously changing radiation diffusion trends and dose distribution maps, the monitoring data includes historical and current radiation dose monitoring data and meteorological observation data; the other data includes population distribution at different times and shielding factors of various buildings.

[0027] Preferably, the predicted data is generally provided by an existing nuclear accident consequences assessment system (such as JRODOS or other nuclear accident consequences and decision support systems).

[0028] 101 Data Assimilation and Fusion Stage: Assimilation and fusion calculations are performed on the predicted data and monitoring data to generate continuously changing first fused predicted data.

[0029] Preferably, the assimilation and fusion calculation of the predicted data and real-time data includes: 1011: Field deformation alignment based on non-equilibrium dynamic optimal transport (UOT), i.e., the spatiotemporal field corresponding to the predicted data. Observational constraint field constructed from sparse monitoring data In the spatial domain Deformation matching is performed internally, and then the non-equilibrium extension of the Benamou–Brenier dynamic formula is solved to obtain the deformation trajectory with minimum "transport work". , and final state , making the spacetime field Under constraints satisfying non-conservation of mass (settlement, decay, etc.), it is continuously deformed into the observation constraint field. The uniform nearest neighbors are represented as: ; ; in, ; Represents the deformation velocity field; , indicating spatial location; This represents the dose field corresponding to the predicted data; This represents the field obtained by interpolating monitoring data, used to represent the consistency target of observations; This represents the density or field value during deformation; This represents non-conservative source and sink terms to characterize sedimentation and decay effects; For the weighting factor; Constraint functions representing consistency with observations (such as weighted residuals or sum of squared residuals at monitoring points). , Indicates the first Location of each monitoring point For quality weights; , representing the consistency tolerance threshold, limits the upper bound of consistency between the alignment result and the observation, in units of and . Maintain consistency. Let the model field after deformation alignment be denoted as . .

[0030] 1012: Using model fields For low-fidelity references, monitoring data is high-fidelity information; for residuals... Modeling was performed, and the Gaussian process Co-Kriging method was used to analyze the residuals. Recursively incorporate heteroscedastic noise, as follows: ; in, , This indicates the high-fidelity field corresponding to the monitoring data. Representing a Gaussian process ; This represents the kernel function, used to characterize spatial-temporal dependencies; This represents the heteroscedasticity term related to position and time, reflecting differences in sensor quality and operating conditions. This represents the cross-fidelity scaling coefficient, used to correct the multiplicative bias of low-fidelity fields relative to high-fidelity observations; according to Obtain the residual posterior mean With variance Thus, the statistical correction field is obtained, expressed as: ; in, This represents the high-fidelity field of the monitoring data; This indicates a relatively low-fidelity residual. ; Residual The posterior mean of the Gaussian process, and Residual The posterior variance of the Gaussian process; This application represents the correction of the low-fidelity model field. and Equivalent, used to refer to a low-fidelity reference after UOT alignment.

[0031] 1013: Uniform correction of partial differential equations (PDEs) based on Physically Constrained Neural Networks (PINNs), i.e., using statistical correction fields. For the data term, a physical residual constraint is introduced from the advection-diffusion-decay equation, expressed as: ; in, Indicates the field to be merged; This represents the partial derivative with respect to time; This represents the advection velocity field, given by the meteorological wind field, in units of... ; Represents the anisotropic diffusion coefficient, in units of ; Represents radioactive decay, unit ; The value represents the source term; a positive value indicates a release source, and a negative value indicates sedimentation.

[0032] The joint loss is constructed based on the physical residual constraints and is expressed as follows: ; in, The weighting coefficients for the physical residuals are used to weigh “data fit” against “physical consistency”. Indicates the weights of the boundary constraint terms; This represents a measure of the deviation between the boundary conditions and the initial conditions, for example, on the set of sampling points at the boundary and the initial time. Weighted sum of norms.

[0033] 1014: Train the neural network using joint loss to obtain a fusion solution that satisfies the consistency between the physical equations and observations. , That is, the first fused prediction data; among which, To achieve the optimal network output field with the best parameters, Achieve a minimum while simultaneously satisfying both data consistency and physical consistency.

[0034] Preferably, in scenarios involving sensor malfunction or monitoring distribution drift, a Wasserstein sphere is used. For an uncertain set, perform sub-Bruker optimization to suppress the impact of extreme monitoring on the first fused prediction data, expressed as: ; in, Represents the space of feasible function hypotheses, such as the set of fused fields parameterized by a neural network; This represents the distribution of candidate data, taken from an empirical distribution. Centered on, with radius Wasserstein's ball; Relative to the distribution The expectation operator; Representation and observation constraint field Mismatch losses, such as the weighted squared error at the location of the monitoring point; An empirical distribution composed of historical observation samples; Used to control the robust radius against distribution drift. The resulting... Substitute it back into the physical constraint neural network method as an initialization or regularization term to improve the stability and confidence interval estimation capability of the fusion under low availability observations and outliers.

[0035] 102 Safe Operation Time Assessment Phase: Based on the first fusion prediction data and other data, the preliminary safe operation time of different dynamically changing zones is analyzed and calculated.

[0036] Preferably, the analysis and calculation of the preliminary safe operation time for dynamically changing different partitions based on the first fused prediction data and other data includes: 1021: Determine the dose threshold according to radiation protection standards. Preferably, the dose threshold for a single mission should not exceed 50 mSv under normal emergency conditions.

[0037] 1022: Predict the environmental dose rate at each location by combining an atmospheric diffusion model with current meteorological observation data (wind direction, wind speed, precipitation, etc.).

[0038] 1023: Introducing a building shielding factor to correct the environmental dose rate yields the effective dose rate, expressed as: ; in, This represents the effective dose rate (real-time dose rate after shielding and operating condition correction). Indicates the environmental dose rate. This represents the shielding attenuation coefficient.

[0039] If there is shielding at a certain location, the effective dose rate will be calculated based on the shielding attenuation factor. Reduce the shielding attenuation coefficient It can be determined based on building materials and thickness using empirical formulas or half-value layer data, expressed as: ; in, Represents a linear decay function (cm) -1 ), This indicates the thickness of the shielding material (cm).

[0040] Exponential masking attenuation is expressed as: ; in, This indicates the radiation intensity after shielding. This represents the incident radiation intensity. This formula is used to quantitatively describe the effect of materials on... The exponential decay law of radiation intensity (half-value layer, tenth-value layer, etc. can be derived from...) (Pushed).

[0041] 1024: The initial safe operating time is obtained based on the effective dose rate, expressed as: ; in, Indicates the initial safe operation time. This indicates the dose limit for this mission. This indicates the cumulative dosage.

[0042] Specifically, assuming the single-mission dose limit for emergency personnel is 25 Rem (≈0.25 Sv), and the highest handle dose rate measured at a certain work site is approximately 10 Rem / h, then the safe stay time in that area is approximately 25 / 10 = 2.5 hours. It should be noted that weather changes can lead to... Dynamic fluctuations can be updated through rolling forecasts and real-time monitoring, and are calculated... The worst-case scenario should be considered (peak dose rate or a certain safety margin).

[0043] Preferably, population distribution factors are used to assess collective risk, and operational strategies for different areas can be optimized by combining spatial density and population flow heat maps: in densely populated areas, the restrictions can be appropriately tightened. or shorten To reduce the overall dose to the population, and to prioritize necessary long-duration operations in sparsely populated areas. Through joint modeling of these multiple factors, this approach can output preliminary safe operating time recommendations for each area, enabling dynamic and safe management of complex emergency situations.

[0044] Therefore, this application can automatically calculate the shielding effect of a building on radiation dose for specific areas, taking into account the building structure, materials, and shielding factor. This allows for adjustments to the safe working hours for personnel within that area. For example, in areas with protective barriers or thick building structures, the safe working hours can be extended, providing more precise protection recommendations for personnel. Specifically, the dynamic calculation of the building shielding effect is based on the fact that the shielding effect of the building structure on radiation follows an exponential decay law, which can be expressed using a formula. Quantization representation. For example, for an energy of approximately 1 MeV... For X-rays, the half-value layer thickness of concrete is approximately 6.6 cm, steel is approximately 2.1 cm, and lead is approximately 1.2 cm. This means that, for the same thickness, lead has the strongest shielding effect; a lead plate of just a few centimeters can reduce strength by several orders of magnitude, while concrete requires a much greater thickness to achieve the same attenuation (e.g., ...). Figure 2 As shown, the plumb line decreases the fastest, while the concrete line decreases the slowest.

[0045] In practical applications, the shielding effect can be accurately assessed by simulating the transmission of rays through complex structures using the Monte Carlo method. For example, programs like MCNP or PHITS can be used to track rays particle by particle based on input parameters such as the density, composition, thickness, and geometry of the building materials. The shielding factor of a building is determined by the penetration and scattering of photons or neutrons. Studies show that the shielding factor of radiation with different energy spectra in buildings can change significantly over time (e.g., due to radioactive deposition, decay altering the environment). (Energy spectrum distribution), therefore, pre-calculation of building shielding is required for multiple representative energies. Dickson et al. used 16 monoenergetic energy distributions in the range of 0.1–3 MeV. X-ray Monte Carlo simulations were used to derive a general formula for building shielding factors that varies with energy, and the consistency with direct Monte Carlo calculations was verified. This demonstrates that dynamic shielding calculations require real-time adjustments based on nuclide type and building structure: during emergency response, on-site building structure material and dimensional parameters (based on engineering drawings or on-site surveys) are obtained; appropriate attenuation models or simulation methods are selected to calculate the shielding effect of different buildings and orientations; and the shielding factor is updated as the accident progresses (changes in source terms and radiation spectrum) to provide operators with more accurate dose protection assessment data.

[0046] Shielding parameters can also be intelligently initialized via remote sensing: key parameters of the building (including material type, wall thickness, floor height, orientation, etc.) are extracted by inverting high-resolution remote sensing data (such as high-resolution satellite imagery, synthetic aperture radar SAR, and oblique photography) to generate formulas. Medium linear attenuation coefficient and the thickness of the shielding material The prior distribution. Then, combined with the measured dose data obtained by the mobile monitoring vehicle inside and outside the building, a Bayesian method is used to analyze... and The system is updated to obtain a posterior estimate, and the building shielding attenuation coefficient is output accordingly. (i.e., the building's shielding factor) and its confidence interval. This intelligent initialization method enables the initialization of the building even in the absence of structural blueprints. To make an effective estimate, such as Figure 3As shown, this allows for the automatic acquisition of the shielding factor of each building during emergency response, thereby improving the adaptability of the method of this invention to complex urban environments.

[0047] Preferably, the data assimilation and fusion stage and the safe operation time assessment stage can adopt a federated assimilation-assessment integrated method to achieve their functions: under the premise that the data does not leave the domain (each monitoring station retains its local data), a federated learning framework (such as the classic algorithm FedAvg) is used to combine data assimilation (such as ensemble Kalman filter EnKF or four-dimensional variational 4D-Var) with dynamic safe operation time assessment. Each monitoring station performs dose data assimilation processing and masking correction locally, extracts the updated summary of the model in this application (such as parameter gradients or updated weights), and uploads it to the central server; the server periodically aggregates the model summaries of each station and updates the global model parameters, and then distributes the updated model to each station, thereby forming a federated closed loop, such as... Figure 4 As shown. This approach achieves unified operational time assessment across multiple monitoring points while ensuring the privacy and security of data from each site. The federalized assimilation-assessment scheme fully utilizes information from each site without directly exchanging raw monitoring data, enhancing the generalization ability of the proposed method in different regional environments and strengthening collaboration during emergency response.

[0048] 103 Feedback and Dynamic Adjustment Phase: Abnormal data in the current monitoring data is captured and corrected. Then, the corrected abnormal data is fused with the first fusion prediction data to generate the second fusion prediction data. The preliminary safe operation time is updated in real time based on the second fusion prediction data to obtain the dynamically updated safe operation time.

[0049] Preferably, the process of generating the second fused prediction data includes: 1031: Stack the time series corresponding to multiple monitoring points into a time data matrix by time sliding window. For time data matrix Robust principal component decomposition is performed to separate the stationary background. With sparsity anomaly Then we get the denoised background field. and anomaly mask , is represented as: ; in, i.e., time Sliding window data matrix (sensor × time step); stable background Indicates a low-rank component; Represents the nuclear norm. express Norm; This represents the sparse regularization weight, used to balance the decomposition ratio between low-rank background and sparse anomalies. The larger the size, the more likely it is to absorb the mutation. .

[0050] To meet emergency practicality requirements, incremental updates using online robust PCA with single samples or small batches can be performed to maintain linear complexity in terms of memory and computation, thereby obtaining the denoised background field. and anomaly mask .

[0051] 1032: Dose rate sequences corresponding to radiation dose monitoring data at each monitoring point using an online Bayesian change point detection method. The current running length (i.e., the observation sequence from time 1 to time t) is calculated posteriorly to locate the abrupt change point, and is represented as: ; in, This indicates the runtime, i.e., from the last mutation point to the current point. The length of the segment; This represents the posterior probability of the run length given all current observations, used to determine "whether a turning point occurs and when it occurs"; The prior transition probability (defined by hazard / arrival rate) represents the run length; if the point remains unchanged, then... If the point changes ; It represents the predicted likelihood based on historical samples within the current segment, reflecting whether the new observations are consistent with the statistical patterns within the segment; This represents the posterior length of the previous step, used for recursion; when Exceeding the threshold If a point is reached, that point is marked as a mutation point, and the data corresponding to that point is considered anomalous data, which is used to trigger subsequent risk reassessment and strategy replanning.

[0052] The above results of locating the mutation point directly triggered the following correction process: (1) assigning a lower weight to the sample at that moment or increasing the observation noise in robust decomposition / statistical modeling; (2) reducing the attention weight of the abnormal node in graph attention propagation. (3) Switch / shorten the training window and quickly re-evaluate in LSTM feedback and PINNs consistency, so as to complete the automatic correction of anomalies and the instant update of subsequent fusion.

[0053] 1033: Construct a sensor graph with monitoring stations as nodes and meteorological correlation, geographical proximity, and wind direction coupling as edges. For the background field Perform graph attention propagation to obtain spatially consistent dose field estimates. , is represented as: ; ; ; in, This represents the set of nodes, i.e., the set of monitoring points; Represents a set of edges (such as correlations or proximity relationships); Represents attention weights and can be used for anomaly masks. Punishment is imposed to suppress the effects of contaminated sensors; Represents a node The updated vector representation (i.e., features that aggregate neighborhood information); This represents the transpose of the attention parameter vector, used for scoring; Represents a node The set of neighboring nodes any node in ; Indicates a node original features The result of applying a linear transformation, Represents the learnable weight matrix; This represents a non-linear activation function (such as ELU / ReLU / Tanh, etc.). Representing neighboring nodes Features after linear transformation; Obtained from the graph readout or decoding operator. This indicates that the node embedding is mapped to a location. The field value can be obtained by basis function reconstruction, kernel interpolation, or multilayer perceptron decoding.

[0054] For dose field estimation Threshold exceeding probability field Calculations are performed to cross the threshold probability field. The probabilistic boundaries that form high-risk zones, short-operation zones, and safe zones are represented as: ; in, Indicates the partition threshold (dose rate or dose threshold, such as mSv / h); Indicates the significance tolerance, with a confidence level of 1. .

[0055] 1034: Based on the location of abrupt changes, an LSTM network is used to learn and provide feedback on the temporal changes in radiation dose monitoring data. This allows for the automatic identification and correction of abrupt changes and abnormal data, resulting in corrected radiation dose monitoring data. Specifically, the LSTM network can capture the dynamic patterns of radiation dose monitoring data. When radiation dose monitoring data deviates significantly from historical data, the LSTM network can detect the abrupt change and automatically adjust the model's output to correct the feedback data. This process is described by the following formula: ; ; in, This represents the hidden state of the LSTM network. This represents the input data at the current time step. This represents historical radiation dose monitoring data. This indicates the radiation dose monitoring data that has been corrected and fed back through the LSTM network; The weight matrix represents the recursion (from hidden state to hidden state); This represents the bias vector. This mechanism can identify abnormal fluctuations, monitoring point malfunctions, or data errors in radiation dose monitoring data in real time, improving the accuracy and stability of the fused data.

[0056] Specifically, for time-series monitoring data of radiation dose, deep learning models (such as Long Short-Term Memory networks (LSTM) or Gated Recurrent Units (GRUs) can learn its nonlinear dynamic characteristics and detect abnormal fluctuations or abrupt trends (such as...). Figure 5 and Figure 6 As shown, Figure 5As can be seen, when the measured dose suddenly increases at point 50, the deep learning model, after judging the changes before and after, considers this sudden increase to be an abnormal working feedback value of the measured data, and therefore does not modify the radiation dose fusion data at that time. For example, LSTM networks capture the change pattern of normal dose rate over time through memory units. When a new monitored value deviates sharply from the model prediction and rises or falls sharply, the LSTM network can identify this moment as an abnormal change point and issue a warning. Studies have shown that hybrid models based on LSTM networks (such as LSTM-CNN) have shown good performance in data stream anomaly detection and can successfully detect abnormal changes in time-series data. Similarly, self-attention models such as Transformer are also suitable for extracting the implicit change patterns in dose rate sequences. In terms of model-assisted decision optimization, reinforcement learning algorithms can be introduced: for example, using a deep Q-network (DQN) to treat continuous dose rate readings as the environmental state and instantaneous dose exceeding the limit as a penalty signal, the agent is trained to automatically adjust the working time strategy. When the monitored data shows nonlinear changes, the above intelligent model can identify and trigger automatic adjustment of safe working time in real time to ensure that personnel dose does not exceed the threshold.

[0057] 1035: Corrected radiation dose monitoring data The probabilistic boundaries of each partition and the first fused prediction data are fused to generate the second fused prediction data, such as... Figure 7 As shown in (b), the schematic diagram of the predicted data for the diffusion of radioactive contaminants from the accident is as follows. Figure 7 As shown in (a), the corrected radiation dose monitoring data reduces or re-estimates the weights of anomalous samples on the statistical, spatial, and physical sides, thereby explicitly connecting "location" and "correction" to ensure that the second fused prediction data is robust to anomalies and drift.

[0058] Preferably, the step of updating the preliminary safe operation time in real time based on the second fusion prediction data to obtain a dynamically updated safe operation time includes: In the prediction time domain (i.e., the prediction time domain length (steps or hours) of the rolling optimization) minimizes the multi-objective cost "time-dose-congestion", and applies probabilistic constraints and tail risk constraints to obtain the dynamically updated safe operating time, expressed as: ; ; in, This indicates the dose increment along the recommended course of action. ; Indicates the cost of time. This indicates a penalty for localized congestion. , , All represent weighting coefficients; Represents the discrete time step; Indicates the current time; This represents a probability operator that constrains the probability of an event occurring to not exceed a given tolerance. This indicates that "single-step dose exceeds the threshold". The probability of "is no greater than" ; Indicates the dose threshold, used for single-step or local risk control (units and...). Consistent, such as mSv); It represents the set of feasible strategies, including all strategies that satisfy action rules, work procedures, routes / traffic / personnel grouping, time windows, and safety constraints; This indicates the risk tolerance under opportunity constraints, for example, 0.05; the smaller the value, the more conservative the risk tolerance. Indicates tail risk constraint The upper limit is used to control the extreme tail expectation of the cumulative dose within the prediction time domain; The variable represents the optimization variable, which represents the joint strategy in the prediction time domain (such as personnel scheduling, travel path, start and end times of the operation, etc.). Tail risk constraints are represented; probabilistic constraints are sampled using a scenario-based approach to account for weather and source term uncertainties and provide feasibility confidence guarantees. Optimization is performed online using rolling replanning (MPC, Model Predictive Control), with real-time updates when a mutation point is triggered. The final output includes the maximum safe operating time, allowed task types, priority evacuation, entry order, and corresponding contingency plans for each protected zone.

[0059] 104 Risk Level Classification Phase: Based on the second fusion prediction data and dynamically updated safe operation time, the operation zones and risk levels are classified.

[0060] Preferably, the step of dividing the work areas and risk levels based on the second fusion prediction data and dynamically updated safe work time includes: 1041: The risk level of each zone is calculated and expressed as follows: ; in, Indicates the area The risk level, Indicates the effective dose rate. Indicates the area The dynamically updated safe operating time. This indicates that the risk assessment function, which combines effective dose rate and safe operating time, can take monotonic combinations, such as... ,in The weights can be used, or segmentation, logistic regression, or other methods can be employed. By setting different risk thresholds, the area is divided into safe work zones, short-term work zones, and high-risk zones.

[0061] Specifically, based on radiation dose and environmental changes, the method described in this application classifies high-risk zones, safe operation zones, and short-term operation zones, and marks different levels of areas on a zoning map to help the command center identify and make timely decisions. Specifically, risk level assessment and protection zoning involves dividing emergency sites into different risk level zones based on real-time dose rates and estimated doses, dynamically adjusting protection strategies. The risk level classification is mainly based on the radiation dose rate and the cumulative dose that personnel may receive in that area. Various authoritative institutions have provided reference thresholds for classification; for example, when the environmental dose rate is ≥10R / h (≈100mSv / h), the area is considered a "Dangerous Fallout Zone"; another example is the German Radiation Protection Regulation, which stipulates that if the avoidable dose exceeds 100mSv within 7 days, evacuation should be carried out promptly (>10mSv, on-site shelter is recommended). Combining these standards, accident sites can generally be divided into: (1) High-risk zone: The dose rate is extremely high, and personnel can reach a dangerous dose within a short period of time (within tens of minutes). This area usually corresponds to the vicinity of the accident source or the main deposition zone of the radioactive plume. Personnel entry should be strictly restricted and only allowed to perform special tasks such as life rescue.

[0062] (2) Short-term operation zone: The dose rate is moderate, and personnel can work for a short time with proper protection, but should not stay for a long time. For example, if the safe stay time in a certain area is less than 30 minutes, it is marked as a red zone, which is a high-risk short-term operation zone; if the safe stay time is between 30 minutes and several hours, it is a medium-risk zone (corresponding to different color level markings).

[0063] (3) Safe working zone: The dose rate is low, and personnel can work for a long time (more than several hours) without exceeding the dose limit. It is a low-risk area.

[0064] Dynamic partitioning mechanism: Based on fused real-time radiation field data, the model updates the partition boundaries on the GIS electronic map over time. When the radiation dose rate in a certain area rises and exceeds a preset threshold, the system automatically upgrades its risk level (e.g., from a safe work zone to a short-term work zone) and changes its map display color and warning markers; conversely, when the radiation level decreases, the risk level is reduced. This dynamically generated protection zoning map clearly marks high-risk, warning, and relatively safe work zones, assisting on-site commanders in rationally allocating emergency resources and formulating personnel protection measures. This ensures that high-risk zones are strictly controlled and low-risk zones are fully utilized, thereby reducing radiation risk to an acceptable level.

[0065] 1042: Using the fast travel method and the Eikonal approximation equation, the "shortest propagation metric" is solved in anisotropic media. The propagation cost is adjusted based on environmental factors such as wind speed and road network speed. The minimum dose propagation path to the target area is calculated and expressed as: ; in, This represents the gradient of the shortest path to the target region. This indicates the minimum dose propagation index reaching the target area; This represents the weighting of local optical length (e.g., the weighting of factors such as wind speed and road network speed). Indicates in the target set Boundary conditions on, i.e. .

[0066] Specifically, when When coupled with the velocity field, the shortest propagation path from the starting point to the target region is calculated using the Eikonal equation, thereby obtaining the minimum dose propagation path.

[0067] 1043: Minimizing dose exposure in the minimum dose propagation path using the HJ reachability model, and considering obstacle avoidance, yields the optimal path from the starting point to the target area, expressed as: ; in, Indicates at time At that time, from the starting point To the target area The minimum dose path value function; Represents the velocity field along the path; This represents the dose at each point along the path. The transpose of the gradient of the value function, i.e., the row vector, is used in relation to the velocity field. do inner product , For controllable direction and speed, Its feasible region; Represents a set of obstacles. This indicates that crossing the obstacle is infinitely costly (unreachable). The HJ reachability model helps assess the risk to workers by solving this equation to calculate the optimal path and its corresponding minimum dose exposure.

[0068] 1044: The dose field is calibrated to the upper quantile using distribution-independent conservative uncertainty calibration (such as conformal prediction) to obtain conservative upper quantile dose values, expressed as: ; in, This represents the conservative upper quantile dose value. Representing a path Dosage increments, This indicates the set dose limit. This indicates risk tolerance, with a confidence level of 1. This step ensures that a robust risk assessment can still be provided even when monitoring data is incomplete or contains outliers.

[0069] 1045: Based on the risk levels of each partition, the work area is divided according to the optimal path and the upper quantile dose value. Then, topological persistence is used to remove unstable small-scale structures at the work area boundaries, ensuring the stability and interpretability of the partition boundaries. Furthermore, a topological persistence method is used to screen out transient structures caused by noise at the work area boundaries at multiple scales, retaining regions with sufficient "persistence." This yields the final partition boundaries of the work areas, represented as follows: ; in, This indicates the final partition boundary of the job partition. This represents the persistence threshold, which is the threshold for retaining only the lifetime data. Topological features such as connected components / holes. This represents a set of topological features (such as stable connected components, loops, and holes) that are obtained through persistent homology and persist throughout the multi-scale filtering process. This indicates that the time to survival in a persistent barcode exceeds [a certain value]. The desirable features are preserved, while the rest are considered noise and removed. This step makes the partitioning results more engineering-executable by stabilizing the partition boundaries and eliminating unreasonable islands and holes caused by data noise or short-term anomalies.

[0070] Specifically, for the generation of protective action zones, based on the results of risk assessment, the model automatically generates operational recommendations for different protective zones, including evacuation, short-term operations, and long-term operations. These zoning plans are adjusted in real-time based on multi-source data, providing intelligent and dynamic protective zoning recommendations for emergency response.

[0071] The dynamic safe operation time assessment results generated by the method of this application are as follows: Figure 8 As shown, Figure 8 Figure (a) shows the estimated safe working time a hours after the accident is released. Figure 8Figure (b) shows the estimated safe working time b hours after the accident release. It can be seen that as the accident progresses, the area affected by the pollutants expands, and the high-dose area increases. In this embodiment, the red area is designated as a high-risk zone, and the blue area as a short-term work zone (where the high-risk zone is the affected area in this embodiment, and the short-term work zone is the area near the accident site). Six emergency work locations are set. It can be seen that each emergency work location displays its safe working time at different times, and its safe working time is dynamically updated as the accident progresses. When the accident case time changes from a to b, the area affected by the pollution increases, causing area 5 to change from a short-term work zone to a high-risk zone. Simultaneously, the safe working time at each emergency work location is shortened to varying degrees.

[0072] Preferably, by combining public protection action dosage standards and emergency personnel dosage constraints, safe operating times and risk levels are set for each zone, and dynamically configured according to actual accident scenarios. parameter.

[0073] In nuclear accident emergency response, areas are divided into different risk levels based on public dose limits and emergency personnel workload limits. According to the different risk levels of each zone (e.g., high-risk zones, suitable work zones, short-term work zones), corresponding safe work times and task priorities are configured. ; These parameters align with existing nuclear emergency preparedness action guidelines (such as the EPA PAG Manual and the IAEA EPR series) to ensure compliance and consistency of risk assessments in nuclear accident emergency response.

[0074] By employing the methods described above, combined with the Eikonal approximation, the HJ reachability model, conservative uncertainty calibration, and topology persistence, this application provides an efficient, accurate, and dynamic method for nuclear accident risk assessment and operation time calculation. This method can adjust operational strategies based on real-time monitoring data, ensuring the efficiency of emergency response and personnel safety.

[0075] The nuclear accident dynamic safe operation time assessment system based on data fusion described in this application includes a data acquisition module, a data assimilation and fusion module, a safe operation time assessment module, a feedback and dynamic adjustment module, and a risk level classification module.

[0076] The data acquisition module acquires nuclear accident consequence assessment prediction data and other data within the nuclear accident assessment area, and acquires monitoring data at the nuclear accident site; wherein, the prediction data includes continuously changing radiation diffusion trends and dose distribution maps, the monitoring data includes historical and current radiation dose monitoring data and meteorological observation data; the other data includes population distribution at different times and shielding factors of various buildings; The data assimilation and fusion module performs assimilation and fusion calculations on the predicted data and the monitoring data to generate continuously changing first fused predicted data. The safe operation time assessment module analyzes and calculates the preliminary safe operation time for different dynamically changing zones based on the first fused prediction data and other data. The feedback and dynamic adjustment module captures and corrects abnormal data in the current monitoring data, then merges the corrected abnormal data with the first fusion prediction data to generate the second fusion prediction data, and updates the preliminary safe operation time in real time based on the second fusion prediction data to obtain the dynamically updated safe operation time. The risk level classification module classifies work zones and risk levels based on the second fusion prediction data and dynamically updated safe operation time.

[0077] The following embodiments, based on a nuclear accident scenario, detail how to use the method described in this application to assess the dynamic safe working time in a nuclear accident, and to evaluate the safe working time of different personnel in different operating rooms. Specifically, this embodiment assumes that after a nuclear accident, several emergency personnel need to reach different operating rooms to perform operations. We will combine simulated data with actual calculation processes to demonstrate how to dynamically assess the safe working time of each worker.

[0078] (1) Background assumptions Assuming a nuclear power plant accident occurs, the radiation spread in the accident area is complex. On-site radiation levels, weather conditions, building shielding effectiveness, and other environmental factors are subject to change at any time. Emergency response personnel need to make rapid decisions based on real-time assessments to ensure personnel safety and execute necessary operational procedures.

[0079] Accident Area: The nuclear accident is set to occur in an area of ​​approximately 10 square kilometers, which contains multiple operating rooms (such as equipment maintenance rooms, control rooms, etc.). Personnel working in each area need to go to a specific operating room to perform tasks.

[0080] Operation rooms: For example, operation rooms A, B, C, etc., each with different task types, risk levels, and operation times.

[0081] Emergency personnel: Five emergency personnel are required to travel to different operating rooms to perform tasks within a specified time. The safe working time for each personnel is calculated based on real-time radiation dose and other environmental factors.

[0082] (2) Data reception and preprocessing Following a nuclear accident, the following data were first received from sensors and weather stations at the scene: Nuclear accident prediction data: radiation diffusion trends, dose distribution maps, etc., obtained from existing nuclear accident consequence assessment systems (such as JRODOS and ARGOS).

[0083] Real-time monitoring and other data: on-site radiation dose, meteorological conditions (such as wind speed, wind direction, etc.), building shielding factor, personnel distribution, etc.

[0084] For example, suppose the following data is received: Radiation dose: In operating room A, the real-time radiation dose is 5 mSv / h; in operating room B, it is 8 mSv / h; and in operating room C, it is 10 mSv / h.

[0085] Meteorological conditions: wind speed of 3 m / s and wind direction of 45° affect the diffusion of radiation.

[0086] Building shielding factor: The shielding factor of operating room A is 0.6, that of operating room B is 0.8, and that of operating room C is 0.5.

[0087] These data will be preprocessed and prepared for further analysis.

[0088] (3) Data assimilation and fusion Data assimilation algorithms (such as ensemble Kalman filtering and variational assimilation methods) are used to fuse real-time monitoring data and predicted data to obtain the first fused predicted data. The effective dose rate is then calculated using the following formula: ; in, For effective dose rate, For environmental dose rate, This is the shielding attenuation coefficient.

[0089] For example, assuming the ambient dose rate of operating room A is 5 mSv / h, the shielding attenuation coefficient is... The effective dose rate of operating room A is: ; (4) Assessment of safe working time Based on the first fusion prediction data, the safe operation time for each area is calculated. The formula for calculating the operation time is: ; in, Indicates safe working time. For the mission dose limit, For the cumulative dose, This represents the effective dose rate.

[0090] Assuming the emergency personnel's task dose limit is 50 mSv and the accumulated dose is 5 mSv, the safe operating time for operating room A is calculated as follows: ; Similarly, the safe operating time for operating rooms B and C can be calculated.

[0091] (5) Dynamic feedback and adjustment The LSTM network and graph attention propagation algorithm provide dynamic feedback on real-time monitoring data, identify abnormal data, and update work time and risk assessment. For example, when a sudden increase in radiation dose rate is detected in operating room A (assuming it rises to 20 mSv / h, and the personnel in operating room A have accumulated a radiation dose of 10 mSv), the LSTM network can automatically identify the sudden change and adjust the work time in a timely manner.

[0092] The adjusted safe working time for personnel in Operation Room A is 2 hours. If the task in Operation Room A is to be completed in 6 hours, it would be recommended to send 2 more emergency personnel to Operation Room A to complete the task (analysis suggests that if new, untreated emergency personnel are sent to the operation, the safe working time is expected to be 2.5 hours, so 2 more personnel are needed).

[0093] (6) Risk level classification and work area adjustment Based on the second fusion prediction data and dynamic safe operation time, a dose-budget-based reachability-avoidance zoning method is used to divide the operation zones. For example, the risk level of each zone is calculated using the Eikonal approximation and the HJ reachability model, and the zones are divided into high-risk zones, short-term operation zones, and safe operation zones.

[0094] ; in, Indicates the area The risk level, This represents the effective dose rate in the region. This refers to the safe operating time in this area.

[0095] (7) Generation of assignment suggestions Based on the above assessment results, operational recommendations are generated for each emergency personnel. For example, originally, emergency personnel 1 was required to go to operation room A with a safe working time of 15 hours. Now, the safe working time for operation room A is only 2 hours, so at least 2 more emergency personnel (emergency personnel 4 and 5) need to be dispatched to the operation. Emergency personnel 2 is required to go to operation room B with a safe working time of 10 hours, and so on.

[0096] The method described in this application allows for real-time updates and optimization of work time assessments for different personnel, ensuring that workers can make informed decisions based on real-time radiation levels, safe work times, and risk assessments.

[0097] The above are exemplary embodiments of this application, and the scope of protection of this application is defined by the claims and their equivalents.

Claims

1. A method for assessing dynamic safe working time in nuclear accidents based on data fusion, characterized in that, include: Data acquisition phase: Acquire nuclear accident consequence assessment prediction data and other data within the nuclear accident assessment area, and acquire monitoring data at the nuclear accident site; wherein, the prediction data includes continuously changing radiation diffusion trends and dose distribution maps, the monitoring data includes historical and current radiation dose monitoring data and meteorological observation data; the other data includes population distribution at different times and shielding factors of various buildings; Data assimilation and fusion stage: Assimilation and fusion calculations are performed on the predicted data and monitoring data to generate continuously changing first fused predicted data; Safe operation time assessment phase: Based on the first fusion prediction data and other data, the preliminary safe operation time of different zones with dynamic changes is analyzed and calculated; Feedback and dynamic adjustment phase: Abnormal data in the current monitoring data is captured and corrected. Then, the corrected abnormal data is fused with the first fusion prediction data to generate the second fusion prediction data. The preliminary safe operation time is updated in real time based on the second fusion prediction data to obtain the dynamically updated safe operation time. Risk level classification phase: Based on the second fusion prediction data and dynamically updated safe operation time, the operation zones and risk levels are classified.

2. The method for assessing dynamic safe working time in nuclear accidents based on data fusion as described in claim 1, characterized in that, The assimilation and fusion calculation of predicted data and real-time data includes: The spatiotemporal field corresponding to the predicted data Observational constraint field constructed from sparse monitoring data In the spatial domain Deformation matching is performed internally, and then the non-equilibrium extension of the Benamou–Brenier dynamic formula is solved to obtain the deformation trajectory with minimum "transport work". , and final state , making the spacetime field Under the constraint of non-conservation of mass, it is continuously deformed into the observation constraint field. The uniform nearest neighbors are represented as: ； ； in, ; Represents the deformation velocity field; , indicating spatial location; This represents the dose field corresponding to the predicted data; This represents the field obtained by interpolating monitoring data, used to represent the consistency target of observations; This represents the density or field value during deformation; This represents non-conservative source and sink terms to characterize sedimentation and decay effects; For the weighting factor; The constraint function representing consistency with observations. , Indicates the first Location of each monitoring point For quality weights; , representing the consistency tolerance threshold; let the model field after deformation alignment be . ; with model field For low-fidelity references, monitoring data is high-fidelity information; for residuals... Modeling was performed, and the Co-Kriging method was used to analyze the residuals. Recursively incorporate heteroscedastic noise, as follows: ； in, , This indicates the high-fidelity field corresponding to the monitoring data; Representing a Gaussian process , Represents the kernel function; This represents the heteroscedasticity term related to location and time. Indicates the cross-fidelity scale coefficient; according to Obtain the residual posterior mean With variance Thus, the statistical correction field is obtained, expressed as: ； in, This represents the high-fidelity field of the monitoring data; ; Residual The posterior mean of the Gaussian process, and Residual The posterior variance of the Gaussian process; This indicates the correction of the low-fidelity model field, and ; Statistical correction field For the data term, a physical residual constraint is introduced from the advection-diffusion-decay equation, expressed as: ； in, Indicates the field to be merged; This represents the partial derivative with respect to time; Represents the advection velocity field; Indicates the anisotropic diffusion coefficient; Indicates radioactive decay; Indicates the source term; The joint loss is constructed based on the physical residual constraints and is expressed as follows: ； in, Represents the weighting coefficient of the physical residual term; Indicates the weights of the boundary constraint terms; This represents a measure of the deviation between the boundary conditions and the initial conditions. The neural network is trained using a joint loss function to obtain a fusion solution that satisfies the consistency between the physical equations and the observations. , That is, the first fused prediction data; among which, To achieve the optimal network output field with the best parameters, To achieve a minimum while simultaneously satisfying both data consistency and physical consistency; In scenarios involving sensor anomalies or monitoring distribution drift, the Wasserstein sphere is used. For an uncertain set, perform sub-Bruker optimization to suppress the impact of extreme monitoring on the first fused prediction data, expressed as: ； Represent the hypothesis space of feasible functions; This represents the distribution of candidate data, taken from an empirical distribution. Centered on, with radius Wasserstein's ball; Relative to the distribution The expectation operator; Representation and observation constraint field The loss due to mismatch; An empirical distribution composed of historical observation samples; Used to control the robust radius against distribution drift.

3. The method for assessing dynamic safe working time in nuclear accidents based on data fusion as described in claim 2, characterized in that, The analysis and calculation of the preliminary safe operation time for dynamically changing different partitions based on the first fusion prediction data and other data includes: Dose thresholds are determined according to radiation protection standards; The environmental dose rate at each location is predicted by combining an atmospheric diffusion model with current meteorological observation data. By introducing a building shielding factor to correct for the environmental dose rate, the effective dose rate is obtained, expressed as: ； in, Indicates the effective dose rate. Indicates the environmental dose rate. Indicates the shielding attenuation coefficient; The initial safe operating time, derived from the effective dose rate, is expressed as follows: ； in, Indicates the initial safe operation time. This indicates the dose limit for this mission. This indicates the cumulative dosage.

4. The method for assessing dynamic safe working time in nuclear accidents based on data fusion as described in claim 3, characterized in that, ;in, Represents a linear decay function. This indicates the thickness of the shielding material.

5. The method for assessing dynamic safe working time in nuclear accidents based on data fusion as described in claim 4, characterized in that, The process of generating the second fused prediction data includes: The time series corresponding to multi-point monitoring are stacked into a time data matrix according to time sliding windows. For time data matrix Robust principal component decomposition is performed to separate the stationary background. With sparsity anomaly Then we get the denoised background field. and anomaly mask , is represented as: ； in, Represents the nuclear norm. express Norm; This represents the sparse regularization weight; The dose rate sequence corresponding to the radiation dose monitoring data at each monitoring point was obtained using the online Bayesian change point detection method. The current runtime is calculated posteriorly to locate mutation points, represented as: ； in, This indicates the runtime, i.e., from the last mutation point to the current point. The length of the segment; This represents the posterior probability of the run length given all current observations. The prior transition probability represents the length of the operation; This indicates the predicted likelihood based on historical samples within the current segment; This represents the posterior length of the previous run; when Exceeding the threshold If a point is reached, that point is marked as a mutation point, and the data corresponding to that point is considered abnormal data. Construct a sensor map with monitoring stations as nodes and meteorological correlation, geographical proximity, and wind direction coupling as edges. For the background field Perform graph attention propagation to obtain spatially consistent dose field estimates. , is represented as: ； ； ； in, This represents the set of nodes, i.e., the set of monitoring points; Represents the set of edges; Represents attention weights and can be used for anomaly masks. Punishment is imposed to suppress the effects of contaminated sensors; Represents a node The updated vector representation; This represents the transpose of the attention parameter vector; Represents a node The set of neighboring nodes any node in ; Indicates a node original features The result of applying a linear transformation, Represents the learnable weight matrix; Represents a nonlinear activation function; Representing neighboring nodes Features after linear transformation; This indicates that the node embedding is mapped to a location. The field value; For dose field estimation Threshold exceeding probability field Calculations are performed to cross the threshold probability field. The probabilistic boundaries that form high-risk zones, short-operation zones, and safe zones are represented as: ； in, Indicates the partition threshold; Indicates the significance tolerance, with a confidence level of 1. ; Based on the location of abrupt change points, an LSTM network is used to learn and provide feedback on the temporal changes in radiation dose monitoring data. This allows for the automatic identification and correction of abrupt change points and abnormal data, resulting in corrected radiation dose monitoring data, represented as follows: ； ； in, This represents the hidden state of the LSTM network. This represents the input data at the current time step. This represents historical radiation dose monitoring data. This indicates the radiation dose monitoring data that has been corrected and fed back through the LSTM network; The weight matrix represents the recursion. Represents the bias vector; Corrected radiation dose monitoring data The probabilistic boundaries of each partition and the first fusion prediction data are fused to generate the second fusion prediction data.

6. The method for assessing dynamic safe working time in nuclear accidents based on data fusion as described in claim 5, characterized in that, The step of updating the preliminary safe operation time in real time based on the second fusion prediction data to obtain the dynamically updated safe operation time includes: In the prediction time domain The above minimizes the multi-objective cost "time-dose-congestion," and applies probabilistic and tail risk constraints to obtain the dynamically updated safe operating time, expressed as: ； ； in, This indicates the dose increment along the recommended course of action. ; Indicates the cost of time. This indicates a penalty for localized congestion. , , All represent weighting coefficients; Represents the discrete time step; Indicates the current time; This represents a probability operator that constrains the probability of an event occurring to not exceed a given tolerance. This indicates that "single-step dose exceeds the threshold". The probability of "is no greater than" , Indicates the dose threshold; Represents the set of feasible strategies; This indicates the risk tolerance under opportunity constraints; Indicates tail risk constraint The upper limit; Represents the optimization variable; This represents the tail risk constraint; the probability constraint samples weather and source term uncertainties using the scenario method and provides a feasibility confidence guarantee.

7. The method for assessing dynamic safe working time in nuclear accidents based on data fusion as described in claim 6, characterized in that, The process of dividing work zones and risk levels based on the second fusion prediction data and dynamically updated safe work time includes: The risk level of each zone is calculated and expressed as follows: ； in, Indicates the area The risk level, Indicates the effective dose rate. Indicates the area Dynamically updated safe operating time This represents a risk assessment function that combines effective dose rate and safe operating time. The minimum dose propagation path to the target area is calculated using the fast travel method and the Eikonal approximation equation, and is expressed as follows: ； in, This represents the gradient of the shortest path to the target region. This represents the minimum dose propagation index reaching the target area. Represents the local optical length weight. Indicates in the target set Boundary conditions on, i.e. ; By minimizing dose exposure in the minimum dose propagation path using the HJ reachability model and considering obstacle avoidance, the optimal path from the starting point to the target area is obtained, expressed as: ； in, Indicates at time At that time, from the starting point To the target area The minimum dose path value function; Represents the velocity field along the path; This represents the dose at each point in the path; This represents the transpose of the gradient of the value function; The dose field is calibrated to the upper quantile using distribution-independent conservative uncertainty calibration, resulting in conservative upper quantile dose values, expressed as: ； in, This represents the conservative upper quantile dose value. Representing a path Dosage increments, This indicates the set dose limit. Indicates risk tolerance; Based on the risk levels of each zone, the work zones are divided according to the optimal path and the upper quantile dose value. Then, unstable small-scale structures in the work zone boundaries are removed by topological persistence techniques, and transient structures caused by noise at the work zone boundaries are screened out at multiple scales using persistent cohomology methods. The final zone boundaries of the work zones are obtained, represented as: ； in, This indicates the final partition boundary of the job partition. Indicates the persistence threshold; This represents the set of topological features obtained through persistent cohomology and that persists throughout the multi-scale filtering process. This indicates that the time to survival in a persistent barcode exceeds [a certain value]. The features are preserved, while the rest are treated as noise and removed.

8. A dynamic safety operation time assessment system for nuclear accidents based on data fusion, characterized in that, The evaluation system is used in the evaluation method according to any one of claims 1-7, and the evaluation system comprises: The data acquisition module acquires nuclear accident consequence assessment prediction data and other data within the nuclear accident assessment area, and acquires monitoring data at the nuclear accident site; wherein, the prediction data includes continuously changing radiation diffusion trends and dose distribution maps, the monitoring data includes historical and current radiation dose monitoring data and meteorological observation data; the other data includes population distribution at different times and shielding factors of various buildings; The data assimilation and fusion module performs assimilation and fusion calculations on the predicted data and the monitoring data to generate continuously changing first fused predicted data. The safe operation time assessment module analyzes and calculates the preliminary safe operation time for different dynamically changing zones based on the first fused prediction data and other data. The feedback and dynamic adjustment module captures and corrects abnormal data in the current monitoring data, then merges the corrected abnormal data with the first fusion prediction data to generate the second fusion prediction data, and updates the preliminary safe operation time in real time based on the second fusion prediction data to obtain the dynamically updated safe operation time. The risk level classification module classifies work zones and risk levels based on the second fusion prediction data and dynamically updated safe operation time.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the evaluation method as described in any one of claims 1 to 7.

10. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the evaluation method as described in any one of claims 1 to 7.

Images