A three-dimensional imaging method and system for reactor neutron radiation

Abstract

The present invention discloses a three-dimensional imaging method and system for reactor neutron radiation, and relates to the field of three-dimensional imaging of reactor radiation fields. The method comprises: obtaining off-core detector data of a detection area, inputting a trained reconstruction network model including a mapping relationship between off-core detector data and neutron radiation distribution inside the core for prediction, obtaining a predicted neutron radiation distribution inside the core, and performing a visualized three-dimensional display. The scheme arranges the radiation distribution inside and outside the core according to an image through a reconstruction network model, forms a mapping relationship between off-field detection data and neutron radiation distribution inside the core, and can accurately and efficiently reconstruct the neutron distribution inside the core based on the off-core detector data of the detection area.

Description

Technical Field

[0001] The invention relates to the field of three-dimensional imaging of reactor radiation fields, and in particular to a three-dimensional imaging method and system for reactor neutron radiation. Background Art

[0002] The monitoring of neutron distribution is closely related to the safety and economy of the reactor. At present, during the operation of nuclear power plants, mobile micro-ionization chambers and self-powered detectors placed in the reactor for a long time are usually used to measure the neutron radiation field. However, under accident conditions, the in-core detectors cannot operate normally, and they mainly rely on atmospheric monitoring devices to diagnose or predict the spatial and temporal distribution of released radioactive substances in the environmental medium. It is impossible to detect changes in the core state and formulate a reasonable core treatment strategy.

[0003] For example, in a nuclear accident, after the core melts, if the changes in the core neutron radiation field can be reconstructed in real time, corresponding measures can be formulated in advance to avoid subsequent nuclear disasters. In future advanced reactors, the ionization chambers and self-powered detectors in the reactor will face harsh service environments and will be difficult to deploy in large quantities (high temperature, high irradiation, corrosion, strong magnetic fields, and narrow spaces). For example, the core temperature of a lead-based reactor is 400°C, and that of a high-temperature gas-cooled reactor can reach 1000°C. Lead-based reactors and molten salt reactors also face serious corrosion problems, and fusion reactors will also face a strong magnetic field service environment; the core of a miniaturized nuclear energy system is more compact, and if space is reserved for in-core detectors, the economic efficiency of the reactor will be seriously affected. In addition, in the design and commissioning of new reactor types, it is necessary to continuously monitor parameters such as the neutron flux distribution in the reactor. At this time, the architectural parameters of the in-core design are uncertain. If in-core detectors are used, the detector layout needs to be continuously adjusted, wasting manpower and material resources. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a three-dimensional imaging method and system for reactor neutron radiation in view of the deficiencies in the prior art.

[0005] The technical solution of the present invention to solve the above technical problems is as follows:

[0006] A three-dimensional imaging method of reactor neutron radiation, comprising:

[0007] Obtaining the data of the external detector in the detection area, and inputting the data into the trained reconstruction network model including the mapping relationship between the external detector data and the neutron radiation distribution in the pile for prediction;

[0008] The predicted neutron radiation distribution in the pile is obtained and visualized in three dimensions.

[0009] The beneficial effect of the present invention is that this scheme arranges the radiation distribution inside and outside the pile according to an image by reconstructing the network model, forming a mapping relationship between the off-site detection data and the neutron radiation distribution inside the pile, and according to the off-site detector data in the detection area, the neutron distribution inside the pile can be accurately and efficiently reconstructed.

[0010] Under the premise of controlling the number of detectors and meeting the design space requirements, the reactor core neutron radiation distribution imaging can be further realized. The core neutron radiation imaging has strong practical value in accident conditions or severe operating environments.

[0011] Furthermore, it also includes:

[0012] Converting the neutron distribution in the training data set into a two-dimensional image vector or a three-dimensional image vector;

[0013] Training fully convolutional neural networks of different dimensions according to the two-dimensional image vector or the three-dimensional image vector, forming a mapping relationship between the off-site detection data and the neutron radiation distribution in the pile in the fully convolutional neural network, and obtaining a reconstructed network model after training;

[0014] The training accuracy of the trained reconstructed network model is evaluated, and when the training accuracy meets the requirement, the trained reconstructed network model is saved.

[0015] The beneficial effect of adopting the above further scheme is: this scheme uses the advantages of convolutional neural network feature extraction, arranges the out-core detection data and the in-core neutron radiation distribution data to be measured according to the spatial structure, extracts features, and can sensitively perceive the changes in the out-core detection data.

[0016] Furthermore, it also includes:

[0017] Establish a reactor model including all detection points according to the type of reactor;

[0018] Obtaining an out-of-core detector response matrix according to the reactor model by combining Monte Carlo and inversion calculations;

[0019] According to the off-pillar detector response matrix, the pre-acquired measurement data is labeled, and a training data set is generated by combining the labeled measurement data with existing label data.

[0020] The beneficial effect of adopting the above further scheme is that the present invention avoids the high cost of acquiring and labeling a large amount of data based on the semi-supervised learning technology, makes full use of a small amount of labeled data and unlabeled samples, and generates a large amount of available data sets.

[0021] Furthermore, it also includes: according to a combined optimization algorithm, in combination with the ex-core detector response matrix, obtaining an optimized ex-core detector response matrix;

[0022] The step of labeling the pre-acquired measurement data according to the off-pile detector response matrix and generating a training data set by combining the labeled measurement data with existing label data specifically includes:

[0023] The pre-acquired measurement data are labeled according to the optimized out-of-core detector response matrix, and a training data set is generated by combining the labeled measurement data with existing label data.

[0024] Furthermore, it also includes:

[0025] When the training accuracy does not meet the requirements, the structure of the fully convolutional neural network is adjusted, and training and training accuracy evaluation are performed again.

[0026] Another technical solution of the present invention to solve the above technical problems is as follows:

[0027] A three-dimensional imaging system for reactor neutron radiation, comprising: a detection data prediction module and a visual display;

[0028] The detection data prediction module is used to obtain the external detector data of the detection area, and input the data into the trained reconstruction network model including the mapping relationship between the external detector data and the neutron radiation distribution in the pile for prediction;

[0029] The visualization display is used to obtain the predicted neutron radiation distribution within the pile and perform a three-dimensional visualization display.

[0030] The beneficial effect of the present invention is that this scheme arranges the radiation distribution inside and outside the pile according to an image by reconstructing the network model, forming a mapping relationship between the off-site detection data and the neutron radiation distribution inside the pile, and according to the off-site detector data in the detection area, the neutron distribution inside the pile can be accurately and efficiently reconstructed.

[0031] Under the premise of controlling the number of detectors and meeting the design space requirements, the reactor core neutron radiation distribution imaging can be further realized. The core neutron radiation imaging has strong practical value in accident conditions or severe operating environments.

[0032] Furthermore, it also includes: a reconstruction network model construction training module for converting the neutron distribution in the training data set into a two-dimensional image vector or a three-dimensional image vector;

[0033] Training fully convolutional neural networks of different dimensions according to the two-dimensional image vector or the three-dimensional image vector, forming a mapping relationship between the off-site detection data and the neutron radiation distribution in the pile in the fully convolutional neural network, and obtaining a reconstructed network model after training;

[0034] The training accuracy of the trained reconstructed network model is evaluated, and when the training accuracy meets the requirement, the trained reconstructed network model is saved.

[0035] The beneficial effect of adopting the above further scheme is: this scheme uses the advantages of convolutional neural network feature extraction, arranges the out-core detection data and the in-core neutron radiation distribution data to be measured according to the spatial structure, extracts features, and can sensitively perceive the changes in the out-core detection data.

[0036] Furthermore, it also includes: a training set generation module, which is used to establish a reactor model including all detection points according to the type of reactor;

[0037] Obtaining an out-of-core detector response matrix according to the reactor model by combining Monte Carlo and inversion calculations;

[0038] According to the off-pillar detector response matrix, the pre-acquired measurement data is labeled, and a training data set is generated by combining the labeled measurement data with existing label data.

[0039] The beneficial effect of adopting the above further scheme is that the present invention avoids the high cost of acquiring and labeling a large amount of data based on the semi-supervised learning technology, makes full use of a small amount of labeled data and unlabeled samples, and generates a large amount of available data sets.

[0040] Furthermore, it also includes: an optimization module, which is used to: obtain an optimized ex-core detector response matrix based on a combined optimization algorithm and in combination with the ex-core detector response matrix;

[0041] The training set generation module is specifically used to label the pre-acquired measurement data according to the optimized out-of-core detector response matrix, and generate a training data set by combining the labeled measurement data with existing label data.

[0042] Furthermore, it also includes: an adjustment module, which is used to adjust the structure of the full convolutional neural network when the training accuracy does not meet the requirements, and then perform training and training accuracy evaluation.

[0043] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic flow chart of a method for three-dimensional imaging of reactor neutron radiation provided by an embodiment of the present invention;

[0045] Figure 2 A structural framework diagram of a three-dimensional imaging system of reactor neutron radiation provided by an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of reactor detection channel detector placement and core division provided in other embodiments of the present invention;

[0047] Figure 4 A schematic diagram of a basic flow chart of a system provided for other embodiments of the present invention;

[0048] Figure 5 A simplified arrangement of the out-of-core detectors provided in other embodiments of the present invention;

[0049] Figure 6 A schematic diagram of three-dimensional imaging of core neutron radiation provided for other embodiments of the present invention. DETAILED DESCRIPTION

[0050] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention but not to limit the scope of the present invention.

[0051] like Figure 1 As shown, a three-dimensional imaging method of reactor neutron radiation provided by an embodiment of the present invention includes:

[0052] S1, obtaining the data of the external detector in the detection area, and inputting the data into the trained reconstruction network model including the mapping relationship between the external detector data and the neutron radiation distribution in the pile for prediction;

[0053] It should be noted that obtaining the reconstructed network model may include:

[0054] Converting the neutron distribution in the training data set into a two-dimensional image vector or a three-dimensional image vector;

[0055] Training fully convolutional neural networks of different dimensions according to the two-dimensional image vector or the three-dimensional image vector, forming a mapping relationship between the off-site detection data and the neutron radiation distribution in the pile in the fully convolutional neural network, and obtaining a reconstructed network model after training;

[0056] The training accuracy of the trained reconstructed network model is evaluated, and when the training accuracy meets the requirement, the trained reconstructed network model is saved.

[0057] According to the mapping relationship between the off-site detection data and the neutron radiation distribution inside the pile, the off-site detector data in the detection area is predicted.

[0058] S2, obtain the predicted neutron radiation distribution in the pile and perform a three-dimensional visualization. Figure 6 shown.

[0059] It should be noted that S2 specifically includes: transmitting the energy spectrum data of the predicted neutron radiation distribution in the reactor to the final neutron distribution three-dimensional imaging system. For a specific reactor, the placement of its core components is as follows: Figure 3 As shown, Figure 5 As shown, the following steps can be followed to complete the imaging of the core neutron flux distribution and core source term:

[0060] 1. According to actual needs, the neutron radiation distribution field is divided into five energy intervals, which range from [0MeV, 1E-06MeV], (1E-06MeV, 1E-03MeV], (1E-03MeV, 1E-01MeV], (1E-01MeV, 1E+01MeV] and (1E+01MeV, 2E+01MeV]. Considering the symmetry of the reactor, only half of the reactor model is constructed for training in the simulation. According to the component structure of the reactor core, the 86 core components are divided into three sections radially, totaling 258 spatial partitions, and combined with the energy partitions to form a total of 1290 source items; in this example, the out-of-core detectors are distributed in two positions, one in the detection channel, with an arrangement of 15*3=45, and the other outside the shielding layer, with an arrangement of 32*4=128;

[0061] 2. Use the Monte Carlo transport calculation program to calculate the response matrix for the above model

[0062] 3. Use the combination optimization algorithm to continuously optimize and iterate until a detector layout scheme that meets the constraint conditions is generated, where the inside of the detection channel is 15*3 and the outside of the shielding layer is 8*4=32. Figure 3 The new scheme is then used to form a simplified response matrix and

[0063] 4. Using the model in this paper, the neutron distribution data in the HBR reactor core under critical state is obtained as the source reference data S = {s1,...,s i ,...,s 1290 The most common core accidents are rod ejection accidents and core undercooling accidents, which are characterized by the difficulty of rapid and accurate detection of local changes. This example considers the symmetry of the core and randomly interferes with the neutron radiation distribution at three local locations of the core, namely the outer layer of the core, the middle of the core, and the core of the core, to form S condition_n ;

[0064] 5. Φ after interference condition_n Calculate the neutron flux distribution Φ in the area to be measured condition_n ; Combined response matrix R simplify Calculate the reading F of the detection area outside the corecondition_n Finally, a large amount of labeled data (S, Φ, F) is generated, and finally the data set D is formed;

[0065] 6. In order to achieve reconstruction from two-dimensional to three-dimensional, the data pairs need to be rearranged and expanded into multi-label data pairs. Where S is the core source term distribution, Φ 3D It means that the core neutron flux distribution is arranged in the format of a three-dimensional matrix; It means that the neutron flux distribution of the outermost components of the core is arranged in a three-dimensional matrix; It indicates that the neutron flux distribution of the outermost components of the core is arranged in a two-dimensional matrix; F indicates the two-dimensional arrangement of the detection data outside the core;

[0066] 7. Import the above arranged data set into the multi-dimensional fully convolutional neural network model for training, and use F as input in the two-dimensional convolution. As output, it is trained 1500 times iteratively until the average relative deviation of the prediction is less than 0.5%; in the 3D convolution As input, Φ 3D As output, the training is iterated 1500 times until the average relative deviation of the prediction is less than 0.5%.

[0067] In another embodiment, if Figure 3 As shown, Figure 5 As shown, three sets of multi-threshold detectors are arranged in each detection channel of the HBR-2 reactor, and corresponding driving mechanisms are added to construct an out-of-core detector subsystem that can move within a certain range. For example, they can be 238 U ionization chamber, 235 U ionization chamber, cadmium-coated boron-coated ionization chamber Nat B and recoil proton detectors. Thus, measurements of different energy thresholds in different areas outside the pile can be obtained;

[0068] The measured values ​​are transmitted to the online spectrum decomposition subsystem, and the preset spectrum is constructed through the ratio relationship of the measured values ​​of detectors with different thresholds. Combined with the existing few-channel spectrum decomposition method in the field, the measurement of the neutron energy spectrum at the detector position is realized.

[0069] This scheme arranges the radiation distribution inside and outside the pile according to an image by reconstructing the network model, forming a mapping relationship between the off-site detection data and the neutron radiation distribution inside the pile. According to the off-site detector data in the detection area, the neutron distribution inside the pile can be reconstructed accurately and efficiently.

[0070] Under the premise of controlling the number of detectors and meeting the design space requirements, the reactor core neutron radiation distribution imaging can be further realized. The core neutron radiation imaging has strong practical value in accident conditions or severe operating environments.

[0071] Optionally, in any of the above embodiments, it further includes:

[0072] Converting the neutron distribution in the training data set into a two-dimensional image vector or a three-dimensional image vector;

[0073] Training fully convolutional neural networks of different dimensions according to the two-dimensional image vector or the three-dimensional image vector, forming a mapping relationship between the off-site detection data and the neutron radiation distribution in the pile in the fully convolutional neural network, and obtaining a reconstructed network model after training;

[0074] The training accuracy of the trained reconstructed network model is evaluated, and when the training accuracy meets the requirement, the trained reconstructed network model is saved.

[0075] This scheme uses the advantages of convolutional neural network feature extraction to arrange the out-of-core detection data and the in-core neutron radiation distribution data to be measured according to the spatial structure, extract features, and can sensitively perceive the changes in the out-of-core detection data.

[0076] Optionally, in any of the above embodiments, it further includes:

[0077] Establish a reactor model including all detection points according to the type of reactor;

[0078] Obtaining an out-of-core detector response matrix according to the reactor model by combining Monte Carlo and inversion calculations;

[0079] According to the off-pillar detector response matrix, the pre-acquired measurement data is labeled, and a training data set is generated by combining the labeled measurement data with existing label data.

[0080] The present invention is based on semi-supervised learning technology, which avoids the high cost of acquiring and labeling a large amount of data, and makes full use of a small amount of labeled data and unlabeled samples to generate a large amount of available data sets.

[0081] Optionally, in any of the above embodiments, the method further comprises: obtaining an optimized ex-core detector response matrix according to a combined optimization algorithm in combination with the ex-core detector response matrix;

[0082] The step of labeling the pre-acquired measurement data according to the off-pile detector response matrix and generating a training data set by combining the labeled measurement data with existing label data specifically includes:

[0083] The pre-acquired measurement data are labeled according to the optimized out-of-core detector response matrix, and a training data set is generated by combining the labeled measurement data with existing label data.

[0084] Optionally, in any of the above embodiments, it further includes:

[0085] When the training accuracy does not meet the requirements, the structure of the fully convolutional neural network is adjusted, and training and training accuracy evaluation are performed again.

[0086] In a certain embodiment, a three-dimensional imaging system for reactor neutron distribution based on non-in-situ measurement uses a small number of labeled samples and a large amount of unlabeled data to create a semi-supervised model. At the same time, the system uses feature extraction methods in image processing to arrange the radiation distribution inside and outside the reactor according to the image, forming a mapping relationship between off-site detection data and in-field neutron radiation distribution. Based on the non-in-situ data, the neutron distribution in the reactor can be accurately and efficiently reconstructed.

[0087] It should be noted that the semi-supervised model uses the manually calculated transformation matrix as the supervisory information and the manually annotated information as the supervisory information. Among them, feature extraction serves the semi-supervised module, which is to perform image-like edge and feature point special areas on the radiation field data to extract the discrete feature information of the radiation field. The off-field detection data mainly includes the energy spectrum, neutron intensity distribution and neutron capacity distribution inside the field, which are obtained based on actual measurements outside the field. The neutron radiation distribution inside the field is obtained based on calculations.

[0088] Specifically, in one embodiment, if Figure 4 As shown, a reactor neutron distribution three-dimensional imaging system based on non-in-situ measurement comprises:

[0089] 1) Detector optimization module: According to the type of reactor, a reactor model M containing all possible detection points is established, and the off-reactor detector response matrix R is constructed. Then, the combinatorial optimization algorithm is used to obtain the optimal detector arrangement scheme to obtain the simplified response matrix R. simplify Among them, the combinatorial optimization algorithm uses the method of calculating the eigenvalues ​​of the response matrix and the underdetermination of the selected arrangement method to obtain the optimal arrangement scheme.

[0090] 2) Semi-supervised reconstruction learning module: This module combines the R simplify A large amount of unlabeled data is labeled, i.e., off-site measurement data, and combined with a small amount of labeled data to generate a training data set D, where the small amount of labeled data is a small amount of off-site and corresponding on-site data obtained by Monte Carlo calculation. In addition, the neutron distribution of the reactor in data set D is converted into a two-dimensional or three-dimensional image vector, and trained with fully convolutional neural networks of different dimensions, and the relationship between the off-site detection data and the neutron distribution inside the pile is learned from the inside to the outside, and the learning results are evaluated; wherein, the semi-supervised model is obtained by training the neural network on unlabeled data. The off-site detection data includes: off-site energy spectrum, energy and intensity data. The neutron distribution inside the pile includes: on-site energy spectrum, energy and intensity distribution data.

[0091] 3) Core radiation imaging module: obtain the readings of the external detector and input them into the loaded reconstruction network model, finally obtain the predicted neutron distribution in the core, and visualize it. The reconstruction network model replaces the Monte Carlo calculation to obtain the neutron distribution information in the field. The model is a semi-supervised model obtained by the fully convolutional neural network mentioned above.

[0092] Specifically, the detailed steps of step 1) include:

[0093] 1.1) Establish the reactor neutron transport calculation model M and set the neutron energy partition E (e1, ..., e E}, the reactor core is divided into R spatial partitions to form a set of spatial partitions {s1, ..., s R}, forming a total of N = E * R neutron sources, each with an intensity of s i , then the core source term is S = {s1, ..., s i , ..., s N}; where M represents the detection point model.

[0094] 1.2) Monte Carlo transport calculation is used to calculate the detection response matrix of D different detection areas outside the neutron reactor under N (N neutron sources) unit source strengths according to the reactor neutron transport calculation model M:

[0095]

[0096] 1.3) The purpose of the combinatorial optimization algorithm is to use a given number A of detectors, of which there are M types, to design an arrangement that can have a good response to the pile after the positions are determined. The sum of the responses of the detectors arranged outside the pile to the nth neutron source in the pile In the best case, the neutron source response at the center of the pile is maximized, so the optimization goal is:

[0097]

[0098] in, Neutron source response at the center of the pile, A detectors, Represents the value ri of each detector and calculates the sum cn; where i≤A.

[0099] 1.4) Generate a simplified response matrix R based on the optimized detector layout scheme simplify .

[0100] Specifically, the detailed steps of step 2) include:

[0101] 2.1) Perform neutron transport calculations, select some source items, that is, select source items under typical and special conditions, and simulate the neutron radiation distribution Φ={φ1,...,φ i , ..., φ N} and the off-heap detection area reading F = {f1, ..., f i , ..., f D}, forming the labeled data (S, Φ, F). The remaining source items (source items of all other unselected measurement calculations) form the unlabeled data (S condition_n ,? ,? ); where S is the core source term, and the other items of unlabeled data are the distribution and readings to be measured.

[0102] 2.2) Take S condition_n , combined with the response matrix R of the out-of-pile detection area simplify , calculate the corresponding core neutron flux Φ at the position to be measured condition_n And the reading of the external detector F condition_n , forming label data (S condition_n ,φ condition_n ,F condition_n ), and finally form a data set D for network training, where the data set D represents the label data (S, Φ, F) of some source items and the label data of the remaining source items.

[0103] 2.3) Preprocess the multi-label sample data. The sample data here means that since some labels are calculated, there may be overlaps (that is, multiple labels), for these data; wherein, the preprocessing includes soft labels, hard threshold hard labels, etc. Vectorization processing: Similar to images, each counting point of the reactor is regarded as a pixel point, and the energy group distribution of the unit is the thickness of the image (similar to RGB). The detection area in the present invention uses a single channel because there is only one energy group. Therefore, the composed 4D tensor is (samples, height, weight, energy_depth), where samples is the number of training set samples, height is the height of a certain radiation field in the core after discretization, weight is the width of a certain radiation field in the core after discretization, and energy_depth represents the number of discretized energy groups. The components of the core part are composed of 5D tensors (samples, length, height, weight, energy_depth). Normalization processing: The original neutron radiation distribution data are relatively small in magnitude and can reach 10 -7 ~10 -8 Therefore, there are a lot of singular sample data. For this type of data, this paper uses normalization processing to limit the data to the range of [0,1].

[0104] 2.4) Establish the readings of D detectors outside the pile F = {f1,...,f i ,...,f D} and the neutron radiation distribution of the outermost components of the reactor core According to the structure of the data, the D detector values ​​are arranged in two dimensions as follows:

[0105]

[0106] Where D = h*w, for the neutron radiation distribution of the outermost P components of the reactor core Each energy group of is also arranged in two dimensions:

[0107]

[0108] Where P = a*b. Construct a two-dimensional fully convolutional neural network, input two-dimensional data, train, calculate the loss function, and form a two-dimensional mapping relationship Among them, the two-dimensional fully convolutional neural network is in the form of feature extraction + FPN + head, the purpose of which is to extract the abstract features of the input data and obtain the mapping relationship of the desired output data.

[0109] 2.5) Establish the neutron radiation distribution of the outermost components of the core and the neutron radiation distribution in the core Φ 3D The three-dimensional mapping relationship between the core and the neutron radiation distribution of the outermost components and the core source term S 3D The neutron radiation distribution of the outermost components of the reactor is arranged in three dimensions, and the neutron distribution in the core is also arranged in three dimensions. A three-dimensional fully convolutional neural network is constructed, and three-dimensional data is input for training to form a three-dimensional mapping relationship. and The function of the three-dimensional fully convolutional neural network is to extract three-dimensional abstract features. The full convolution has the ability to reversely iterate and can accurately calculate the mapping information.

[0110] 2.6) In the architecture training of the fully convolutional neural network, the loss function used for training is as follows:

[0111]

[0112] L FCN =L 2D +L 3D ,

[0113] Among them, LFCN is the total loss function, L2D is the two-dimensional loss function, and L3D is the three-dimensional loss function

[0114] NM represents the length in the two-dimensional direction, f(xi) represents the core data calculated by the network output, and yi is the accurate core data obtained by numerical calculation, that is, the label data in training.

[0115] The loss function is used to calculate the gradient of loss attenuation and perform back propagation iteration of the network to train and obtain appropriate network parameters.

[0116] 2.7) Input validation set data to evaluate the trained model parameters;

[0117] 2.8) Determine whether the evaluation value meets the requirements of reactor monitoring, that is, whether the error between the result inferred on the verification set and the measured value meets the calculation requirements, and whether the error meets the required standard. If it does not meet the requirements, adjust the network structure or retrain the parameters, and jump to step 2.7); otherwise, save the learned network model.

[0118] 3 Core radiation imaging module includes:

[0119] Get off-pile detector readings;

[0120] Load and reconstruct the network model;

[0121] Predict core neutron distribution, core neutron distribution data output by 3D fully convolutional neural network;

[0122] Three-dimensional imaging display of core radiation distribution.

[0123] 1. The present invention is based on semi-supervised learning technology, which avoids the high cost of acquiring and labeling a large amount of data, and makes full use of a small amount of labeled data and unlabeled samples to generate a large amount of available data sets;

[0124] 2. The present invention uses the advantages of convolutional neural network feature extraction to arrange the out-core detection data and the in-core neutron radiation distribution data to be measured according to the spatial structure, extract features, and can sensitively perceive the changes in the out-core detection data;

[0125] 3. Existing methods are mostly aimed at imaging the power distribution of reactors, and most of them use detectors placed inside the core. The system proposed in this patent can further realize the neutron radiation distribution imaging of the reactor core under the premise of controlling the number of detectors and meeting the design space requirements. The core neutron radiation imaging has strong practical value under accident conditions or severe operating environment conditions.

[0126] In one embodiment, if Figure 2 As shown, a three-dimensional imaging system of reactor neutron radiation includes: a detection data prediction module 1101 and a visualization display 1102;

[0127] The detection data prediction module 1101 is used to obtain the external detector data of the detection area, and input the data into the trained reconstruction network model including the mapping relationship between the external detector data and the neutron radiation distribution in the pile for prediction;

[0128] The visualization display 1102 is used to obtain the predicted neutron radiation distribution within the pile and perform a three-dimensional visualization display.

[0129] This scheme arranges the radiation distribution inside and outside the pile according to an image by reconstructing the network model, forming a mapping relationship between the off-site detection data and the neutron radiation distribution inside the pile. According to the off-site detector data in the detection area, the neutron distribution inside the pile can be reconstructed accurately and efficiently.

[0130] Under the premise of controlling the number of detectors and meeting the design space requirements, the reactor core neutron radiation distribution imaging can be further realized. The core neutron radiation imaging has strong practical value in accident conditions or severe operating environments.

[0131] Optionally, in any of the above embodiments, it further comprises: a reconstruction network model construction training module, which is used to convert the neutron distribution in the training data set into a two-dimensional image vector or a three-dimensional image vector;

[0132] Training fully convolutional neural networks of different dimensions according to the two-dimensional image vector or the three-dimensional image vector, forming a mapping relationship between the off-site detection data and the neutron radiation distribution in the pile in the fully convolutional neural network, and obtaining a reconstructed network model after training;

[0133] The training accuracy of the trained reconstructed network model is evaluated, and when the training accuracy meets the requirement, the trained reconstructed network model is saved.

[0134] This scheme uses the advantages of convolutional neural network feature extraction to arrange the out-of-core detection data and the in-core neutron radiation distribution data to be measured according to the spatial structure, extract features, and can sensitively perceive the changes in the out-of-core detection data.

[0135] Optionally, in any of the above embodiments, it further comprises: a training set generation module, used to establish a reactor model including all detection points according to the type of reactor;

[0136] Obtaining an out-of-core detector response matrix according to the reactor model by combining Monte Carlo and inversion calculations;

[0137] According to the off-pillar detector response matrix, the pre-acquired measurement data is labeled, and a training data set is generated by combining the labeled measurement data with existing label data.

[0138] The present invention is based on semi-supervised learning technology, which avoids the high cost of acquiring and labeling a large amount of data, and makes full use of a small amount of labeled data and unlabeled samples to generate a large amount of available data sets.

[0139] Optionally, in any of the above embodiments, it further includes: an optimization module, used to: obtain an optimized ex-core detector response matrix according to a combined optimization algorithm and in combination with the ex-core detector response matrix;

[0140] The training set generation module is specifically used to label the pre-acquired measurement data according to the optimized out-of-core detector response matrix, and generate a training data set by combining the labeled measurement data with existing label data.

[0141] Optionally, in any of the above embodiments, it also includes: an adjustment module, which is used to adjust the structure of the full convolutional neural network when the training accuracy does not meet the requirements, and then perform training and training accuracy evaluation.

[0142] It can be understood that in some embodiments, some or all of the optional implementation modes in the above embodiments may be included.

[0143] It should be noted that the above embodiments are product embodiments corresponding to the previous method embodiments. For the description of the optional implementation methods in the product embodiments, reference can be made to the corresponding description in the above method embodiments, which will not be repeated here.

[0144] The reader should understand that in the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0145] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the method embodiments described above are only illustrative, for example, the division of steps is only a logical function division, and there may be other division methods in actual implementation, such as multiple steps can be combined or integrated into another step, or some features can be ignored or not executed.

[0146] If the above method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0147] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A three-dimensional imaging method of reactor neutron radiation, characterized in that: include: Obtaining the data of the external detector in the detection area, and inputting the data into the trained reconstruction network model including the mapping relationship between the external detector data and the neutron radiation distribution in the pile for prediction; Obtain the predicted neutron radiation distribution in the pile and perform a three-dimensional visualization; Also includes: Converting the neutron distribution in the training data set into a two-dimensional image vector or a three-dimensional image vector; Training fully convolutional neural networks of different dimensions according to the two-dimensional image vector or the three-dimensional image vector, forming a mapping relationship between the off-site detection data and the neutron radiation distribution in the pile in the fully convolutional neural network, and obtaining a reconstructed network model after training; Evaluate the training accuracy of the trained reconstructed network model, and save the trained reconstructed network model when the training accuracy meets the requirement; Also includes: Establish a reactor model including all detection points according to the type of reactor; Obtaining an out-of-core detector response matrix according to the reactor model by combining Monte Carlo and inversion calculations; According to the off-pillar detector response matrix, the pre-acquired measurement data is labeled, and a training data set is generated by combining the labeled measurement data with existing label data.

2. A three-dimensional imaging method of reactor neutron radiation according to claim 1, characterized in that: Also includes: According to the combined optimization algorithm, in combination with the ex-core detector response matrix, an optimized ex-core detector response matrix is ​​obtained; The step of labeling the pre-acquired measurement data according to the off-pile detector response matrix and generating a training data set by combining the labeled measurement data with existing label data specifically includes: The pre-acquired measurement data are labeled according to the optimized out-of-core detector response matrix, and a training data set is generated by combining the labeled measurement data with existing label data.

3. The method for three-dimensional imaging of reactor neutron radiation according to claim 1, characterized in that: Also includes: When the training accuracy does not meet the requirements, the structure of the fully convolutional neural network is adjusted, and training and training accuracy evaluation are performed again.

4. A three-dimensional imaging system of reactor neutron radiation, characterized in that: include: Detection data prediction module and visualization display; The detection data prediction module is used to obtain the external detector data of the detection area, and input the data into the trained reconstruction network model including the mapping relationship between the external detector data and the neutron radiation distribution in the pile for prediction; The visualization display is used to obtain the predicted neutron radiation distribution in the pile and perform a three-dimensional visualization display; It also includes: a reconstruction network model construction training module for converting the neutron distribution in the training data set into a two-dimensional image vector or a three-dimensional image vector; Training fully convolutional neural networks of different dimensions according to the two-dimensional image vector or the three-dimensional image vector, forming a mapping relationship between the off-site detection data and the neutron radiation distribution in the pile in the fully convolutional neural network, and obtaining a reconstructed network model after training; Evaluate the training accuracy of the trained reconstructed network model, and save the trained reconstructed network model when the training accuracy meets the requirement; It also includes: a training set generation module, which is used to establish a reactor model including all detection points according to the type of reactor; Obtaining an out-of-core detector response matrix according to the reactor model by combining Monte Carlo and inversion calculations; According to the off-pillar detector response matrix, the pre-acquired measurement data is labeled, and a training data set is generated by combining the labeled measurement data with existing label data.

5. The three-dimensional imaging system of reactor neutron radiation according to claim 4, characterized in that: Also includes: An optimization module is used to obtain an optimized ex-core detector response matrix according to a combined optimization algorithm and in combination with the ex-core detector response matrix; The training set generation module is specifically used to label the pre-acquired measurement data according to the optimized out-of-core detector response matrix, and generate a training data set by combining the labeled measurement data with existing label data.

6. The three-dimensional imaging system of reactor neutron radiation according to claim 4, characterized in that: Also includes: The adjustment module is used to adjust the structure of the fully convolutional neural network when the training accuracy does not meet the requirements, and then perform training and training accuracy evaluation.

Images