An unknown source term radiation field reconstruction method, storage medium and system

Abstract

The present application relates to a kind of unknown source term radiation field reconstruction method, comprising the steps of: obtaining input parameter and generating the radiation field data under different source term assumption;The generated radiation field data is preprocessed and data is augmented;Variational auto-encoding neural network model is constructed;Loss function is constructed to tune variational auto-encoding neural network model, and according to the variational auto-encoding neural network model after tuning, obtain gamma radiation field reconstruction result.The present application also provides a kind of storage medium and a kind of unknown source term radiation field reconstruction system, using the unknown source term radiation field reconstruction method, storage medium and system described in the present application can be in the case where existing measurement means and field condition do not support accurate measurement, realize gamma radiation field accurate reconstruction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of radiation field reconstruction, and particularly relates to an unknown source term radiation field reconstruction method, a storage medium and a system. BACKGROUND

[0002] Before the decommissioning of nuclear facilities, the radiation characteristics of the scene need to be investigated, and a complete database is extremely important for the development and optimization of subsequent decommissioning plans. However, due to the limitations of on-site measurement conditions and the shortcomings of existing measurement methods, there are often unknown or highly uncertain radiation source terms in the decommissioning scene, and the position information, nuclide type and proportion, distribution, and other data are necessary input parameters for radiation field and dose simulation calculation, which greatly affect the accurate radiation assessment of decommissioning operations.

[0003] Existing source term measurement methods, such as those based on high-purity germanium or NaI detectors, cannot accurately measure irregularly shaped and non-uniformly distributed source terms. At the same time, due to the limitations of on-site conditions, some source terms or radioactive contamination do not have the conditions for measurement, resulting in the lack of relevant data, which is an indispensable part of the radiation characteristic database investigation during the early stage of nuclear facility decommissioning. For example, the radiation source terms in the nuclear island room are often complex and non-uniformly distributed, and existing technical means cannot accurately measure them.

[0004] Therefore, a method is needed that can accurately reconstruct the gamma radiation field in the absence of accurate source term information, when traditional methods cannot calculate three-dimensional dose distribution, and existing measurement methods and on-site conditions do not support accurate measurement. SUMMARY

[0005] In view of the defects in the prior art, the purpose of the present application is to provide an unknown source term radiation field reconstruction method, a storage medium and a system to realize the accurate construction of gamma radiation field in the unknown source term scene.

[0006] To achieve the above purpose, the technical solution adopted by the present application is as follows: an unknown source term radiation field reconstruction method, comprising the steps of: obtaining input parameters and generating radiation field data under different source term assumptions; preprocessing and data augmentation of the generated radiation field data; constructing a variational auto-encoding neural network model; constructing a loss function to optimize the variational auto-encoding neural network model, and obtaining the gamma radiation field reconstruction result according to the optimized variational auto-encoding neural network model.

[0007] Further, the radiation field data includes the three-dimensional coordinates of the radiation source term, the proportion of different nuclides, the shape and size of the non-point source, and the non-uniformity of the source term distribution.

[0008] Further, after obtaining the input parameters, the assumed radiation field data under different combinations is simulated and calculated by point kernel integration and source term inversion algorithm.

[0009] Further, the generated radiation field data is preprocessed as normalization processing.

[0010] Further, the variational auto-encoding neural network model comprises a sampling layer, an encoding layer, a decoding layer and a hidden layer, the sampling layer is used for feature selection, the hidden layer is used for abstracting the selected features, the encoding layer is used for encoding high-dimensional input into a low-dimensional hidden vector, and the decoding layer is used for restoring the hidden vector to the initial dimension.

[0011] Further, the output of the encoding layer is two variables, which are respectively the mean and variance vectors of the hidden layer encoding, and sampling is performed from the inverse square distance distribution defined by the mean and variance vectors of the hidden layer encoding as the input of the decoding layer.

[0012] Further, the loss function comprises two parts of calculating the difference between the generated data and the original data sample and comparing the distribution of the hidden layer encoding vector with the inverse square distance from the source term.

[0013] Further, the calculation formula of the loss function is:

[0014]

[0015] Wherein, x represents the original data, y represents the generated data, and λ represents a weight parameter, The distribution of the hidden layer encoding vector z generated by the learning of the original data through the encoding layer is established, thereby establishing the relationship between z and x; p(z) is a quadratic decay distribution of the hidden layer encoding vector z distribution; E p(x) is the mathematical expectation value.

[0016] The application also provides a storage medium, wherein the storage medium stores a computer program, and the computer program is arranged to execute the unknown source term radiation field reconstruction method as described above when running.

[0017] The application also provides an unknown source term radiation field reconstruction system, comprising: a sample generation unit for generating initial samples for training according to input parameters; a data processing unit for preprocessing and data augmentation of the initial samples; a model training unit for training using the processed data set to form a variational auto-encoding neural network model; and a model optimization unit for constructing a loss function and optimizing the variational auto-encoding neural network model through the loss function.

[0018] The application has the effect that the prior knowledge of known source term information can be fully utilized to realize accurate reconstruction of the gamma radiation field under the premise that the three-dimensional radiation field distribution cannot be measured, and the measurement means and the on-site conditions do not support accurate measurement. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a step flow chart of a method for reconstructing a radiation field of an unknown source term according to the present application;

[0020] Figure 2 is a schematic diagram of a preliminary radiation field data generation module in a method for reconstructing a radiation field of an unknown source term according to the present application;

[0021] Figure 3 is a schematic diagram of an augmented flow after data preprocessing in a method for reconstructing a radiation field of an unknown source term according to the present application;

[0022] Figure 4 is a schematic diagram of a structure module of a variational auto-encoding neural network model;

[0023] Figure 5 is a schematic diagram of a module of a system for reconstructing a radiation field of an unknown source term according to the present application. DETAILED DESCRIPTION

[0024] The present application will be further described below in conjunction with the drawings and specific embodiments.

[0025] As shown in Figures 1-4 , the present application provides a method for reconstructing a radiation field of an unknown source term, which comprises the steps of:

[0026] S1, obtaining input parameters and generating preliminary radiation field data under different source terms;

[0027] Specifically, generating preliminary radiation field data under different source terms, i.e., the preliminary radiation field data under each source term, needs to be based on the geometric structure of the site, the source term investigation equipment and the working experience of the personnel, and a series of evaluations are made from the three-dimensional coordinates of each radiation source term, the proportion of different nuclides, the shape and size of non-point sources, the non-uniformity of source term distribution, etc., to form different combinations of input parameters. That is, the input parameters include the three-dimensional coordinates of the radiation source term, the proportion of different nuclides, the shape and size of non-point sources, and the non-uniformity of source term distribution.

[0028] It can be understood that, due to the limitations of measurement means and site conditions, the preliminary radiation field data can only be inaccurate and hypothetical data.

[0029] In an embodiment, the three-dimensional coordinates of the radiation source term can correspond to the geometric structure of the site, i.e., the radiation source term is located at a coordinate point of the site. The proportion of different nuclides, the shape and size of non-point sources, and the non-uniformity of source term distribution can correspond to the source term investigation equipment and / or the working experience of the personnel.

[0030] It can be understood that the radiation hot spots or source items of the region of interest of the nuclear facility site radiation protection personnel are relatively clear, and the position, distribution and number of the radiation hot spots or source items can be understood. The approximate parameter range can be provided. For the decommissioning scene containing multiple unknown or uncertain radiation source items, the source item investigation equipment is used, and the experience of the staff is used to maximize the use of the existing information.

[0031] After obtaining the above input parameters, the assumed radiation scene data under different combinations is simulated and calculated by means of point kernel integration and source item inversion algorithm, and a data set is formed.

[0032] It can be understood that the principle of the source item inversion algorithm is to use the measurement data of part of the radiation field, combine the three-dimensional geometric model and other known information of the source item, and first calculate the source item activity in reverse, and then use it as input to simulate the three-dimensional radiation field data of the place.

[0033] S2, preprocessing and data augmentation of the generated preliminary radiation field data;

[0034] Specifically, after generating the preliminary radiation field data and forming the data set, the radiation scene data needs to be preprocessed first to prepare for the subsequent steps.

[0035] In one embodiment, the radiation scene data is normalized, and the z-score standardization method is used to preprocess the sample data, for example:

[0036]

[0037] Wherein, u represents the original radiation scene data, x represents the average value, and σ represents the standard deviation of x.

[0038] Through normalization processing, the process of finding the optimal solution can be obviously smoothed, so that the optimal solution can be more easily converged, and the accuracy can be improved.

[0039] Data augmentation is a scheme to increase training data to alleviate the lack of data of machine learning algorithm. For the data set of numerical simulation, the amount of data input into the neural network training is limited. More data can make the neural network have better generalization performance, so it is necessary to increase the amount of data through data augmentation.

[0040] In one specific embodiment, when the size of the data batch is n, n test values are randomly selected from the entire initial data set. Then, n custom length segments are randomly divided from each selected test data to form a batch that meets the training size.

[0041] S3, constructing a variational auto-encoding neural network model;

[0042] Specifically, the data set formed in the above steps is used for training, so as to construct a variational auto-encoding neural network model, and the sampling layer, the encoding layer, the decoding layer and the hidden layer are optimized.

[0043] In one specific embodiment, the step is based on a variational auto-encoding neural network architecture, and the sampling layer, the encoding layer, the decoding layer and the hidden layer are designed respectively. The function of the sampling layer is to select features, the function of the hidden layer is to abstract the selected features, the encoding layer encodes the high-dimensional input into a low-dimensional hidden vector, and the decoding layer restores the hidden vector to the initial dimension.

[0044] Among them, the output of the encoding layer will be designed into two variables, which are the mean and variance vectors of the hidden layer coding z. Then, sampling is performed from the inverse square distance distribution defined by the two variables as the input of the decoding layer.

[0045] S4, constructing a loss function to optimize the variational auto-encoding neural network model, and obtaining the gamma radiation field reconstruction result according to the optimized variational auto-encoding neural network model;

[0046] Specifically, the variational auto-encoding neural network has two goals: one is to restore the original data, and the other is to make the coding vector of the hidden layer follow a specific distribution. Therefore, the loss function of the variational auto-encoding neural network needs to be divided into two parts: the first part is to calculate the difference between the generated data and the original data sample. This embodiment adopts the Mean Square Error (MSE) function, which refers to the expected value of the square difference between two variables. MSE can evaluate the difference between the original data and the data restored by the neural network. The smaller the value of MSE, the better the performance of the neural network in restoring the original data.

[0047] The second part is to compare the loss value between the distribution of the hidden layer coding vector and the inverse square distance from the source term, that is, the difference between the two distributions, which applies KL divergence. In order to make the distribution of the hidden layer coding vector z close to a certain distribution that needs to be followed, the key is to measure the difference between the two distributions q and p, so KL divergence is introduced to measure the difference between the two probability distributions. The closer the two distributions, the smaller the KL value, otherwise the larger. If there are two unknown distributions p(x) and q(x), the KL divergence can be obtained by the following equation:

[0048]

[0049] This is called relative entropy, that is, KL divergence, or KL divergence between p(x) and q(x).

[0050] Therefore, the final loss function of the variational auto-encoding neural network model is as follows:

[0051]

[0052] Wherein, x represents the original data, y represents the generated data, and lambda represents the weight parameter, The distribution of the hidden layer coding vector Z generated by the learning of the original data through the encoder, so as to establish the relationship between z and x, and p(z) represents the distribution of Z, which is a quadratic decay distribution in the above case, E p(x) Is the mathematical expectation value.

[0053] After the loss function of the variational auto-encoding neural network is constructed, the reconstruction result of the gamma radiation field can be obtained by using the variational auto-encoding neural network model.

[0054] The application also provides a storage medium, which stores a computer program, wherein the computer program is arranged to execute the above method steps when running. The storage medium can include, for example, floppy disks, optical disks, DVDs, hard disks, flash memories, U disks, CF cards, SD cards, MMC cards, SM cards, Memory Sticks, XD cards, etc.

[0055] The computer software product is stored in the storage medium, and includes a plurality of instructions for causing one or more computer devices (which can be personal computer devices, servers or other network devices, etc.) to execute all or part of the steps of the method of the application.

[0056] Please refer to Figure 5 The application also provides an unknown source term radiation field reconstruction system, which includes a sample generation unit 10, a data processing unit 20, a model training unit 30 and a model tuning unit 40.

[0057] The sample generation unit 10 generates initial samples for training according to input parameters.

[0058] That is, based on the geometry of the scene, the source term investigation equipment and the working experience of the personnel, a series of evaluations are performed from the three-dimensional coordinates of the radiation source term, the proportion of different nuclides, the shape and size of the non-point source, the non-uniformity of the source term distribution, etc., to form different combinations of input parameters, and then the input parameters are simulated and calculated by means of point kernel integration and source term inversion algorithm to simulate and calculate the assumed radiation scene data under different combinations, and form a data set.

[0059] The data processing unit 20 pre-processes and data augments the initial samples.

[0060] The model training unit 30 trains using the processed data set to form a variational auto-encoding neural network model.

[0061] The model tuning unit 40 constructs a loss function and tunes the variational auto-encoding neural network model through the loss function.

[0062] As can be seen from the above examples, in the unknown scene, the known approximate information sample is used to form a variational auto-encoding neural network model, after the model is optimized and the loss function is constructed for tuning, the gamma radiation field can be accurately reconstructed, the accuracy of the obtained gamma radiation field information is high, the requirements for the on-site conditions and the measurement means are relatively reduced, and the measurement cost of the radiation source term investigation in the early stage of the decommissioning of the nuclear facility is reduced.

[0063] The device described in the application is not limited to the examples described in the specific embodiments, and other embodiments can be derived by those skilled in the art according to the technical solutions of the application, which also belong to the technical innovation range of the application.

Claims

1. A method for reconstructing a radiation field of unknown source terms, characterized in that, The method comprises the following steps: acquiring input parameters, and simulating and calculating preliminary radiation field data under different source term assumptions by point kernel integration and source term inversion algorithm, wherein the radiation field data comprises three-dimensional coordinates of the radiation source term, different nuclide proportions, shapes and sizes of non-point sources, and non-uniformity of source term distribution; preprocessing and data augmentation are performed on the generated preliminary radiation field data, wherein the preprocessing specifically comprises normalization processing on the generated radiation field data; a variational auto-encoding neural network model is constructed, wherein the variational auto-encoding neural network model comprises a sampling layer, an encoding layer, a decoding layer and a hidden layer, the sampling layer is used for feature selection, the hidden layer is used for abstracting the selected features, the encoding layer is used for encoding high-dimensional input into a low-dimensional hidden vector, and the decoding layer is used for restoring the hidden vector to the initial dimension; the output of the encoding layer is two variables, which are the mean and variance vectors of the hidden layer encoding, and sampling is performed from the distance square inverse distribution defined by the mean and variance vectors of the hidden layer encoding as the input of the decoding layer; a loss function is constructed to optimize the variational auto-encoding neural network model, and a gamma radiation field reconstruction result is obtained according to the optimized variational auto-encoding neural network model.

2. The unknown source term radiation field reconstruction method according to claim 1, wherein: the loss function comprises two parts, which are the difference between the generated data and the original data samples, and the loss value of comparing the distribution of the hidden layer encoding vector with the distance square inverse ratio of the source term.

3. The unknown source term radiation field reconstruction method according to claim 1, wherein: the calculation formula of the loss function is: Wherein, x represents the original data, y represents the generated data, and λ represents the weight parameter, The distribution of the hidden layer coding vector z generated by the original data through the learning of the coding layer is represented, thereby establishing the relationship between z and x; p(z) refers to the quadratic decay distribution of the hidden layer coding vector z distribution; E p(x) is the mathematical expectation value.

4. A storage medium, comprising: The storage medium stores a computer program, wherein the computer program is configured to execute the unknown source term radiation field reconstruction method described in any one of claims 1-3 when running.

5. An unknown source term radiation field reconstruction system, comprising: The system executes the unknown source term radiation field reconstruction method described in any one of claims 1-3 when running, and the system comprises: a sample generation unit configured to generate initial training samples according to input parameters; a data processing unit configured to preprocess and data augment the initial samples; a model training unit configured to train the processed data set to form a variational auto-encoding neural network model; a model optimization unit configured to construct a loss function and optimize the variational auto-encoding neural network model through the loss function.

Images