Nuclear radiation index early warning system based on multi-algorithm fusion

Abstract

The invention discloses a nuclear radiation index early warning system based on multi-algorithm fusion, and the system comprises an environment data collection module which is used for collecting gamma-ray dose rate data, wind speed data, wind direction data and geographic position information, and constructing a nuclear radiation environment data set; the dose rate prediction module is used for predicting a future change trend of gamma ray dose rate data based on a sparse recurrent neural network model; the anomaly detection module is used for identifying an abnormal mutation point and positioning the position of a high-risk sensor; the diffusion region modeling module is used for calculating a nuclear radiation diffusion influence region; the grid risk evolution module is used for simulating time sequence evolution of a regular grid risk state based on the multi-state cellular automaton; and the fusion judgment module is used for fusing output results and generating a final early warning grade and a spatial risk distribution result. According to the invention, accurate prediction of the nuclear radiation risk and dynamic space early warning are realized, and the early warning accuracy and response efficiency are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear radiation index warning, and particularly to a nuclear radiation index warning system based on multi-algorithm fusion. Background Art

[0002] In the field of nuclear radiation safety prevention and control, how to accurately predict the changing trend of nuclear radiation dose rate and timely identify high-risk areas has always been an important research direction in environmental monitoring and emergency management. The existing nuclear radiation warning systems mainly rely on physical sensors at fixed sites to collect gamma-ray dose rate data, supplemented by a threshold judgment method based on statistical analysis to detect abnormal fluctuations in the monitoring data. However, such methods have poor sensitivity in abnormal identification and lack in-depth modeling of the time series change law, and are prone to missing sudden high-dose events or misreporting natural fluctuations as abnormal events. At the same time, traditional systems mostly adopt static geographical diffusion models and fail to fully consider meteorological factors, especially the actual influence of wind speed and wind direction on the nuclear radiation diffusion path and range, resulting in large deviations in spatial warning results and being difficult to support dynamic and real-time emergency response decisions.

[0003] In recent years, some studies have tried to introduce machine learning models, such as support vector machines, random forests or convolutional neural networks, for the prediction and abnormal discrimination of nuclear radiation dose rate, and certain accuracy improvements have been achieved. However, most of these models are black-box structures with poor interpretability and fail to fully utilize long-term dependence information when facing time series modeling, resulting in insufficient prediction stability. At the same time, most current methods directly use data abnormal points as risk points for simple spatial diffusion simulation, ignoring the quantification of abnormal degree and the propagation directionality characteristics, and there are potential risks of misjudging the direction and area in the diffusion simulation. Especially in the aspect of multi-source model collaborative judgment, the existing systems lack a unified fusion framework and are difficult to effectively integrate the outputs of different algorithms, resulting in one-sided or unstable overall warning results.

[0004] Aiming at the above deficiencies, the present invention proposes a nuclear radiation index warning system based on multi-algorithm fusion. By constructing a systematic processing flow including modules such as environmental data collection, dose rate prediction, abnormal detection, diffusion modeling, risk evolution and fusion determination, it solves problems such as inaccurate prediction, inaccurate diffusion modeling, isolated risk identification, and lack of fusion mechanism in the prior art. The sparse recurrent neural network combined with the time window attention mechanism is used to improve the prediction accuracy of gamma-ray dose rate time series data. The residual analysis is used to construct high-risk position data. Further, the exponential propagation algorithm corrected based on wind direction offset is used to achieve non-uniform spatial diffusion modeling. The multi-state cellular automaton is used to simulate the dynamic evolution of grid risks. Finally, a historical accuracy weighted strategy is introduced to fuse the outputs of each model to generate a nuclear radiation warning level result with higher timeliness and spatial accuracy, significantly improving the overall prediction accuracy and decision support ability of the system.

[0005] The proposed system not only overcomes the structural defects of traditional methods in data processing and model linkage, but also realizes the ability expansion from single-point prediction to the assessment of the dynamic risk distribution in the whole region. It has good promotion value and engineering adaptability, and is applicable to the nuclear radiation safety early warning tasks in nuclear power plants, the surrounding environment of nuclear facilities and urban emergency protection scenarios.

[0006] Therefore, how to provide a nuclear radiation index early warning system based on multi-algorithm fusion is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose a nuclear radiation index early warning method based on multi-algorithm fusion. The present invention fully combines a sparse recurrent neural network, an anomaly detection module, an exponential propagation algorithm with wind direction offset correction, and a multi-state cellular automaton, and details the whole process of nuclear radiation risk from data collection, trend prediction, anomaly identification, diffusion modeling to grid evolution and fusion early warning, with the advantages of high prediction accuracy, reasonable spatial distribution, fast response speed and adaptability to complex meteorological conditions.

[0008] The nuclear radiation index early warning system based on multi-algorithm fusion according to an embodiment of the present invention includes: An environmental data collection module: used to collect γ-ray dose rate data, wind speed data, wind direction data and geographical location information, and construct a nuclear radiation environment data set; A dose rate prediction module: used to predict the future change trend of γ-ray dose rate data based on a sparse recurrent neural network model; An anomaly detection module: used to identify anomaly mutation points and locate the positions of high-risk sensors; A diffusion area modeling module: used to calculate the nuclear radiation diffusion influence area; A grid risk evolution module: used to divide the monitoring area into regular grids and simulate the temporal evolution of the risk states of the regular grids based on a multi-state cellular automaton; A fusion determination module: used to generate the final early warning level and the spatial risk distribution result.

[0009] Optionally, the modules are implemented by the following method: S1. Collect γ-ray dose rate data, wind speed data, wind direction data and geographical location information in the monitoring area, and construct a nuclear radiation environment data set; S2. For the γ-ray dose rate data, use a sparse recurrent neural network model to perform trend prediction and generate a nuclear radiation dose rate prediction result matrix; S3. Input the nuclear radiation dose rate prediction result matrix and historical γ-ray dose rate data into the anomaly detection module to identify the anomaly mutation points of the nuclear radiation dose rate in the nuclear radiation dose rate prediction result matrix and generate high-risk position data; S4. Input the high-risk location data, wind speed data, and wind direction data into the improved exponential propagation algorithm to generate the data of the nuclear radiation diffusion influence area; S5. Divide the monitoring area into regular grids. Based on the multi-state cellular automaton model, according to the nuclear radiation diffusion influence area data and the high-risk location data, simulate the risk state evolution process of each regular grid, and output the grid risk level results; S6. Perform weighted fusion on the output results of the sparse recurrent neural network model, the anomaly detection module, the exponential propagation algorithm, and the multi-state cellular automaton model, dynamically adjust the weights, and generate the final warning level and the spatial risk distribution results.

[0010] Optionally, the S1 specifically includes: arranging fixed sensor nodes in the set monitoring area, obtaining the geographical location information of each sensor node, and in the sampling period, collecting the γ-ray dose rate data, wind speed data, and wind direction data in real time. Construct the γ-ray dose rate data, wind speed data, and wind direction data of each node at the same time point into a data matrix, perform time alignment, missing value filling, outlier removal, and normalization processing on the data matrix to obtain the standardized nuclear radiation environment data matrix, and combine it into a nuclear radiation environment data set in chronological order.

[0011] Optionally, the S2 specifically includes: S21. Based on the nuclear radiation environment data set, construct the standardized γ-ray dose rate time series data for each sensor node. Set the sensor node. For each sensor node, collect the standardized γ-ray dose rate data at each sampling time point, arrange them in chronological order, and construct the standardized γ-ray dose rate time series of the sensor node; S22. Set the sliding window length and prediction step. Perform sliding sampling on the standardized γ-ray dose rate time series of each sensor node to construct the sequence for model input. Introduce the time window attention mechanism into the sequence for model input. Based on the weight coefficients at each time step, perform weighted processing on the γ-ray dose rate values at different time points in the sequence for model input to generate the attention weighted input sequence for model training; S23. Based on the attention weighted input sequence, construct a sparse recurrent neural network model and improve it. Adopt the dynamic pruning strategy to control the sparsity and iterative update of the hidden state connection structure in the network to obtain the sparse connection state matrix for prediction, complete the model training, use the trained sparse recurrent neural network model to predict the γ-ray dose rate time series of each sensor node, obtain the γ-ray dose rate prediction results of all sensor nodes at multiple future time steps, and combine the prediction results of each sensor node at the corresponding time point in chronological order and node number to construct the nuclear radiation dose rate prediction result matrix.

[0012] Optionally, S3 specifically includes: S31. Based on the nuclear radiation dose rate prediction result matrix and the standardized γ-ray dose rate time series, for each sensor node , extract the prediction sequence from time step to from the nuclear radiation dose rate prediction result matrix, denoted as , extract the observed subsequence at the same time step from the standardized γ-ray dose rate time series, denoted as , and form an input sample pair with and ; S32. Input each input sample pair into the anomaly detection module, calculate the residual values at each time step between the predicted values of the prediction sequence and the observed values of the observed subsequence, and generate a residual sequence , where each residual value is defined as , represents the predicted time step index; S33. Set the anomaly detection threshold for each sensor node, judge each time step in the residual sequence. If the residual value at a certain time step is greater than the corresponding anomaly detection threshold, mark this time step as an anomaly point, record the node number and time point of the anomaly point, and form an anomaly point set; S34. Perform spatial clustering processing on the anomaly point set, use the density clustering algorithm to aggregate the anomaly points with close geographical locations to form a high-risk area, extract the sensor node numbers corresponding to the clustering centers, construct a high-risk sensor node set, and output high-risk location data.

[0013] Optionally, S4 specifically includes: S41. Based on the high-risk location data, set the high-risk sensor node set as , where represents the number of high-risk nodes, and each high-risk node includes the node geographical coordinates , the sampling time point , the γ-ray dose rate residual value , and obtain the wind speed data and the wind direction data at the corresponding time point; S42. Based on each high-risk node , construct an improved exponential propagation algorithm with wind direction offset correction, and calculate the diffusion influence value of the node on any grid point coordinate in the monitoring area; S43. Traverse each high-risk node , and calculate the diffusion influence value for all grid points within the influence range; S44. Superimpose the diffusion influence values of all high-risk nodes in the spatial dimension to obtain a superimposed matrix ; S45. Set a diffusion influence threshold , and determine all grid points that satisfy as diffusion influence area points, and extract this set to form nuclear radiation diffusion influence area data.

[0014] Optionally, the specific steps of S5 are as follows: S51. Divide the monitoring area into regular grids according to the geographical boundary information of the monitoring area, and define the set of regular grids as , where each grid cell corresponds to a unique central coordinate ; S52. Based on the superimposed matrix , classify the diffusion influence value corresponding to the central coordinate of each grid cell into risk levels. The diffusion influence threshold is , the gradient interval is , and construct an initial risk state vector , where: If , then set , indicating a low-risk state; If , then set , indicating a medium-risk state; If , then set , indicating a high-risk state; S53. Based on the initial state vector , establish a multi-state cellular automaton model, and for each evolution round , update the state value according to the following state update rules: If the current state , and the number of adjacent units with a state value of 2 is not less than the threshold , then update to ; If the current state , and the number of adjacent units with a state value of 2 is not less than the threshold , then update to ; If the current state , then keep it unchanged; In other cases, keep the original state, that is ; Among them, , is the adjacency state trigger threshold, preset as a constant parameter; S54. Set the maximum number of evolution rounds , for the state vector perform rounds of iteration to obtain the final state vector , where represents the final risk level of the grid cell ; S55. Bind the final state vector to each grid coordinate and output the grid risk level result including the geospatial distribution.

[0015] Optionally, the S6 specifically includes: collecting the nuclear radiation dose rate prediction result matrix, high-risk location data, nuclear radiation diffusion impact area data, and grid risk level results, uniformly summarizing the output results, dynamically assigning weights, and performing weighted fusion on each output result to obtain the comprehensive risk score of each grid cell. Divide the comprehensive score into low warning level, medium warning level, and high warning level, and combine the geographical location information to output the final warning level and spatial risk distribution result of the nuclear radiation index for warning display and response deployment.

[0016] The beneficial effects of the present invention are: First, the present invention introduces a sparse recurrent neural network model and combines it with a time window attention mechanism to achieve accurate prediction of γ-ray dose rate time series data, overcoming the problem of insufficient time dependence modeling in traditional methods and effectively improving the timeliness and accuracy of prediction.

[0017] Second, the present invention identifies high-risk locations by constructing an anomaly detection module and performs directional modeling on the nuclear radiation diffusion path by combining an exponential propagation algorithm with wind direction offset correction, which is more in line with the actual environmental diffusion process than the existing static diffusion methods and significantly improves the reliability of the spatial warning results.

[0018] In addition, the grid risk evolution simulation implemented by the present invention based on a multi-state cellular automaton, combined with a multi-model fusion decision mechanism, comprehensively considers the output results of each model and historical accuracy for dynamic weighting, effectively improving the comprehensiveness and stability of the nuclear radiation warning level judgment, and having stronger practicability and adaptability. Description of the Drawings

[0019] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is the method flow chart of the nuclear radiation index warning system based on multi-algorithm fusion proposed by the present invention; Figure 2 This is a schematic diagram of the module structure of the nuclear radiation index early warning system based on multi - algorithm fusion proposed by the present invention. Specific implementation manner

[0020] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0021] Reference Figure 1 and Figure 2 , the nuclear radiation index early warning system based on multi - algorithm fusion includes: Environmental data acquisition module: used to collect gamma - ray dose rate data, wind speed data, wind direction data and geographical location information, and construct a nuclear radiation environment data set; Dose rate prediction module: used to predict the future change trend of gamma - ray dose rate data based on a sparse recurrent neural network model; Abnormal detection module: used to identify abnormal mutation points and locate the positions of high - risk sensors; Diffusion area modeling module: used to calculate the nuclear radiation diffusion influence area; Grid risk evolution module: used to divide the monitoring area into regular grids and simulate the temporal evolution of the risk states of the regular grids based on a multi - state cellular automaton; Fusion decision - making module: used to generate the final early warning level and spatial risk distribution results.

[0022] The present invention realizes the data perception, intelligent modeling and collaborative analysis of the whole process of nuclear radiation early warning by constructing six major functional modules including an environmental data acquisition module, a dose rate prediction module, an abnormal detection module, a diffusion area modeling module, a grid risk evolution module and a fusion decision - making module. The data flow between modules is closed - loop, and the functional logics are complementary, effectively solving the problems of information isolation, one - sided modeling and lagging response in the prior art. The system has the advantages of clear structure, clear division of labor, high processing efficiency and strong scalability, significantly improving the intelligent level and practical performance of nuclear radiation monitoring and early warning.

[0023] In this implementation manner, the modules are realized through the following methods: S1. Collect gamma - ray dose rate data, wind speed data, wind direction data and geographical location information in the monitoring area, and construct a nuclear radiation environment data set; S2. For the gamma - ray dose rate data, use a sparse recurrent neural network model to perform trend prediction and generate a nuclear radiation dose rate prediction result matrix; S3. Input the nuclear radiation dose rate prediction result matrix and the historical gamma-ray dose rate data into the anomaly detection module to identify the abnormal mutation points of the nuclear radiation dose rate in the nuclear radiation dose rate prediction result matrix and generate high-risk location data; S4. Input the high-risk location data, wind speed data, and wind direction data into the improved exponential propagation algorithm to generate nuclear radiation diffusion impact area data; S5. Divide the monitoring area into regular grids. Based on the multi-state cellular automata model, according to the nuclear radiation diffusion impact area data and the high-risk location data, simulate the risk state evolution process of each regular grid and output the grid risk level result; S6. Perform weighted fusion on the output results of the sparse recurrent neural network model, the anomaly detection module, the exponential propagation algorithm, and the multi-state cellular automata model, dynamically adjust the weights, and generate the final warning level and spatial risk distribution results.

[0024] The nuclear radiation index warning method based on multi-algorithm fusion proposed by the present invention constructs a full-process technical system from multi-source data collection, time series prediction, anomaly detection, diffusion modeling, risk evolution to fusion determination, realizing the organic connection of the algorithm functions in each stage. Compared with the existing nuclear radiation warning methods that focus on single-point processing and lack model coordination, the present invention improves the data processing accuracy, the authenticity of diffusion modeling, and the reliability of warning results through modular design and dynamic fusion strategy, and has the advantages of clear structure, accurate prediction, and adaptability to complex environments, providing systematic technical support for constructing an efficient and intelligent nuclear radiation warning system.

[0025] In this embodiment, the S1 specifically includes: arranging fixed sensor nodes in the set monitoring area to obtain the geographical location information of each sensor node, and in the sampling period, collecting gamma-ray dose rate data, wind speed data, and wind direction data in real time. The gamma-ray dose rate data is measured by a radiation detector in real time and recorded at a second-level frequency. The wind speed data is obtained by an ultrasonic anemometer to obtain the wind speed magnitude, and the wind direction data is measured by a wind vane or an internal wind direction sensor to measure the current wind direction angle. The three types of data are collected synchronously and stored with time stamps. The gamma-ray dose rate data, wind speed data, and wind direction data of each node at the same time point are constructed into a data matrix. After aligning the data according to the time stamp, interpolation is used to fill in the missing values, the out-of-limit abnormal points are removed, and finally, the data is normalized to a unified scale by column to obtain a standardized nuclear radiation environment data matrix, which is combined into a nuclear radiation environment data set in chronological order.

[0026] The present invention ensures the continuity, integrity, and analyzability of data input by deploying fixed sensor nodes in the monitoring area to collect gamma-ray dose rate, wind speed, wind direction, and geographical location information, and constructing a structured and standardized nuclear radiation environment dataset. Compared with the traditional nuclear radiation monitoring method that only relies on a single physical quantity or non-standardized data, the present invention realizes the synchronous acquisition and standard preprocessing of multi-source monitoring data, provides a unified input format for subsequent prediction and analysis, and improves the data quality and scalability of the overall system.

[0027] In this embodiment, step S2 specifically includes: S21. Based on the nuclear radiation environment dataset, construct a standardized gamma-ray dose rate time series data for each sensor node. Set the sensor node. For each sensor node, collect the standardized gamma-ray dose rate data at each sampling time point, arrange them in chronological order, and construct the standardized gamma-ray dose rate time series of the sensor node. S22. Set the sliding window length and prediction step. Perform sliding sampling on the standardized gamma-ray dose rate time series of each sensor node to construct a sequence for model input. Introduce a time window attention mechanism into the sequence for model input. Based on the weight coefficients at each time step, perform weighted processing on the gamma-ray dose rate values at different time points in the sequence for model input to generate an attention-weighted input sequence for model training. S23. Based on the attention-weighted input sequence, construct a sparse recurrent neural network model. By introducing a connection sparsification mechanism, reduce the redundant connections between neurons, reduce the computational complexity while retaining the time series modeling ability, and make improvements. Adopt a dynamic pruning strategy to perform sparsity control and iterative update on the hidden state connection structure in the network. Specifically, in each round of training, evaluate the importance of each connection between hidden states according to the amplitude of the connection weight and the gradient sensitivity, dynamically prune the connections with smaller weights or low contribution degrees to improve the network sparsity, and at the same time retain a recoverable mechanism for some of the pruned connections. Dynamically adjust the connection structure in combination with the network error feedback in subsequent iterations to achieve the coexistence of connection addition and removal, thereby controlling the parameter scale while maintaining the model expression ability, improving the training efficiency and generalization ability, obtaining a sparse connection state matrix for prediction, completing the model training, using the trained sparse recurrent neural network model to predict the gamma-ray dose rate time series of each sensor node, obtaining the gamma-ray dose rate prediction results of all sensor nodes at multiple future time steps, and combining the prediction results of each sensor node at the corresponding time points according to the time order and node number to construct a nuclear radiation dose rate prediction result matrix.

[0028] In the process of predicting the gamma-ray dose rate, the present invention adopts a sparse recurrent neural network model with an attention mechanism, which effectively captures the long-term dependence information in the time series, and reduces the redundant connections of the model through dynamic pruning, improving the computational efficiency and generalization ability. Compared with the traditional LSTM or black-box neural network prediction methods, the present invention greatly reduces the model complexity while ensuring the prediction accuracy, and enhances the response ability to abnormal change trends, effectively supporting the early perception of sudden nuclear radiation events.

[0029] In this embodiment, the S3 specifically includes: S31. Based on the nuclear radiation dose rate prediction result matrix and the standardized gamma-ray dose rate time series, for each sensor node , extract the prediction sequence from time step to from the nuclear radiation dose rate prediction result matrix, denoted as , extract the observed subsequence of the same time step from the standardized gamma-ray dose rate time series, denoted as , and combine and to form an input sample pair; S32. Input each input sample pair into the anomaly detection module, calculate the residual values at each time step between the predicted values of the prediction sequence and the observed values of the observed subsequence, and generate a residual sequence , where each residual value is defined as , represents the predicted time step index; S33. Set the anomaly detection threshold for each sensor node, judge each time step in the residual sequence. If the residual value at a certain time step is greater than the corresponding anomaly detection threshold, mark this time step as an anomaly point, record the node number and time point of the anomaly point, and form an anomaly point set; S34. Perform spatial clustering processing on the anomaly point set, use the density clustering algorithm to aggregate the anomaly points with close geographical locations to form high-risk areas, extract the sensor node numbers corresponding to the clustering centers, construct a high-risk sensor node set, and output high-risk location data.

[0030] The present invention constructs a residual analysis mechanism, compares the prediction results with the historical observed values point by point to form an anomaly score sequence, and combines the spatial clustering method to identify high-risk nodes, effectively improving the sensitivity and positioning accuracy of anomaly detection. Compared with the existing anomaly detection methods that rely on fixed thresholds or single-point judgments, the present invention can comprehensively consider the time deviation and spatial distribution characteristics, accurately identify sudden or local dose rate abnormal change areas, and improve the accuracy and stability of early warning.

[0031] In this embodiment, the S4 specifically includes: S41. Based on the high-risk location data, set the high-risk sensor node set as , where represents the number of high-risk nodes, and each high-risk node includes the node's geographic coordinates , the sampling time point , the residual value of the gamma-ray dose rate , and obtain the wind speed data and the wind direction data at the corresponding time point; S42. Based on each high-risk node , construct an improved exponential propagation algorithm with wind direction offset correction, and calculate the diffusion influence value of the node on any grid point coordinate in the monitoring area, which is defined as: ; where represents the Euclidean distance between the grid point and the node , represents the direction angle of the grid point relative to the node, represents the wind direction data of the node at the time point , is the wind direction offset penalty coefficient, is the diffusion scale parameter, which is controlled by the wind speed , represents the exponential mapping function; The diffusion influence value formula dynamically corrects the radiation influence intensity based on the angle offset between the wind direction and the radiation source diffusion path. By introducing the wind direction offset angle to adjust the exponential decay coefficient, the diffusion path is enhanced in the downwind direction and attenuated in the upwind direction, so as to more realistically simulate the asymmetric propagation behavior of nuclear radiation under different meteorological conditions and improve the spatial accuracy and physical rationality of the diffusion area modeling; S43. Traverse each high-risk node , and calculate the diffusion influence value for all grid points within the influence range; S44. Superimpose the diffusion influence values of all high-risk nodes in the spatial dimension to obtain the superimposed matrix , that is: ; S45. Set the diffusion influence threshold , and determine all grid points that satisfy as the diffusion influence area points, and extract this set to form the nuclear radiation diffusion influence area data.

[0032] ​​​​​​The wind direction deviation correction index propagation algorithm proposed by the present invention comprehensively considers the intensity of the radiation source term, distance attenuation, and wind direction deviation angle, and more realistically restores the actual diffusion path of nuclear radiation under complex meteorological conditions through asymmetric diffusion modeling. Compared with the traditional isotropic diffusion model or static diffusion map, the present invention improves the spatial accuracy and directional expression ability of the diffusion area simulation, which helps to more accurately identify the possible affected areas and guide the response deployment in nuclear accident emergencies.

[0033] In this embodiment, step S5 specifically includes: S51. Divide the monitoring area into regular grids according to the geographical boundary information of the monitoring area, and define the set of regular grids as , where each grid cell corresponds to a unique central coordinate ; S52. According to the superposition matrix , classify the diffusion influence value corresponding to the central coordinate of each grid cell into risk levels. The diffusion influence threshold is , the gradient interval is , and construct the initial risk state vector , where: If , then set , indicating a low-risk state; If , then set , indicating a medium-risk state; If , then set , indicating a high-risk state; S53. Based on the initial state vector , establish a multi-state cellular automaton model, and for each evolution round , update the state value according to the following state update rules: If the current state , and the number of adjacent units with a state value of 2 is not less than the threshold , then update to ; If the current state , and the number of adjacent units with a state value of 2 is not less than the threshold , then update to ; If the current state , then keep it unchanged; In other cases, keep the original state, that is ; Among them, 、 is the adjacency state trigger threshold, preset as a constant parameter; S54. Set the maximum number of evolution rounds , for the state vector perform rounds of iteration to obtain the final state vector , where represents the final risk level of the grid cell ; S55. Bind the final state vector with each grid coordinate to output the grid risk level result including the geospatial distribution.

[0034] Through the construction of a multi-state cellular automaton model, the present invention iteratively evolves and simulates the risk states of grid cells, and combines the initial diffusion influence value and the neighborhood state propagation rule to realize the dynamic modeling of the risk evolution process over time. Compared with the existing technology that only makes static judgments on the current risk, the present invention can reflect the process characteristics of the risk spreading from point to surface, improve the coherence, expandability and predictability of the nuclear radiation risk area identification, and provide effective support for spatial continuous early warning.

[0035] In this embodiment, the S6 specifically includes: collecting the nuclear radiation dose rate prediction result matrix, high-risk location data, nuclear radiation diffusion influence area data and grid risk level results, uniformly summarizing the output results, dynamically assigning weights, and performing weighted fusion on each output result to obtain the comprehensive risk score of each grid cell. The comprehensive score is divided into low warning level, medium warning level and high warning level, specifically including: when the fused comprehensive score is less than or equal to the first threshold, it is determined as the low warning level, indicating that there is no significant abnormality in the area currently and the risk is controllable; when the fused comprehensive score is greater than the first threshold and less than or equal to the second threshold, it is determined as the medium warning level, indicating that there are potential risks in the area and need to be paid more attention; when the fused comprehensive score is greater than the second threshold, it is determined as the high warning level, indicating that there may be significant nuclear radiation abnormalities in the area and timely response and intervention are required. Combining with the geographical location information, the final warning level of the nuclear radiation index and the spatial risk distribution result are output for early warning display and response deployment. The specific process is as follows: the system maps the fused score result to the grid map of the monitoring area, and uses different colors to identify the low, medium and high warning levels to visually present the risk states of each area. Users can view the geographical locations, risk levels and their associated model score data of each risk point through the interface; at the same time, the system automatically pushes the high warning areas to the emergency management platform to link relevant units to execute the response mechanism, such as starting on-site investigation, adjusting the monitoring frequency or issuing personnel evacuation warnings. The whole process realizes the closed-loop control from risk identification to emergency response, improving the timeliness and pertinence of nuclear radiation event handling.

[0036] The present invention constructs a fusion determination algorithm to dynamically weight and fuse the dose rate prediction results, anomaly detection data, diffusion area information, and grid risk levels. The weights are automatically adjusted according to historical prediction accuracy, avoiding model bias and the instability of a single algorithm. Compared with the traditional method of using static weights or manual experience to assign weights in a multi-model combination, the present invention improves the scientificity and robustness of the warning level output, significantly enhancing the overall warning decision-making ability and practicality of the system.

[0037] Example 1: To verify the feasibility of the present invention during implementation, the present invention is applied to the radiation safety monitoring scenario around a nuclear facility. There are multiple fixed monitoring nodes in this area. The original radiation warning mechanism is mainly based on a single-point judgment mode with a set threshold, lacking the ability to predict data trends and the linkage judgment mechanism for multi-source information, resulting in problems such as recognition delay, high false alarm rate, and insufficient spatial response ability in emergency events. To address the above problems, a nuclear radiation index warning system based on multi-algorithm fusion proposed by the present invention is deployed to comprehensively model and judge nuclear radiation risks in real time.

[0038] The system first deploys 60 fixed sensor nodes in the monitoring area. Each node collects data such as γ-ray dose rate, wind speed, and wind direction every 5 minutes. The data is normalized, time-aligned, missing value filled, and anomaly removed through edge computing nodes, and a standardized nuclear radiation environment dataset is constructed. In the experimental area, the dose rate of a node gradually increases from a normal 0.12 μSv / h to 0.46 μSv / h within four consecutive hours, and then fluctuates and reaches 0.59 μSv / h. Based on this, the system uses a sparse recurrent neural network model for time series prediction and successfully predicts the rising trend of the risk that its dose rate will exceed 0.50 μSv / h two time steps (10 minutes) in advance.

[0039] Through the residual anomaly detection model, by comparing historical data with prediction results, the system accurately identifies abnormal mutation points and locates the specific node numbers. Within the same time window, the system identifies a total of 4 adjacent nodes with significant abnormal residuals (the mean residual reaches 0.27 μSv / h, and the standard deviation is 0.05), and it is judged as a high-risk aggregation area through a spatial clustering algorithm. Subsequently, the central node position of this area, wind speed (1.9 m / s), and wind direction (southeast by east) are input into the exponential diffusion model with wind direction offset correction. The system calculates that under the existing meteorological conditions, the radiation impact area expands in an elliptical shape in the northeast direction, with a maximum diffusion distance of approximately 980 meters and 143 affected grids.

[0040] During the risk evolution process, based on grid division and initial diffusion influence values, the system assigns initial states to all grid cells. Among them, the number of low-risk grids is 893, the number of medium-risk grids is 116, and the number of high-risk grids is 27. Subsequently, a multi-state cellular automaton model is used for risk state evolution, setting an adjacent trigger threshold to simulate the process of risk spreading from the center to the boundary. After three rounds of evolution, finally, the high-risk area expands outward by approximately 160 meters, and the changing trend of the risk level within the grid is highly consistent with the changing trend of the actual radiation value.

[0041] The fusion determination module comprehensively analyzes the dose rate prediction output, anomaly detection results, diffusion influence matrix, and grid evolution status, and automatically assigns weights according to the historical model accuracy to generate a fusion score result. Using 0.35 and 0.55 as the grading thresholds to divide the early warning levels, the system finally determines 43 grids as high early warning levels, 82 grids as medium early warning levels, and the rest as low early warning levels. Comparing this result with the subsequent manual inspection data, it is found that the coincidence rate between the high early warning area and the actually measured exceeded area reaches 92%, and the prediction accuracy of the medium early warning area reaches 87%, effectively reducing false alarms and missed alarms.

[0042] This invention demonstrates excellent trend prediction ability, anomaly response speed, and spatial diffusion simulation effect in this implementation scenario, solving problems such as response lag, rough spatial early warning, and simplified risk diffusion models in traditional monitoring modes. The overall system is superior to existing solutions in terms of data processing efficiency, risk positioning accuracy, and comprehensive early warning ability, and has good promotion adaptability and engineering practicality.

[0043] Table 1: Comparison Table of Nuclear Radiation Multi-Model Fusion Early Warning Results and Actual Observation Data

[0044] The above "Comparison Table of Nuclear Radiation Multi-Model Fusion Early Warning Results and Actual Observation Data" shows the comparison between the system's prediction ability of γ-ray dose rate, anomaly detection effect, diffusion influence modeling ability, and risk evolution results and the actual observation situation in different grid cells. It can be seen from the table that the early warning levels of the system are highly consistent with the actual observation levels in most grids. For example, grid cells such as G001, G017, and G082 are all determined as high early warning levels by the system, which is consistent with their actual observation results of exceeding the normal dose rate, verifying the accuracy of the system in high-risk identification.

[0045] In the medium-risk grids such as G005, G029, and G101, the predicted values, residual values, and diffusion influence values are at medium levels. The system identifies them as the medium warning level, which is also highly consistent with the actual situation, indicating that the weighted processing of the multi-model results by the fusion judgment algorithm has strong stability and hierarchical recognition capabilities. In the low-risk areas such as G045 and G120, all the system scores are low, and the final judgment result is the low warning level, without over-warning, reflecting the good false alarm control ability of the system.

[0046] Based on the comprehensive analysis of the data in the table, it can be seen that the multi-algorithm fusion mechanism described in the present invention has good risk identification capabilities in complex spaces. The fusion results can effectively fit the actual monitoring situation, and have high engineering practicability and promotion value.

[0047] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A nuclear radiation index early warning system based on multi-algorithm fusion, characterized in that, Including: Environmental data acquisition module: It is used to collect gamma-ray dose rate data, wind speed data, wind direction data and geographical location information, and construct a nuclear radiation environment dataset; Dose rate prediction module: It is used to predict the future change trend of gamma-ray dose rate data based on a sparse recurrent neural network model; Abnormal detection module: It is used to identify abnormal mutation points and locate the positions of high-risk sensors; Diffusion area modeling module: It is used to calculate the nuclear radiation diffusion influence area; Grid risk evolution module: It is used to divide the monitoring area into regular grids and simulate the temporal evolution of the risk states of the regular grids based on a multi-state cellular automaton; Fusion decision module: It is used to generate the final warning level and spatial risk distribution results.

2. The nuclear radiation index early warning system based on multi-algorithm fusion according to claim 1, characterized in that The modules are implemented through the following methods: S1. Collect gamma-ray dose rate data, wind speed data, wind direction data and geographical location information in the monitoring area, and construct a nuclear radiation environment dataset; S2. For the gamma-ray dose rate data, use a sparse recurrent neural network model for trend prediction to generate a nuclear radiation dose rate prediction result matrix; S3. Input the nuclear radiation dose rate prediction result matrix and historical gamma-ray dose rate data into the abnormal detection module to identify the abnormal mutation points of the nuclear radiation dose rate in the nuclear radiation dose rate prediction result matrix and generate high-risk position data; S4. Input the high-risk position data, wind speed data and wind direction data into an improved exponential propagation algorithm to generate nuclear radiation diffusion influence area data; S5. Divide the monitoring area into regular grids, and based on the multi-state cellular automaton model, according to the nuclear radiation diffusion influence area data and high-risk position data, simulate the risk state evolution process of each regular grid and output the grid risk level results; S6. Perform weighted fusion on the output results of the sparse recurrent neural network model, abnormal detection module, exponential propagation algorithm and multi-state cellular automaton model, dynamically adjust the weights, and generate the final warning level and spatial risk distribution results.

3. The nuclear radiation index early warning system based on multi-algorithm fusion according to claim 2, characterized in that, The specific content of the above S1 includes: Fix sensor nodes in the set monitoring area, obtain the geographical location information of each sensor node, and in the sampling period, collect gamma-ray dose rate data, wind speed data and wind direction data in real time. Construct the gamma-ray dose rate data, wind speed data and wind direction data of each node at the same time point into a data matrix, and perform time alignment, missing value filling, outlier removal and normalization processing on the data matrix to obtain a standardized nuclear radiation environment data matrix, and combine it into a nuclear radiation environment dataset in chronological order.

4. The nuclear radiation index early warning system based on multi-algorithm fusion according to claim 3, characterized in that The specific content of the above S2 includes: S21. Based on the nuclear radiation environment dataset, construct standardized gamma-ray dose rate time series data for each sensor node, set the sensor node, and for each sensor node, collect the standardized gamma-ray dose rate data at each sampling time point, arrange them in chronological order and construct the standardized gamma-ray dose rate time series of the sensor node; S22. Set the sliding window length and prediction step size, perform sliding sampling on the normalized gamma-ray dose rate time series of each sensor node to construct a sequence for model input, introduce a time window attention mechanism to the sequence for model input, and based on the weight coefficients at each time step, perform weighted processing on the gamma-ray dose rate values at different time points in the sequence for model input to generate an attention-weighted input sequence for model training; S23. Based on the attention-weighted input sequence, construct a sparse recurrent neural network model and improve it. Adopt a dynamic pruning strategy to control the sparsity and iterative update of the hidden state connection structure in the network to obtain a sparse connection state matrix for prediction, complete model training, use the trained sparse recurrent neural network model to predict the gamma-ray dose rate time series of each sensor node, obtain the gamma-ray dose rate prediction results of all sensor nodes at multiple future time steps, and combine the prediction results of each sensor node at the corresponding time points according to the time sequence and node number to construct a nuclear radiation dose rate prediction result matrix.

5. The nuclear radiation index early warning system based on multi-algorithm fusion according to claim 4, wherein, The specific content of S3 includes: S31. For each sensor node, based on the nuclear radiation dose rate prediction result matrix and the standardized γ-ray dose rate time series , extract the prediction sequence from time step to from the nuclear radiation dose rate prediction result matrix, denoted as , extract the observed subsequence of the same time step from the standardized γ-ray dose rate time series, denoted as , and form an input sample pair by combining and . S32. Input each input sample into the anomaly detection module, calculate the residual values at each time step between the predicted values of the prediction sequence and the observed values of the observed subsequence, and generate a residual sequence , where each residual value is defined as , represents the predicted time step index; S33. Set the anomaly detection threshold for each sensor node, judge each time step in the residual sequence. If the residual value at a certain time step is greater than the corresponding anomaly detection threshold, mark this time step as an anomaly point, record the node number and time point of the anomaly point, and form an anomaly point set; S34. Perform spatial clustering processing on the anomaly point set, use a density clustering algorithm to aggregate the anomaly points with close geographical locations to form high-risk areas, extract the sensor node numbers corresponding to the clustering centers, construct a high-risk sensor node set, and output high-risk location data.

6. The nuclear radiation index early warning system based on multi-algorithm fusion according to claim 5, characterized in that The specific content of S4 includes: S41. Based on the high-risk location data, set the high-risk sensor node set as , where represents the number of high-risk nodes, and each high-risk node includes the node geographical coordinates , the sampling time point , the γ-ray dose rate residual value , and obtain the wind speed data and the wind direction data at the corresponding time point; S42. Based on each high-risk node , construct an improved exponential propagation algorithm with wind direction offset correction, and calculate the diffusion influence value of the node on any grid point coordinates in the monitoring area; S43. Traverse each high-risk node , and calculate the diffusion influence value for all grid points within the influence range; S44. Superimpose the diffusion influence values of all high-risk nodes in the spatial dimension to obtain a superimposed matrix ; S45. Set the diffusion impact threshold , and determine all grid points that satisfy as the points in the diffusion impact area, and extract this set to form the data of the nuclear radiation diffusion impact area.

7. The nuclear radiation index early warning system based on multi-algorithm fusion according to claim 6, characterized in that, The specific content of S5 includes: S51. Divide the monitoring area into regular grids according to the geographical boundary information of the monitoring area, and define the set of regular grids as , where each grid cell corresponds to a unique central coordinate ; S52. According to the superposition matrix , for each grid cell , the central coordinate corresponding to the diffusion influence value is used to divide the risk level. The diffusion influence threshold is , the gradient interval is , and the initial risk state vector is constructed, where: If , then set , indicating a low-risk state; If , then set , indicating a medium-risk status; If , then set , indicating a high-risk state; S53. Based on the initial state vector , establish a multi-state cellular automaton model, and for each evolution round , update the state value according to the following state update rules: If the current state , and the number of adjacent units with a state value of 2 is not less than the threshold , then update to ; If the current state , and the number of adjacent units with a state value of 2 is not less than the threshold , then update to ; If the current state , then it remains unchanged; Keep other conditions in the original state, i.e., ; Among them, and are the adjacency state trigger thresholds, preset as constant parameters; S54. Set the maximum number of evolution rounds , for the state vector perform rounds of iteration to obtain the final state vector , where represents the final risk level of the grid cell ; S55. Bind the final state vector with each grid coordinate to output the grid risk level result containing the geospatial distribution.

8. The nuclear radiation index early warning system based on multi-algorithm fusion according to claim 7, characterized in that The specific content of S6 includes: Collect the nuclear radiation dose rate prediction result matrix, high-risk location data, nuclear radiation diffusion impact area data, and grid risk level results, uniformly summarize the output results, dynamically assign weights, perform weighted fusion on each output result to obtain the comprehensive risk score of each grid unit, divide the comprehensive score into low warning level, medium warning level, and high warning level, and combine the geographical location information to output the final warning level and spatial risk distribution result of the nuclear radiation index for warning display and response deployment.

Images