A rapid nuclear radiation field inversion method and inversion equipment based on artificial intelligence

Abstract

This invention discloses a rapid nuclear radiation field inversion method and equipment based on artificial intelligence. The method involves three-dimensional modeling of the possible operating environment and shielding structure; calculating radiation field distribution data using Monte Carlo particle transport or deterministic particle transport programs; constructing training and testing sets using the radiation field data to train a neural network, and testing it on the testing set; using a radiation measuring instrument for on-site detection, processing the on-site detection data, and further training the neural network model; and using the optimized neural network model to rapidly invert the radiation distribution of the entire area. This invention, combined with limited on-site radiation detection data, enables rapid inversion of the radiation field distribution of the entire operating area under conditions of incomplete source term information, providing a scientific means for optimizing operational plans and assessing personnel radiation dose during nuclear activities.

Description

Technical Field

[0001] This invention relates to the field of nuclear radiation field inversion technology, and in particular to a rapid nuclear radiation field inversion method and inversion equipment based on artificial intelligence. Background Technology

[0002] With the rapid development of nuclear energy and nuclear technology utilization, radiation protection has become increasingly important. To reduce the radiation dose to workers in radiation fields, the cumulative dose to personnel during operations can be assessed through methods such as pre-simulation calculations and radiation dosimeter detection, providing a reference for optimizing operational plans.

[0003] The calculation of radiation fields mainly employs the Monte Carlo method and deterministic methods, each with its own advantages and disadvantages. The Monte Carlo method obtains the entire radiation field distribution by randomly simulating particle transport processes, but it requires knowledge of the source information and is time-consuming. Deterministic methods include the point kernel integral method and the discrete ordinate method. The point kernel integral method is an analytical calculation method based on the accumulation factor, while the discrete ordinate method approximates the solution of the neutron transport equation by discretizing the variables in the neutron transport equation. Deterministic methods are computationally fast, but they lack the ability to handle complex scenarios and require relatively clear structural information about the radioactive source system. They are unsuitable for situations where the internal structure and source information of the radioactive source are unknown, and their errors are significant.

[0004] To address the need for rapidly constructing radiation field scenarios and to accurately and quickly assess the rationality of personnel operation plans and ensure personnel radiation safety, it is urgent to further study rapid radiation distribution inversion methods, providing a rapid scientific means for the design and optimization of personnel operation plans. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a rapid nuclear radiation field inversion method based on artificial intelligence. It combines artificial intelligence with nuclear simulation calculation programs, establishes a neural network model through preliminary simulation calculations, and optimizes the neural network model based on a small amount of data detected in the field, thereby achieving rapid inversion of radiation field data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, including:

[0007] A rapid inversion method for nuclear radiation fields based on artificial intelligence includes the following steps:

[0008] S1, Scene Modeling: Establishing a model for the application scenario of the nuclear radiation field to be inverted, including models of various facilities in the scenario and known source terms;

[0009] S2, simulation calculation, sets the source terms, and calculates radiation field data using the Monte Carlo particle transport method or the deterministic particle transport method;

[0010] S3 preprocesses the radiation field data and divides it into training and testing sets;

[0011] S4. Establish a neural network and train it using a training set; the input data of the neural network is the location information of the measurement point, and the output data is the predicted neutron / photon flux of that measurement point.

[0012] S5. Test the neural network using the test set to verify the accuracy of the model. If the accuracy does not meet the requirements, modify the hyperparameters of the neural network and repeat step S4 until the accuracy meets the requirements or the number of iterations reaches the set number. Generate and save the neural network model.

[0013] S6. Optimize the neural network model generated in step S5. Use a detection device to perform on-site detection to obtain the neutron / photon flux data at the detection location. Compare this data with the predicted neutron / photon flux at the same location output by the neural network model. If the error between the predicted and detected neutron / photon fluxes is within acceptable limits, the neural network model is used directly. If the error exceeds acceptable limits, the detected neutron / photon flux at the location is preprocessed and added to the training set. The neural network model generated in step S5 is then further trained using this training set, and the model is updated and saved.

[0014] S7 uses the optimized neural network model to invert the radiation field data.

[0015] Preferably, in step S1, the data required for scene modeling includes:

[0016] Geometric data, including the dimensions and locations of various facilities in the scene;

[0017] Material data, including the material composition, material density, nuclide composition, and nuclide cross-section database of each facility in the scenario;

[0018] Source term data, including the shape of the radioactive source, particle type, energy, direction, and source term intensity.

[0019] Preferably, step S2 is performed as follows: the source term intensity is set, the entire radiation field is gridded, the neutron / photon flux of each grid is counted, and the radiation field distribution data is calculated using a Monte Carlo particle transport program or deterministic particle transport software.

[0020] Preferably, in step S3, the radiation field data is preprocessed, as shown below:

[0021] First, take the logarithm of the radiation field data to the base 10, and then normalize the logarithmic radiation field data.

[0022] Preferably, in step S6, the allowable error range between the predicted neutron / photon flux and the detected neutron / photon flux is within 1%.

[0023] Preferably, in step S5, the neural network hyperparameters include: the number of hidden layers, the number of nodes, and the activation function.

[0024] Preferably, the neural network is specifically as follows:

[0025] The neural network has a hidden layer with b nodes. An optimization algorithm is used to find the optimal neural network topology to minimize the loss function of the training set.

[0026] The activation functions for both the input and hidden layers use ReLU functions;

[0027] The optimization algorithm uses Adam, an extension of the stochastic gradient descent algorithm.

[0028] The mean squared error (MSE) is used as the loss function for the training set.

[0029]

[0030] Among them, y i To obtain the neutron / photon flux at position i in the training set, The predicted neutron / photon flux at the i-th position in the output of the neural network model; n represents the amount of data in the training set.

[0031] The mean percentage error (MAPE) was used as the final evaluation metric for the test set to verify model accuracy.

[0032]

[0033] Among them, y j To measure the neutron / photon flux at position j in the test set, The predicted neutron / photon flux at the j-th position in the output of the neural network model; m represents the amount of data in the test set.

[0034] Preferably, the neural network has two hidden layers, each with 50 nodes.

[0035] The present invention also provides an inversion device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the rapid nuclear radiation field inversion method of steps S1 to S7.

[0036] In step S6, the data and location information of the on-site detection equipment are sent to the inversion equipment.

[0037] Preferably, the inversion device includes a display for displaying the inverted radiation field data.

[0038] The advantages of this invention are:

[0039] (1) By combining artificial intelligence with nuclear simulation calculation programs, and leveraging the advantages of precise nuclear simulation calculations and the strong learning capabilities of artificial intelligence, a neural network model is established through preliminary simulation training. In the later stages, the radiation field can be quickly inverted based on a small amount of data detected on-site. This can provide a reference for the design and optimization of radiation operation plans, enabling "dosage-aware" operations, training operators to be familiar with radiation distribution, and rationally arranging operation time, which is of great significance for ensuring personnel radiation safety.

[0040] (2) The method of the present invention combines a small amount of on-site radiation detection data to quickly invert the radiation field distribution of the entire operation area under the condition of incomplete source term information, providing a scientific means for optimizing operation plans and assessing personnel radiation dose during nuclear activities. Attached Figure Description

[0041] Figure 1 This is a flowchart of a rapid nuclear radiation field inversion method based on artificial intelligence according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] In radiation field calculations, source term information is a prerequisite, serving as input. However, in practical applications, source term information is often incomplete. Other information, however, is usually familiar to field radiation technicians. This includes measurements of the radiation field dose rate, including dose rate readings and the coordinates of each measurement location; the locations of possible sources, including the energy, shape, number, and coordinates of the source terms; and the location and parameters of the shielding structure, such as its position, orientation, thickness, and material structure in the area of ​​interest. Providing this information allows for the inversion of the radiation field across the entire area.

[0044] Example 1

[0045] Depend on Figure 1 As shown, the present invention provides a method for rapid inversion of nuclear radiation fields based on artificial intelligence, which specifically includes the following steps:

[0046] S1, Scene Modeling: Establishing a model for the application scenario of the nuclear radiation field to be inverted, including models of various facilities in the scenario and known source terms.

[0047] In step S1, the data required for scene modeling includes:

[0048] Geometric data, including the dimensions and locations of all facilities in the scene;

[0049] Materials data, including the material composition of each facility, material density, nuclide composition, nuclide cross-section database, etc.;

[0050] Source term data includes the shape, particle type, energy, direction, and source strength of a potential radioactive source.

[0051] In this embodiment, the scene modeling described in step S1 specifically requires the following data:

[0052] The entire radiation field measures 10m × 10m × 3m. At the center is a neutron source with an energy of 2 MeV, which flies in an isotropic direction. The outer perimeter is surrounded by a concrete cylinder with an outer diameter of 2m, an inner diameter of 1.9m, and a height of 3m.

[0053] S2, simulation calculation: The source term intensity is set, and the entire radiation field is gridded with a scale of 0.1m × 0.1m × 0.1m, resulting in a total of 3000 cubic grids. The neutron flux of each grid is calculated. The radiation field data is calculated using Monte Carlo particle transport software or deterministic particle transport software.

[0054] S3 preprocesses the radiation field distribution data and splits the data into training and testing sets in an 8:2 ratio.

[0055] In step S3, the preprocessing method is as follows:

[0056] Because the radiation dose distribution varies greatly throughout the radiation field, with the maximum and minimum values ​​differing by several orders of magnitude, this data distribution presents significant challenges for neural network learning. Therefore, we first take the logarithm (base 10) of the radiation field distribution data, shrinking the data distribution to between -20 and 0, while also making the dose changes between adjacent points smoother. Then, we normalize the logarithmized radiation field distribution data. Normalization scales the data changes at different points with varying source intensity uniformly within the range [0,1], thus accelerating network convergence.

[0057] S4. Build and train a neural network. The input data of the neural network is the location information of the test point, and the output data is the predicted neutron flux of the test point. Select the hyperparameters of the neural network, including the number of hidden layers, the number of nodes, the activation function, etc., and train the neural network on the training set.

[0058] In step S4, the neural network is established as follows:

[0059] The neural network takes as input the spatial coordinates of a location in the training set and outputs the predicted flux at that location. An exhaustive optimization algorithm is used to build the neural network with two hidden layers, each containing 50 nodes. The goal is to find the optimal neural network topology that minimizes the loss function of the training set and ensures easy convergence.

[0060] The input and hidden layers use ReLU (Rectified linear unit) activation functions, g(z) = {0, z}. This non-linear activation function enables the neural network to fit non-linear functions.

[0061] The optimization algorithm uses Adam, an extension of the stochastic gradient descent algorithm, with an optimization step size of lr = 0.001.

[0062] The mean squared error (MSE) is used as the loss function for the training set.

[0063]

[0064] Among them, y i To obtain the flux at position i in the training set, is the predicted flux at the i-th position of the neural network model output; n represents the amount of data in the training set.

[0065] The mean percentage error (MAPE) was used as the final evaluation metric for the test set to verify the model's accuracy.

[0066]

[0067] Among them, y j To test the flux at position j in the set, is the predicted flux at the j-th position of the neural network model output; m represents the amount of data in the test set.

[0068] Since the prediction results of neural networks have a certain degree of randomness, each experimental data point is repeated 10 times and the average value is taken as the final prediction result of the neural network.

[0069] In this embodiment, the selected hyperparameters are as follows:

[0070] There are two hidden layers, each containing 50 nodes. The ReLU activation function is used for both the input and hidden layers.

[0071] S5. Test and adjust the neural network. Use the test set to test the neural network and check the accuracy of the model. If the accuracy does not meet the requirements, modify the hyperparameters and repeat step S4 until the accuracy meets the requirements or the number of iterations reaches the set number. Generate and save the neural network model.

[0072] S6. Optimize the neural network model generated in step S5.

[0073] The detection is carried out using neutron and gamma detection equipment. The data and location information of the on-site detection are sent to the intelligent radiation field inversion equipment and compared with the data field inverted by the neural network. If the error is within the allowable range, in this embodiment, if the error is less than 1%, the neural network model is used. If the error is large and exceeds the allowable range, that is, the error is greater than or equal to 1%, the detection data is used to further train the neural network, update and save the neural network model.

[0074] S7, Intelligent Inversion and Display, uses an optimized neural network model to invert radiation field data and displays it on an intelligent radiation field inversion device.

[0075] The method of this invention combines a small amount of on-site radiation detection data to rapidly invert the radiation field distribution of the entire work area under the condition of incomplete source term information, providing a scientific means for optimizing work plans and assessing personnel radiation dose during nuclear activities.

[0076] Example 2

[0077] An inversion device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the rapid nuclear radiation field inversion method in steps S1 to S7. In step S6, data and location information from the field detection device are sent to the inversion device. The inversion device also includes a display for showing the inverted radiation field data.

[0078] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based rapid inversion method for a nuclear radiation field, characterized in that, Comprise the following steps: S1, scene modeling, the application scene of the to be inverted nuclear radiation field is modeled, including the model of each facility in the scene, the known source term; S2, simulation calculation, setting the source term, calculating the radiation field data by using the Monte Carlo particle transport method or the deterministic particle transport method; S3, preprocessing the radiation field data, and dividing the radiation field data into a training set and a test set; S4, establishing a neural network and training the neural network using the training set; the input data of the neural network is the position information of the measuring point, and the output data is the predicted neutron / photon flux of the measuring point; S5, testing the neural network using the test set, verifying the model accuracy, if the accuracy does not meet the requirements, modifying the hyperparameters of the neural network, repeating step S4 until the accuracy meets the requirements or the number of iterations reaches the set number, generating and saving the neural network model; S6, optimizing the neural network model generated in step S5, detecting the detection position using a detection device to obtain detection neutron / photon flux data of the detection position, comparing the detection neutron / photon flux data of the position with the predicted neutron / photon flux of the position output by the neural network model, if the error between the predicted neutron / photon flux and the detection neutron / photon flux of the position is within the allowable range, the neural network model is directly used; if the error exceeds the allowable range, the detection neutron / photon flux of the position is preprocessed and added to the training set, and the training set is further trained based on the neural network model generated in step S5 to update and save the neural network model; S7, using the optimized neural network model to invert the radiation field data; The neural network is specifically as follows: The neural network has a hidden layer, each hidden layer has b nodes, and the optimal neural network topology structure is found by an optimization algorithm to minimize the loss function of the training set; The activation function of the input layer and the hidden layer uses the Relu function; The optimization algorithm uses the extended method Adam of the stochastic gradient descent algorithm; The mean square error MSE is used as the loss function of the training set: where y i is the neutron / photon flux at the i-th position in the training set, is the predicted neutron / photon flux at the i-th position output by the neural network model; n represents the amount of data in the training set; The average percentage error MAPE is used as the final evaluation index of the test set, which is used to verify the model accuracy: where y j is the neutron / photon flux at the jth position in the test set, is the predicted neutron / photon flux at the jth position output by the neural network model; m represents the amount of data in the test set.

2. The method of claim 1, wherein the method is based on artificial intelligence. In step S1, the data required for scene modeling includes: Geometric data, including the size and position of each facility in the scene; Material data, including the material composition, material density, nuclide share, and nuclide cross section database of each facility in the scene; Source term data, including the shape, particle type, energy, direction, and source term intensity of the radioactive source.

3. The method of claim 1, wherein the method is based on artificial intelligence. The specific way of step S2 is as follows: setting the source term intensity, drawing a grid for the entire radiation field, and counting the neutron / photon flux of each grid, and calculating the radiation field distribution data by using the Monte Carlo particle transport program or the deterministic particle transport software.

4. The method of claim 1, wherein the method is based on artificial intelligence. In step S3, the radiation field data is preprocessed, and the specific way is as follows: First, take the logarithm of the radiation field data with base 10, and then normalize the radiation field data after taking the logarithm.

5. The method of claim 1, wherein the method is based on artificial intelligence. In step S6, the allowed range of errors of both the predicted neutron / photon flux and the detected neutron / photon flux is within 1%.

6. The method of claim 1, wherein the method is based on artificial intelligence. In step S5, the neural network hyperparameters include: the number of hidden layers, the number of nodes, and the activation function.

7. The method of claim 6, wherein the method is based on artificial intelligence. Two hidden layers are established in the neural network, and each hidden layer has 50 nodes.

8. An inversion device suitable for use in a method of rapid inversion of a nuclear radiation field based on artificial intelligence according to any one of claims 1 to 7, characterized in that, The inversion device includes a processor, a memory, and a computer program stored on the memory and executable on the processor, and the processor implements the fast inversion method of the nuclear radiation field of steps S1-S7 when executing the computer program. In step S6, the data information and the position information of the field detection device are sent to the inversion device.

9. The inversion device of claim 8, wherein, The inversion device includes a display for displaying the inverted radiation field data.

Images