Intelligent evaluation and early warning method and system for nuclear radiation dose spatial distribution

Abstract

The invention relates to the technical field of intelligent evaluation and early warning methods, in particular to a nuclear radiation dose space distribution intelligent evaluation and early warning method and system, and the method comprises the following steps: obtaining multi-source monitoring data which comprises gamma radiation dose, alpha radiation dose and beta radiation dose measuring instrument data; performing fusion and enhancement processing on the multi-source monitoring data to obtain an enhanced radiation monitoring signal; based on the enhanced radiation monitoring signal, high-precision prediction of nuclear radiation dose spatial distribution is carried out through a deep probabilistic neural network model; according to the high-precision prediction result, in combination with a nuclear accident severity level standard, a radiation alert evaluation result and early warning information are generated, short-term radiation fluctuation can be captured, a long-term change trend can be identified, and the understanding ability of the system to a complex radiation field is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent assessment and early warning methods, and more specifically, to an intelligent assessment and early warning method for nuclear radiation dose spatial distribution and a system thereof. Background Art

[0002] In recent years, as the importance of nuclear energy in the global energy structure has become increasingly prominent, the safe operation of nuclear facilities and nuclear radiation monitoring have become the focus of international attention. The intelligent assessment and early warning system of the spatial distribution of nuclear radiation dose plays a vital role in ensuring nuclear safety and preventing the risk of nuclear accidents. Traditional nuclear radiation monitoring methods mainly rely on discrete data from fixed monitoring points. Although this method is simple and intuitive, it is difficult to fully capture the dynamic changes and spatial distribution characteristics of the radiation field in a complex environment.

[0003] With the advancement of technology, some advanced nuclear radiation monitoring systems have begun to adopt multi-source data fusion and machine learning algorithms to improve prediction accuracy. However, these systems still have some significant shortcomings. First, most existing systems are not effective in processing multi-scale spatiotemporal data, and it is difficult to capture short-term fluctuations and long-term trends at the same time. Second, existing machine learning models often ignore the basic principles of radiation physics, resulting in prediction results that may violate the laws of physics, reducing the credibility and practicality of the model. In addition, existing systems generally lack quantification of prediction uncertainty, which is a serious flaw in such a high-risk field as nuclear safety. Finally, the existing assessment and early warning mechanisms are often too rigid, making it difficult to flexibly adjust response strategies according to different situations.

[0004] The existence of these problems seriously restricts the effectiveness of nuclear radiation monitoring systems and increases the potential risks of nuclear facility operations. Today, when nuclear safety is so important, there is an urgent need for new methods that can comprehensively, accurately and in real time assess the spatial distribution of nuclear radiation doses and provide intelligent early warning. Summary of the invention

[0005] The present invention aims to solve the above technical problems and provide a method and system for intelligent evaluation and early warning of nuclear radiation dose spatial distribution. This method realizes high-precision prediction and intelligent evaluation and early warning of nuclear radiation dose spatial distribution through an innovative deep probabilistic neural network model, combined with multi-source data fusion, multi-scale spatiotemporal feature extraction, physical constraint mechanism and uncertainty quantification.

[0006] The present invention provides an intelligent evaluation and early warning method for nuclear radiation dose spatial distribution, which comprises the following steps:

[0007] Acquiring multi-source monitoring data, wherein the multi-source monitoring data includes gamma radiation dose, alpha radiation dose and beta radiation dose measuring instrument data;

[0008] fusing and enhancing the multi-source monitoring data to obtain an enhanced radiation monitoring signal;

[0009] Based on the enhanced radiation monitoring signal, a deep probabilistic neural network model is used to perform high-precision prediction of the spatial distribution of nuclear radiation dose;

[0010] Based on the high-precision prediction results and combined with the nuclear accident severity level standards, radiation alert evaluation results and warning information are generated.

[0011] Preferably, the deep probabilistic neural network model includes a multi-scale spatiotemporal feature extraction module, an adaptive attention mechanism, a physical constraint enhanced prediction module, an uncertainty-aware integrated learning module and an adaptive multi-task learning optimizer, and these modules process data in sequence as follows:

[0012] The multi-scale spatiotemporal feature extraction module receives the enhanced radiation monitoring signal and outputs fused spatiotemporal features;

[0013] The adaptive attention mechanism dynamically adjusts the importance of different spatiotemporal features based on the fused spatiotemporal features;

[0014] The physical constraint enhanced prediction module uses the output of the adaptive attention mechanism to generate prediction results that conform to physical laws;

[0015] The uncertainty-aware integrated learning module improves the robustness of prediction by integrating multiple models based on the output of the physical constraint-enhanced prediction module;

[0016] The adaptive multi-task learning optimizer integrates the outputs of all the aforementioned modules to achieve adaptive optimization of model parameters.

[0017] Preferably, the mathematical expression of the multi-scale spatiotemporal feature extraction module is:

[0018]

[0019] Among them, X t is the input feature at time t, is the hidden state of the ith time scale, is the feature of the jth spatial scale, is the weight matrix, b i ,c j is the bias term, f,g is the activation function, F t It is the fused spatiotemporal features.

[0020] Preferably, the mathematical expression of the adaptive attention mechanism is:

[0021]

[0022] α t =softmax(e t )

[0023]

[0024] h t =GRU(c t ,h t-1 )

[0025] Among them, F t is the output of the multi-scale spatiotemporal feature extraction module 1, h t-1 is the hidden state of the previous moment, v, W, b are learnable parameters, is the attention score of the k-th feature, α t is the attention weight, c t is the context vector, h t is the hidden state at the current moment. GRU is a gated recurrent unit.

[0026] Preferably, the mathematical expression of the prediction module enhanced by physical constraints is:

[0027] [μ t ,σ t ]=MLP(h t )

[0028]

[0029] L physics =λ||Φ(y t )-Φ(y true )|| 2

[0030] Among them, h t is the output of adaptive attention mechanism 2, MLP is multi-layer perceptron, μ t ,σ t are the predicted mean and standard deviation, respectively. is the initial prediction result, y t is the prediction result after physical constraints, Φ is the physical constraint function, L physics is the physical constraint loss and λ is the trade-off factor.

[0031] Preferably, the mathematical expression of the uncertainty-aware integrated learning module is:

[0032]

[0033] in, represents the i-th model, are the predicted value and uncertainty of the ith model, is the weight of the i-th model, β is the temperature parameter, are the integrated predicted value and uncertainty respectively.

[0034] Preferably, the mathematical expression of the adaptive multi-task learning optimizer is:

[0035] L total =L pred +L physics +L uncertainty

[0036]

[0037] m t =β1m t-1 +(1-β1)g t

[0038]

[0039] η t =η base (1+cos(πt / T)) / 2

[0040]

[0041] Among them, L total is the total loss, L pred To predict the loss, L physics is the physical constraint loss, L uncertainty is the uncertainty loss, g t is the gradient, m t ,v t are the first-order and second-order moment estimates, β1, β2 are momentum parameters, η base is the basic learning rate, T is the period, θ t are model parameters.

[0042] Preferably, the fusion and enhancement process comprises the following steps:

[0043] Correcting the multi-source monitoring data by combining physical methods and statistical regression to obtain calibrated signal data;

[0044] Get sensor posture and baseline data;

[0045] The calibrated signal data is fused with the sensor posture and baseline data to obtain the enhanced radiation monitoring signal.

[0046] Preferably, the nuclear accident severity level standard is based on the accident classification standard of the International Atomic Energy Agency (IAEA), and is combined with a Monte Carlo simulation method to establish a dose-level mapping model, which is used to convert the predicted radiation dose into the corresponding nuclear accident severity level.

[0047] An intelligent evaluation and early warning system for nuclear radiation dose spatial distribution implementing the method comprises:

[0048] A data acquisition module, used to acquire multi-source monitoring data;

[0049] A signal processing module, used for fusing and enhancing the multi-source monitoring data;

[0050] Prediction module, used to make high-precision predictions of the spatial distribution of nuclear radiation doses through a deep probabilistic neural network model;

[0051] An evaluation and warning module is used to generate radiation warning evaluation results and warning information based on the high-precision prediction results and combined with the nuclear accident severity level standards;

[0052] Among them, the prediction module includes a multi-scale spatiotemporal feature extraction unit, an adaptive attention mechanism unit, a physical constraint enhanced prediction unit, an uncertainty-aware integrated learning unit and an adaptive multi-task learning optimization unit, and these units perform data processing and model optimization in the described manner.

[0053] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0054] First, the multi-scale spatiotemporal feature extraction module and adaptive attention mechanism of the present invention realize the effective integration of radiation data at different time and space scales. This design can not only capture short-term radiation fluctuations, but also identify long-term change trends, greatly improving the system's ability to understand complex radiation fields. For example, in the daily operation of a nuclear power plant, the system can simultaneously monitor short-term radiation fluctuations caused by equipment maintenance and long-term trend changes caused by fuel decay.

[0055] Secondly, the physical constraint enhanced prediction module introduced in this invention cleverly integrates the principles of radiation physics into the deep learning model. This innovative design solves the problem that pure data-driven methods may violate physical laws and significantly improves the reliability and interpretability of prediction results. In practical applications, even when data is sparse or noisy, the system can still maintain the physical rationality of the prediction, which is particularly important for nuclear safety supervision.

[0056] Furthermore, the uncertainty-aware integrated learning module of the present invention greatly enhances the robustness and credibility of the system by integrating multiple models and quantifying the prediction uncertainty. This feature is particularly important when facing abnormal or unknown situations. For example, when abnormal radiation levels are detected, the system can not only give a predicted value, but also provide a reliable confidence interval, providing more comprehensive information support for decision makers.

[0057] In addition, the adaptive multi-task learning optimizer of the present invention achieves balanced optimization of multiple objectives, so that the system can maintain physical consistency and reasonable uncertainty estimation while ensuring prediction accuracy. This balance is particularly important in high-risk fields such as nuclear radiation monitoring, which ensures that the system does not sacrifice reliability and stability while pursuing high accuracy.

[0058] Finally, the evaluation and warning module of the present invention is based on an advanced dose-level mapping model, which can automatically adjust the warning level and response strategy according to different radiation levels and uncertainties. This flexible design greatly improves the practicality and effectiveness of the system and can better cope with various possible nuclear accident scenarios.

[0059] In summary, the present invention achieves high-precision prediction, reliable evaluation and intelligent early warning of nuclear radiation dose spatial distribution through the synergy of multiple innovative modules. This comprehensive and in-depth method not only significantly improves the accuracy and reliability of nuclear radiation monitoring, but also provides strong technical support for the safe operation of nuclear facilities and emergency response to potential nuclear accidents, which is of great significance to improving nuclear safety and protecting the environment and public health. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The figure is a logic block diagram of the method of the present invention.

[0061] Figure 2 Flow chart of the fusion and enhancement process of the present invention.

[0062] Figure 3 Schematic diagram of the internal structure of the deep probabilistic neural network model of the present invention.

[0063] Figure 4 It is a logic flow chart of the evaluation and early warning module of the present invention.

[0064] Figure 5 It is a logic block diagram of the system of the present invention. DETAILED DESCRIPTION

[0065] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0066] Unless defined otherwise, 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 belongs.

[0067] Please refer to Figure 1-5 The present invention provides an intelligent evaluation and early warning method for nuclear radiation dose spatial distribution, comprising the following steps: acquiring multi-source monitoring data, the multi-source monitoring data including gamma radiation dose, alpha radiation dose and beta radiation dose measuring instrument data; fusing and enhancing the multi-source monitoring data to obtain an enhanced radiation monitoring signal; based on the enhanced radiation monitoring signal, performing high-precision prediction of nuclear radiation dose spatial distribution through a deep probabilistic neural network model; generating radiation alert evaluation results and early warning information according to the high-precision prediction results and in combination with nuclear accident severity standards.

[0068] In the vicinity of nuclear power plants or nuclear facilities, a plurality of different types of radiation monitoring equipment are usually arranged. In a preferred embodiment of the present invention, these equipment include a gamma radiation dose meter, an alpha radiation dose meter and a beta radiation dose meter. The gamma radiation dose meter is mainly used to detect high-energy photon radiation, and its sensitivity is usually between 0.01μSv / h and 10Sv / h. The alpha radiation dose meter is used to detect helium nuclear particle radiation, and its detection limit can reach 0.1Bq / m 3 The β radiation dose meter is used to detect electron or positron radiation, and its measurement range is usually between 0.1μSv / h and 1Sv / h. By simultaneously acquiring these three types of radiation data, the method of the present invention can fully understand the radiation environment around the nuclear facility.

[0069] After acquiring the multi-source monitoring data, the method of the present invention fuses and enhances these data. In another embodiment of the present invention, this step uses a Kalman filter to perform preliminary processing on the raw data to remove high-frequency noise. The parameter setting of the Kalman filter depends on the specific noise characteristics. Usually, the process noise covariance Q is set to between 0.001 and 0.1, and the measurement noise covariance R is set to between 0.1 and 1. Then, through a multi-sensor data fusion algorithm, such as Bayesian fusion or evidence theory fusion, different types of radiation data are integrated to obtain a more comprehensive and accurate radiation monitoring signal.

[0070] Next, the core of the present invention is to use a deep probabilistic neural network model to perform high-precision prediction of the spatial distribution of nuclear radiation dose. The model includes a multi-scale spatiotemporal feature extraction module 1, an adaptive attention mechanism 2, a physical constraint enhanced prediction module 3, an uncertainty-aware integrated learning module 4, and an adaptive multi-task learning optimizer 5. These modules process data in a specific order to form a complete prediction process.

[0071] The multi-scale spatiotemporal feature extraction module 1 first receives the enhanced radiation monitoring signal and outputs the fused spatiotemporal features. In one embodiment of the present invention, the mathematical expression of the module is:

[0072]

[0073] Among them, X t is the input feature at time t is the hidden state of the ith time scale, is the feature of the jth spatial scale, is the weight matrix, b i ,c j is the bias term, f,g is the activation function, F t is the fused spatiotemporal feature

[0074] In a preferred embodiment of the present invention, 3-5 time scales and 2-3 spatial scales are selected. Specifically, the time scales may include 1 hour, 6 hours, 24 hours, 7 days and 30 days, while the spatial scales may include point level (single monitoring station), regional level (average of multiple adjacent monitoring stations) and global level (overall situation of all monitoring stations). This multi-scale design enables the model to capture short-term fluctuations and long-term trends, as well as local anomalies and global patterns at the same time.

[0075] The adaptive attention mechanism 2 dynamically adjusts the importance of different spatiotemporal features based on the fused spatiotemporal features. In another embodiment of the present invention, the mathematical expression of the mechanism is:

[0076]

[0077] α t =softmax(e t

[0078]

[0079] h t =GRU(c t ,h t-1 )

[0080] Among them, F t is the output of the multi-scale spatiotemporal feature extraction module 1, ht-1 is the hidden state of the previous moment, v, W, b are learnable parameters, is the attention score of the k-th feature, α t is the attention weight, c t is the context vector, h t is the hidden state at the current moment GRU is the gated recurrent unit

[0081] In nuclear radiation monitoring, the adaptive attention mechanism 2 can automatically identify important monitoring points and time periods. For example, if a monitoring point detects an abnormally high radiation value, such as 10 times the background value, the attention mechanism will automatically increase the weight of the monitoring point, making the model pay more attention to this potential danger signal.

[0082] The prediction module 3 enhanced by physical constraints utilizes the output of the adaptive attention mechanism 2 to generate prediction results that conform to physical laws. In one embodiment of the present invention, the mathematical expression of this module is:

[0083] [μ t ,σ t ]=MLP(h t )

[0084]

[0085] L physics =λ||Φ(y t )-Φ(y true )|| 2

[0086] Among them, h t is the output of adaptive attention mechanism 2, MLP is multi-layer perceptron, μ t ,σ t are the predicted mean and standard deviation, respectively. is the initial prediction result, y t is the prediction result after physical constraints, Φ is the physical constraint function, L physics is the physical constraint loss and λ is the trade-off factor.

[0087] In a preferred embodiment of the present invention, the physical constraints include: 1) the radiation intensity cannot be negative; 2) the radiation intensity decays with distance following the inverse square law, that is, the intensity is inversely proportional to the square of the distance; 3) the radiation intensity change in a short period of time (such as within 1 hour) cannot exceed the physically possible maximum value, which is usually set to 5 times the current value. These constraints ensure that the prediction results of the model always conform to the principles of radiation physics, and improve the reliability and interpretability of the prediction.

[0088] The first three modules of the core algorithm of the present invention are described in detail above. These modules work together to form a powerful nuclear radiation dose spatial distribution prediction system. Through the multi-scale spatiotemporal feature extraction module 1, the present invention captures the complex patterns of radiation data; through the adaptive attention mechanism 2, the present invention dynamically focuses on important information; through the physical constraint enhanced prediction module 3, the present invention ensures that the prediction results conform to the principles of radiation physics. This method that combines data-driven and physical models can provide reliable and interpretable results while ensuring prediction accuracy, which is of great significance for nuclear safety monitoring and emergency response. Next, the present invention will describe the uncertainty-aware integrated learning module 4 in detail. This module is designed to improve the reliability and robustness of nuclear radiation dose prediction. In one embodiment of the present invention, the mathematical expression of the module is as follows:

[0089]

[0090] in, represents the i-th model, are the predicted value and uncertainty of the ith model, is the weight of the i-th model, β is the temperature parameter, are the integrated predicted value and uncertainty respectively.

[0091] In a preferred embodiment of the present invention, 5 to 10 different models are used for ensemble. Each model may have a slightly different structure or initialization to capture the uncertainty of the prediction. The temperature parameter β is usually set between 0.1 and 1 to adjust the degree of influence of uncertainty on the weight. A lower β value will make the weight distribution more uniform, while a higher β value will make the weight more concentrated on models with lower uncertainty.

[0092] This module is particularly important in the scenario of nuclear radiation monitoring. For example, when the radiation level in a certain area suddenly rises from the normal 0.1μSv / h to 1μSv / h, different models may give different predictions. Some models may think that this is a temporary fluctuation and predict that it will return to normal levels after 24 hours; while other models may predict that this is the beginning of a continuous rise and may reach 10μSv / h after 24 hours. By integrating these different viewpoints and considering the uncertainty of each model, the present invention can obtain more robust prediction results.

[0093] In another embodiment of the present invention, if all models show a high degree of uncertainty, for example, the standard deviation of the prediction exceeds 50% of the predicted mean, this may mean that an abnormal situation has been encountered, and the system will automatically trigger a more in-depth investigation process, such as increasing the monitoring frequency or starting additional monitoring equipment.

[0094] Next, the present invention describes an adaptive multi-task learning optimizer 5. The design of this optimizer takes into account the complexity of the nuclear radiation monitoring task and the need to optimize multiple objectives simultaneously. In one embodiment of the present invention, the mathematical expression of the optimizer is as follows:

[0095] L total =L pred +L physics +L uncertainty

[0096]

[0097] m t =β1m t-1 +(1-β1)g t

[0098]

[0099] η t =η base (1+cos(πt / T)) / 2

[0100]

[0101] Among them, L total is the total loss, L pred To predict the loss, L physics is the physical constraint loss, L uncertainty is the uncertainty loss, g t is the gradient, m t ,v t are the first-order and second-order moment estimates, β1, β2 are momentum parameters, η base is the basic learning rate, T is the period, θ t are model parameters.

[0102] In a preferred embodiment of the present invention, the momentum parameters β1 and β2 are set to 0.9 and 0.999 respectively. base Can be set to 0.001. The period T is usually set to 1 / 2 or 1 / 4 of the number of training rounds. This optimizer combines the adaptability of the Adam optimizer with the periodicity of the cosine annealing learning rate schedule.

[0103] In the context of nuclear radiation monitoring, this design has several advantages. First, it can adapt the gradient scales of different loss terms, so that prediction accuracy, physical consistency, and uncertainty estimation can be optimized in a balanced manner. Second, periodic learning rate adjustments help the model escape from local optimality, which is particularly useful when dealing with complex radiation patterns.

[0104] In another embodiment of the present invention, the weights of the various loss items are dynamically adjusted according to the specific situation. For example, during normal monitoring, the predicted loss L pred The weight of may be set to 0.6, and the physical constraint loss L physics The weight is 0.3, and the uncertainty loss L uncertainty The weight of L is 0.1. When an abnormal event occurs, such as the detected radiation level exceeds the normal value by 10 times, the system will automatically increase the weight of physical constraints and uncertainty estimates, and physics The weight is increased to 0.4, L uncertainty The weight of is increased to 0.2 to improve the reliability and robustness of the model.

[0105] Next, the present invention describes in detail the steps of fusion and enhancement processing. This process is crucial to improving the quality of nuclear radiation monitoring data. First, the present invention corrects the multi-source monitoring data by combining physical methods and statistical regression. The physical method may include energy spectrum correction based on the response function of the detector, or efficiency correction based on the geometric structure of the detector. Statistical regression can use historical data to establish a mapping relationship between detector readings and actual radiation intensity.

[0106] In one embodiment of the present invention, for a gamma radiation dose measuring instrument, the following correction formula is used:

[0107] D corrected =a·D measured +b·E+c

[0108] Among them, D corrected is the corrected dose rate, D measured is the raw reading of the measurement, E is the energy of the gamma ray, and a, b, and c are parameters obtained through regression analysis. Typically, the value of a is between 0.9 and 1.1, the value of b is between -0.01 and 0.01, and the value of c is between -0.1 and 0.1. These parameters need to be determined and updated through regular calibration experiments.

[0109] Next, the present invention obtains sensor attitude and baseline data. The sensor attitude data can help understand the orientation of the detector, which is important for accurately estimating the location of the radiation source. For example, if the detector is tilted more than 30 degrees, its measurement results may require additional correction. The baseline data represents the background radiation level under normal circumstances, which helps to identify abnormal situations. In a preferred embodiment of the present invention, the baseline data is obtained by calculating the moving average of the past 30 days and is updated once a day.

[0110] Finally, the present invention fuses the calibrated signal data with the sensor attitude and baseline data. This fusion process can use a data fusion algorithm, such as Kalman filtering or Bayesian fusion. In another embodiment of the present invention, an extended Kalman filter is used for data fusion, and its state vector includes the corrected radiation dose rate, the detector attitude angle, and the current background radiation level. In this way, the enhanced radiation monitoring signal obtained by the present invention not only accurately reflects the current radiation level, but also contains information about the spatial position and historical background.

[0111] In terms of nuclear accident severity assessment, the present invention is based on the accident classification standard of the International Atomic Energy Agency (IAEA) and combines the Monte Carlo simulation method to establish a dose-level mapping model. The IAEA standard includes 7 levels, from level 0 (deviation) to level 7 (major accident). The model of the present invention needs to convert the predicted radiation dose into the corresponding nuclear accident severity level.

[0112] In one embodiment of the present invention, a Monte Carlo method is used to simulate nuclear accidents of different severity levels to generate a large amount of radiation dose data. These data cover various situations from normal operation to serious accidents. By analyzing these simulated data, the present invention establishes a probability model to give the probability of belonging to each accident level when a certain radiation dose level is observed.

[0113] For example, in a preferred embodiment of the present invention, the following mapping relationship may be obtained:

[0114] When the observed radiation dose rate is lower than 0.1 μSv / h, there is a 99.9% probability that it belongs to the normal operation state (level 0).

[0115] When the observed radiation dose rate is between 10-100 μSv / h, there is an 80% probability that it is a Level 3 event (severe event).

[0116] When the observed radiation dose rate exceeds 1 Sv / h, there is a 95% probability that it is a Level 6 or 7 accident (serious accident or major accident).

[0117] This probabilistic mapping model enables the present invention to more flexibly assess the severity of a nuclear accident while also taking into account the uncertainty of the prediction.

[0118] Finally, the present invention describes an intelligent evaluation and early warning system for nuclear radiation dose spatial distribution that implements the above method. The system includes a data acquisition module 6, a signal processing module 7, a prediction module 8 and an evaluation and early warning module 9.

[0119] The data acquisition module 6 is responsible for collecting raw data from various radiation monitoring devices. In one embodiment of the present invention, this may include fixed monitoring stations, mobile monitoring vehicles, and even radiation detectors carried by drones. This module needs to consider the real-time and reliability of data transmission, and may use 5G or dedicated networks to ensure timely acquisition of data. For example, a fixed monitoring station may transmit data once a minute, while a mobile monitoring device may dynamically adjust the transmission frequency according to its moving speed, ranging from once per second to once per minute.

[0120] The signal processing module 7 receives the raw data and performs the above-mentioned fusion and enhancement processing. In another embodiment of the present invention, this module runs on an edge computing device to reduce data transmission delay and improve the response speed of the system. Specifically, an industrial-grade computer equipped with a GPU, such as the NVIDIA Jetson series, may be used to perform real-time signal processing tasks.

[0121] The prediction module 8 is the core of the system, which runs the deep probabilistic neural network model described in detail above in the present invention. In a preferred embodiment of the present invention, considering the complexity of the model, this module runs on a high-performance GPU server, such as a server equipped with an NVIDIA Tesla V100 GPU. In order to ensure the real-time prediction, the present invention may use model quantization and acceleration technologies, such as TensorRT, to optimize the reasoning speed of the model. Typically, for a nuclear facility covering an area of ​​100 square kilometers, the system of the present invention can generate a high-resolution (10 m × 10 m grid) prediction of the spatial distribution of radiation dose every 5 minutes.

[0122] The evaluation and early warning module 9 receives the prediction results, combines the nuclear accident severity level standard, and generates the final evaluation results and early warning information. In another embodiment of the present invention, this module needs to consider different early warning levels and early warning methods. For example, for minor anomalies (such as local radiation levels exceeding the background value by 2 times but less than 10 times), the system may only send a reminder message to the operator. For possible serious accidents (such as local radiation levels exceeding the background value by 100 times), the system may trigger an automatic emergency response program, send an emergency notice to the relevant departments, and start a pre-set evacuation plan.

[0123] In general, the system of the present invention realizes intelligent evaluation and early warning of nuclear radiation dose spatial distribution through advanced algorithms and comprehensive system design. It can not only provide accurate predictions, but also take into account the uncertainty of predictions and convert the results into actionable early warning information. This is of great significance for the safe operation of nuclear facilities and the timely response to potential nuclear accidents. The system of the present invention can significantly improve the efficiency and reliability of nuclear safety management and provide strong technical support for protecting the environment and public health.

[0124] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent evaluation and early warning method for nuclear radiation dose spatial distribution, characterized in that: The method comprises the following steps: Acquiring multi-source monitoring data, wherein the multi-source monitoring data includes gamma radiation dose, alpha radiation dose and beta radiation dose measuring instrument data; fusing and enhancing the multi-source monitoring data to obtain an enhanced radiation monitoring signal; Based on the enhanced radiation monitoring signal, a deep probabilistic neural network model is used to perform high-precision prediction of the spatial distribution of nuclear radiation dose; Based on the high-precision prediction results and combined with the nuclear accident severity level standards, radiation alert evaluation results and warning information are generated.

2. The method according to claim 1, characterized in that The deep probabilistic neural network model includes a multi-scale spatiotemporal feature extraction module, an adaptive attention mechanism, a physical constraint enhanced prediction module, an uncertainty-aware integrated learning module, and an adaptive multi-task learning optimizer. These modules process data in sequence as follows: The multi-scale spatiotemporal feature extraction module receives the enhanced radiation monitoring signal and outputs fused spatiotemporal features; The adaptive attention mechanism dynamically adjusts the importance of different spatiotemporal features based on the fused spatiotemporal features; The physical constraint enhanced prediction module uses the output of the adaptive attention mechanism to generate prediction results that conform to physical laws; The uncertainty-aware integrated learning module improves the robustness of prediction by integrating multiple models based on the output of the physical constraint-enhanced prediction module; The adaptive multi-task learning optimizer integrates the outputs of all the aforementioned modules to achieve adaptive optimization of model parameters.

3. The method according to claim 2, characterized in that The mathematical expression of the multi-scale spatiotemporal feature extraction module is: Among them, X t is the input feature at time t, is the hidden state of the i-th time scale, is the feature of the jth spatial scale, is the weight matrix, b i ,c j is the bias term, f,g is the activation function, F t It is the fused spatiotemporal features.

4. The method according to claim 2, characterized in that: The mathematical expression of the adaptive attention mechanism is: a t =softmax(e t ) h t =GRU(c t ,h t-1 ) Among them, F t is the output of the multi-scale spatiotemporal feature extraction module 1, h t-1 is the hidden state of the previous moment, v, W, b are learnable parameters, is the attention score of the kth feature, α t is the attention weight, c t is the context vector, h t is the hidden state at the current moment. GRU is a gated recurrent unit.

5. The method according to claim 2, characterized in that: The mathematical expression of the prediction module enhanced by physical constraints is: [m t ,s t ]=MLP(h t ) L physics =λ||Φ(y t )-Φ(y true )|| 2 Among them, h t is the output of adaptive attention mechanism 2, MLP is multi-layer perceptron, μ t ,σ t are the predicted mean and standard deviation, respectively. is the initial prediction result, y t is the prediction result after physical constraints, Φ is the physical constraint function, L physics is the physical constraint loss and λ is the trade-off factor.

6. The method according to claim 2, characterized in that The mathematical expression of the uncertainty-aware integrated learning module is: in, represents the i-th model, are the predicted value and uncertainty of the ith model, is the weight of the i-th model, β is the temperature parameter, are the prediction value and uncertainty after integration, respectively.

7. The method according to claim 2, characterized in that The mathematical expression of the adaptive multi-task learning optimizer is: L total =L pred +L physics +L uncertainty m t =β1m t-1 +(1-β1)g t or t =the base (1+cos(πt / T)) / 2 Among them, L total is the total loss, L pred To predict the loss, L physics is the physical constraint loss, L uncertainty is the uncertainty loss, g t is the gradient, m t ,v t are the first-order and second-order moment estimates, β1, β2 are momentum parameters, η base is the basic learning rate, T is the period, θ t are model parameters.

8. The method according to claim 1, characterized in that The fusion and enhancement process comprises the following steps: Correcting the multi-source monitoring data by combining physical methods and statistical regression to obtain calibrated signal data; Get sensor posture and baseline data; The calibrated signal data is fused with the sensor posture and baseline data to obtain the enhanced radiation monitoring signal.

9. The method according to claim 1, characterized in that: The nuclear accident severity level standard is based on the accident classification standard of the International Atomic Energy Agency (IAEA), and is combined with a Monte Carlo simulation method to establish a dose-level mapping model, which is used to convert the predicted radiation dose into the corresponding nuclear accident severity level.

10. An intelligent evaluation and early warning system for nuclear radiation dose spatial distribution implementing the method according to any one of claims 1 to 9, characterized in that: The system includes: Data acquisition module, used to acquire multi-source monitoring data; A signal processing module, used for fusing and enhancing the multi-source monitoring data; Prediction module, used to make high-precision predictions of the spatial distribution of nuclear radiation doses through a deep probabilistic neural network model; An evaluation and warning module is used to generate radiation warning evaluation results and warning information based on the high-precision prediction results and combined with the nuclear accident severity level standards; Among them, the prediction module includes a multi-scale spatiotemporal feature extraction unit, an adaptive attention mechanism unit, a physical constraint enhanced prediction unit, an uncertainty-aware integrated learning unit and an adaptive multi-task learning optimization unit, and these units perform data processing and model optimization in the manner described in claims 2-7.

Images