Nuclear power radiation environment monitoring and early warning instrument based on regularization space-time attention
By adopting a regularized space-time attention mechanism in the nuclear radiation environment monitoring and early warning instrument, the problem of insufficient accuracy and interpretability in the existing technology is solved, more efficient data processing and analysis is achieved, and the accuracy and real-timeness of early warning are improved.
Patent Information
- Application Number
- CN202510059458.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-15
- Publication Date
- 2025-05-30
AI Technical Summary
The existing nuclear radiation environment monitoring and early warning instruments have shortcomings in terms of accuracy and interpretability, making it difficult to effectively process and analyze nuclear radiation environment data, resulting in inaccurate warning results and poor real-time performance.
A nuclear power radiation environment monitoring and early warning instrument based on regularized space-time attention is used to describe the correlation between variables through the spatial attention module and the temporal attention module, and obtain sparse information through regularized losses, reduce false relationships and highlight relationships in the physical sense.
It improves the accuracy and real-timeness of nuclear radiation environment monitoring and early warning, and enhances the interpretability and credibility of early warning results.
Smart Images
Figure CN120069168A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing, and in particular, to a nuclear power radiation environment monitoring and warning instrument based on regularized spatio-temporal attention. Background Art
[0002] Nuclear radiation environment monitoring and warning is an important technology, which is widely used in the fields of nuclear energy industry, nuclear accident emergency management, environmental protection, etc. The monitoring and warning of the nuclear radiation environment are of great significance for protecting people's health and environmental safety. However, due to the complexity and uncertainty of the nuclear radiation environment, nuclear radiation environment monitoring and warning face many challenges. On the one hand, nuclear radiation environment data may be affected by external interference and changes, resulting in a decrease in the credibility of the data. These interferences and changes may be caused by various factors such as nuclear accidents, weather conditions, and environmental factors. For example, changes in wind direction and speed, unstable meteorological conditions, and the influence of the surrounding geological structure may all affect nuclear radiation environment data. On the other hand, nuclear radiation environment monitoring data may be incomplete and inaccurate, affecting the effectiveness and availability of the data. These problems may be caused by various factors such as the limitations of monitoring equipment, the difficulty of data collection, and errors in monitoring data processing.
[0003] At present, there have been some research works on nuclear radiation environment monitoring and warning, mainly including monitoring methods based on sensor data, prediction methods based on statistics and models, and warning instruments based on machine learning and artificial intelligence. However, these methods have some common limitations in nuclear radiation environment monitoring and warning: on the one hand, they rely on step-by-step association, and these associations do not involve the semantics with physical significance in the nuclear radiation environment monitoring data log, resulting in poor accuracy and interpretability; on the other hand, the attention matrix is often filled with false correlations, which cover up the physically meaningful correlations and further hinder effective interpretation. Therefore, there is an urgent need to develop a new type of nuclear power radiation environment monitoring and warning instrument to overcome the deficiencies of existing methods and improve the accuracy and real-time performance of warning. The instrument should have the ability to effectively process and analyze nuclear radiation environment data, be able to make full use of historical monitoring data and real-time monitoring data, extract key features and make accurate warning judgments. At the same time, the instrument should also have good adaptability and scalability, be able to cope with different environmental conditions and monitoring requirements, and provide reliable monitoring and warning services for the nuclear energy industry and environmental protection. Summary of the Invention
[0004] In order to overcome the deficiencies of the existing nuclear power radiation environment monitoring and warning instruments, such as poor accuracy and redundant relevant information, the purpose of the present invention is to provide a nuclear power radiation environment monitoring and warning instrument based on regularized spatio-temporal attention, which can make full use of the data laws of historical monitoring data and real-time monitoring data, reduce the transmission of redundant information, and achieve credible, accurate and timely warning results.
[0005] The technical solution adopted by the present invention to solve its technical problems is: a nuclear power radiation environment monitoring and warning instrument based on regularized spatio-temporal attention, including a database, a host computer and a warning instrument, which are connected in sequence; the database stores historical monitoring data of the nuclear radiation environment, and the data stored in the database is used by the host computer; the host computer includes a supervision and training module and a monitoring and warning module connected in sequence; the supervision and training module of the host computer uses the historical monitoring data of the nuclear radiation environment to train the monitoring and warning model, and then uploads the trained monitoring and warning model to the monitoring and warning module of the host computer; the warning instrument obtains nuclear environment monitoring data in real time and uploads it to the monitoring and warning module of the host computer, and the monitoring and warning module makes a warning judgment on the real-time monitoring data.
[0006] Further, the database is responsible for storing a large amount of historical monitoring data of the nuclear radiation environment. Given historical data X t =[x t+1 ,x t+2 ,...,x t+T , where x t+1 represents the monitoring data at the (t + 1)-th time step, T represents the length of the monitoring data time window, where D represents the dimension of the monitored data; is the radiation dose rate at the (t + T + H)-th time step, The goal is to learn a function such that
[0007] Further, the supervision and training module includes a neural network monitoring and warning module for predicting future radiation dose rates based on historical data and a regularized attention mechanism training module:
[0008] 1) Neural network monitoring and warning module, which uses the historical monitoring data of the nuclear radiation environment to predict future radiation dose rates. The specific process is as follows:
[0009] 1.1) The spatial attention module describes the correlation between variables by performing message passing on the spatial scale, and each variable is regarded as a node. The input of the spatial attention module is The output is where k represents the message passing in the k-th round, and in and out represent the input and output subscripts. First, generate the outgoing message embedding
[0010]
[0011] where is the outgoing message embedding of the i-th node, MLP represents the multi-layer perceptron, and the MLP layer contains a learnable affine transformation and a linear activation function. Then, define an attention matrix
[0012]
[0013] where is the generalized similarity measure between the i-th node and the j-th node; The division of avoids numerical errors in the optimization process. The SoftMax operation normalizes the attention scores of each node.
[0014] Next, aggregate the outgoing message embeddings by using the attention matrix to generate the incoming message embedding for each node
[0015]
[0016] where the incoming message embedding of the i-th node is the weighted sum of the outgoing message embeddings of its neighbor nodes, and the weights are the degrees of its attention to the neighbor nodes.
[0017] Finally, the state embedding of each node is updated according to its previous state embedding S k,in and the incoming message embedding as:
[0018]
[0019] The state embedding S k,in of the spatial attention module is initialized as S 1,in = Transpose(S 0 ) = Transpose(X t ).
[0020] 1.2) The time attention module extraction and the spatial attention module adopt the same principle. The inputs and outputs of the time attention module are and where k represents the message passing in the k-th round, and in and out represent the input and output subscripts. The attention matrix of the time attention module is defined as The state embedding T k,inis initialized to T k,in = Transpose(S k,out ). And because in the message passing of the k-th round, the spatial attention module and the temporal attention module are in a cascaded relationship, so in step 1.1), the state embedding S k,in where k≠1 is initialized to S k,in = Transpose(T k-1,out ).
[0021] 1.3) Input the output T K,out of the last temporal attention module into the gated recurrent unit and the multi-layer perceptron to obtain the predicted future radiation dose rate, where K represents the total number of rounds of information passing or the total number of spatio-temporal attention modules:
[0022] Z = GRU(T K,out ) (7)
[0023]
[0024] 2) Regularized attention mechanism training module, which trains the neural network monitoring and warning model in step 1). The specific process is as follows:
[0025] 2.1) In the training process, use regularized loss to efficiently obtain sparse information in the dense attention mechanism:
[0026]
[0027] where, ‖·‖ 1 represents the average absolute value of the matrix. The regularized loss strengthens the sparse weights in the training process. This regularization helps to eliminate spurious relationships and highlight physically meaningful relationships, ultimately improving the accuracy and interpretability of the monitoring.
[0028] 2.2) Given the input X t and the label the prediction error is:
[0029]
[0030] where, is the actual radiation dose rate at the t+T+H-th time step, is the predicted radiation dose rate at the t+T+H-th time step.
[0031] 2.3) The learning objective of the neural network is to minimize the prediction error and the message passing density. Combining the losses in step 2.1) and step 2.2) gives the final learning objective as:
[0032]
[0033] Among them, λ controls the strength of the regularization loss.
[0034] 3) Upload the trained neural network monitoring and warning model to the monitoring and warning module.
[0035] Furthermore, the warning device obtains nuclear environment monitoring data in real time and uploads it to the monitoring and warning module. The monitoring and warning module receives the trained neural network monitoring and warning model passed in from the supervised training module and uses this model to perform warning judgment on the real-time monitoring data. The judgment basis is whether there is data that does not exist in the historical normal data or whether it exceeds the threshold data.
[0036] Furthermore, the upper computer further includes a warning display module for displaying the warning judgment result of the monitoring and warning module on the real-time monitoring data.
[0037] The beneficial effects of the present invention are as follows: In response to the problem of nuclear radiation environment monitoring and warning, the present invention uses a spatial attention module to describe the correlation between variables by performing message passing on the spatial scale, uses a temporal attention module to simulate the progressive correlation relationship by performing message passing on the temporal scale, and adopts a regularization loss to efficiently obtain sparse information in the dense attention mechanism, thereby highlighting physically meaningful relationships, and finally realizing real-time intelligent monitoring and warning of the nuclear radiation environment; it has the following advantages: 1. High warning accuracy and strong real-time performance; 2. Strong interpretability and high credibility. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0039] Figure 1 is a schematic diagram of the hardware structure of the instrument proposed by the present invention;
[0040] Figure 2 is a schematic diagram of the functional modules of the upper computer proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0041] The following will describe the present invention in detail with reference to the drawings. Without conflict, the features in the following embodiments and implementation manners can be combined with each other.
[0042] Refer to Figure 1 , Figure 2, A nuclear power radiation environment monitoring and early warning instrument based on regularized spatio-temporal attention, including a database 1, a host computer 2, and an early warning instrument 3. The database 1, the host computer 2, and the early warning instrument 3 are connected in sequence. The database 1 stores historical monitoring data of the nuclear radiation environment and its corresponding radiation dose rate, and the data stored in the database is used by the host computer 2; the host computer 2 includes a supervision and training module 4, a monitoring and early warning module 5, and an early warning display module 6 connected in sequence; the supervision and training module 4 of the host computer 2 uses the historical monitoring data of the nuclear radiation environment and its corresponding radiation dose rate to train a monitoring and early warning model, and then uploads the trained monitoring and early warning model to the monitoring and early warning module 5 of the host computer; the early warning instrument 3 obtains nuclear environment monitoring data in real time and uploads it to the monitoring and early warning module 5 of the host computer. The monitoring and early warning module 5 makes an early warning judgment on the real-time monitoring data; finally, the early warning result is displayed on the early warning display module 6 of the host computer 2.
[0043] Specifically, the database 1 is responsible for storing a large amount of historical monitoring data of the nuclear radiation environment. Given historical data X t =[x t+1 ,x t+2 ,...,x t+T , where x t+1 represents the historical monitoring data at the (t + 1)-th time step, T represents the length of the historical monitoring data time window, where D represents the dimension of the historical monitoring data; is the radiation dose rate at the (t + T + H)-th time step, The goal is to learn a function such that
[0044] Specifically, the supervision and training module 4 includes a neural network monitoring and early warning module 7 for predicting future radiation dose rates based on historical data and a regularized attention mechanism training module 8:
[0045] 1) Neural network monitoring and early warning module 7. The neural network monitoring and early warning module 7 uses the historical monitoring data of the nuclear radiation environment to predict future radiation dose rates. The specific process is as follows:
[0046] 1.1) The typical self-attention method treats each time step as a node and performs message passing on the time scale. This method ignores the correlation between variables with physical significance, resulting in suboptimal accuracy and hindering interpretation in nuclear environment monitoring. To solve this problem, the spatial attention module describes the correlation between variables by performing message passing on the spatial scale, and each variable is regarded as a node. The input of the spatial attention module is The output is where k represents the message passing in the k-th round, and in and out represent the input and output subscripts. First, generate the outgoing message embedding
[0047]
[0048] where is the outgoing message embedding of the i-th node, MLP represents a multi-layer perceptron, and the MLP layer contains a learnable affine transformation and a linear activation function. Then, define an attention matrix
[0049]
[0050] where is the generalized similarity metric between the i-th node and the j-th node; The division of avoids numerical errors in the optimization process. The SoftMax operation normalizes the attention scores for each node.
[0051] Next, aggregate the outgoing message embeddings by using the attention matrix to generate the incoming message embedding for each node
[0052]
[0053] where the incoming message embedding of the i-th node is the weighted sum of the outgoing message embeddings of its neighbor nodes, and the weights are the degrees of its attention to the neighbor nodes.
[0054] Finally, the state embedding of each node is updated according to its previous state embedding S k,in and the incoming message embedding as follows:
[0055]
[0056] The state embedding S of the spatial attention module k,in is initialized as S 1,in = Transpose(S 0 ) = Transpose(X t ).
[0057] 1.2) The spatial attention module focuses on the correlation between variables while ignoring the temporal correlation. However, temporal correlation is crucial for capturing autoregressive features in monitoring logs. The temporal attention module extracts features using the same principle as the spatial attention module, but it relies on the correlation between time steps rather than variables. The temporal attention module simulates this step-by-step correlation by performing message passing on the time scale. The input and output of the temporal attention module are and where k represents the k-th round of message passing, and in and out represent the input and output subscripts. The attention matrix of the temporal attention module is defined as The state embedding T of the temporal attention module k,in is initialized as T k,in = Transpose(S k,out ). And because in the k-th round of message passing, the spatial attention module and the temporal attention module are in a cascaded relationship, the initialization of S k,in with k≠1 in the state embedding of the spatial attention module in step 1.1) is S k,in = Transpose(T k-1,out ).
[0058] 1.3) The output T K,out of the last temporal attention module is input into a gated recurrent unit and a multi-layer perceptron to obtain the predicted future radiation dose rate, where K represents the total number of rounds of information passing or the total number of spatio-temporal attention modules:
[0059] Z = GRU(T K,out ) (7)
[0060]
[0061] 2) The regularized attention mechanism training module 8 trains the neural network monitoring and warning model in step 1). The specific process is as follows:
[0062] 2.1) The associations learned by the self-attention model often have a high density, while the actual relationships in the physical world are sparse. This means the existence of spurious relationships, leading to overfitting and relationships with ambiguous physical meanings, ultimately compromising the accuracy and interpretability of monitoring. Therefore, a regularization loss is adopted during training to efficiently obtain sparse information in the dense attention mechanism:
[0063]
[0064] where, ‖·‖ 1Represents the average absolute value of the matrix. The regularization loss strengthens the sparse weights during the training process. This regularization helps to eliminate spurious relationships and highlight physically meaningful relationships, ultimately improving the accuracy and interpretability of the monitoring.
[0065] 2.2) Given the input X t and the label The prediction error is:
[0066]
[0067] where is the actual radiation dose rate at the (t + T + H)-th time step, is the predicted radiation dose rate at the (t + T + H)-th time step.
[0068] 2.3) The learning objective of the neural network is to minimize the prediction error and the message passing density. Combining the losses in steps 2.1) and 2.2), the final learning objective is:
[0069]
[0070] where λ controls the strength of the regularization loss.
[0071] 3) Upload the trained neural network monitoring and warning model to the said monitoring and warning module.
[0072] Specifically, the early warning device 3 obtains nuclear environment monitoring data in real time and uploads it to the monitoring and warning module 5. The monitoring and warning module 5 receives the trained neural network monitoring and warning model transmitted from the supervised training module 4 and uses this model to make early warning judgments on the real-time monitoring data. The judgment basis is whether there is data that does not exist in the historical normal data or whether it exceeds the threshold data; finally, the early warning result is displayed on the early warning display module 6.
[0073] The above embodiments are used to explain the present invention rather than limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.
Claims
1. A nuclear power radiation environment monitoring and early warning instrument based on regularized spatiotemporal attention, characterized by: It includes a database, a host computer and an early warning instrument; the database stores historical monitoring data of nuclear radiation environment and its corresponding radiation dose rate, and the data stored in the database is used by the host computer; the host computer includes a supervision training module and a monitoring and early warning module connected in sequence; the supervision training module of the host computer uses the historical monitoring data of nuclear radiation environment and its corresponding radiation dose rate to train a monitoring and early warning model, and then uploads the trained monitoring and early warning model to the monitoring and early warning module of the host computer; The early warning instrument acquires nuclear environment monitoring data in real time and uploads the data to the monitoring and early warning module of the host computer. The monitoring and early warning module performs early warning judgment on the real-time monitoring data.
2. The nuclear power radiation environment monitoring and early warning instrument based on regularized spatiotemporal attention as claimed in claim 1, characterized in that: Given historical monitoring data X t =[x t+1 ,x t+2 ,...,x t+T ], where x t+1 represents the historical monitoring data at the t+1th time step, T represents the length of the time window of the historical monitoring data, Where D represents the dimension of historical monitoring data; is the radiation dose rate at the t+T+Hth time step, H represents the time span between the predicted data and the historical monitoring data; The goal is to learn a function Make 3. The nuclear power radiation environment monitoring and early warning instrument based on regularized spatiotemporal attention as claimed in claim 2, characterized in that: The supervised training module includes a neural network monitoring and early warning module and a regularized attention mechanism training module: 1) A neural network monitoring and early warning module, which uses historical monitoring data of the nuclear radiation environment to predict future radiation dose rates; the neural network monitoring and early warning module includes a spatial attention module, a temporal attention module, a gated recurrent unit and a multi-layer perceptron; the specific process is as follows: 1.1) The spatial attention module describes the correlation between variables by performing message passing on the spatial scale. At each time step, the variable x t+1 ,x t+2 ,...,x t+T is regarded as a node; the input of the spatial attention module is The output is Where k represents the message passing of the kth round, in and out represent the input and output indices; first generate the outgoing message embedding in, Indicates that the outgoing message is embedded, The outgoing message of the ith node is embedded as MLP stands for multi-layer perceptron. The MLP layer contains a learnable affine transformation and a linear activation function. Then define an attention matrix in Indicates query, Represents a key value, represents a generalized similarity measure, The generalized similarity measure between the i-th node and the j-th node of is The division avoids numerical errors in the optimization process; the SoftMax operation normalizes the attention score of each node; Next, by using the attention matrix Aggregate the outgoing message embeddings to generate the incoming message embedding for each node Among them, the incoming message embedding of the i-th node is the weighted sum of the outgoing message embeddings of its neighbor nodes, and the weight is the degree of attention it pays to the neighbor nodes; Finally, the state embedding of each node is based on its previous state embedding S k,in and incoming messages embedded Updated to: The spatial attention module state embedding S k,in Initialized to S 1,in =Transpose(X t ); 1.2) The input and output of the temporal attention module are and Where k represents the message passing of the kth round, in and out represent the input and output indices; the attention matrix of the temporal attention module is defined as Temporal attention module state embedding T k,in Initialization is T k,in =Transpose(S k,out ); the spatial attention module and the temporal attention module are in a cascade relationship, and the state of the spatial attention module is embedded in S k,in The initialization of k≠1 is S k,in =Transpose(T k-1,out ); 1.3) The output T of the last temporal attention module K,out Input to the gated recurrent unit and multilayer perceptron to obtain the predicted future radiation dose rate Where K represents the total number of information transfer rounds or the total number of spatiotemporal attention modules: Z=GRU(T K,out ) (7) Z is the output of the gated recurrent unit; 2) A regularized attention mechanism training module, wherein the regularized attention mechanism training module trains the neural network monitoring and early warning model in step 1); the specific process is as follows: 2.1) Use regularization loss during training To capture sparse information in a dense attention mechanism: Among them, ‖·‖1 represents the mean absolute value of the matrix; 2.2) Given input X t and tags The prediction error is: in, is the actual radiation dose rate at the t+T+Hth time step, is the predicted radiation dose rate at the t+T+Hth time step; 2.3) The learning goal of the neural network is to minimize the prediction error and message passing density. Combining the losses in steps 2.1) and 2.2), the final learning goal is: Among them, λ controls the strength of the regularization loss; 3) Upload the trained neural network monitoring and early warning model to the monitoring and early warning module.
4. The nuclear power radiation environment monitoring and early warning instrument based on regularized spatiotemporal attention as claimed in claim 1, characterized in that: The early warning instrument acquires nuclear environment monitoring data in real time and uploads it to the monitoring and early warning module. The monitoring and early warning module receives the trained neural network monitoring and early warning model transmitted from the supervision training module, and uses the model to make early warning judgments on the real-time monitoring data. The judgment is based on whether data that does not exist in historical normal data appears or whether the data exceeds a threshold value.
5. The nuclear power radiation environment monitoring and early warning instrument based on regularized spatiotemporal attention as claimed in claim 1, characterized in that: The host computer also includes an early warning display module, which is used to display the early warning judgment results of the monitoring early warning module on the real-time monitoring data.