Reactor core neutron field ex-situ explicit inversion method based on graph structure

Abstract

The invention discloses a reactor core neutron field ex-situ explicit inversion method based on a graph structure, and belongs to the technical field of reactor core safety monitoring and control. According to the method, reactor core space coordinates, energy spectrum grouping, material physical properties and prior information are fused through an explicit graph structure, a lightweight graph convolutional network architecture is adopted, and accurate inversion of a high-dimensional reactor core neutron field is achieved. And meanwhile, the inversion precision under the high-dimensional resolution can be remarkably improved, and reliable technical support can be provided for safe operation and intelligent monitoring of an advanced reactor type. Compared with an existing deep learning black box model, the method has the advantages that the occurrence probability of local high deviation can be remarkably reduced at different reactor core change positions, and more uniform and stable prediction performance can be realized in the whole space. And under the condition that the reactor core has multi-region change, better inversion performance is still shown. When the middle and low complexity changes, the fine multi-region features of the reactor core can be effectively and stably captured, and the method has better overall error suppression capability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of reactor core safety monitoring and control, and particularly relates to a non-in-situ explicit inversion method for a reactor core neutron field based on a graph structure. BACKGROUND

[0002] In the engineering practice development of the fourth generation advanced nuclear energy system, the reactor in-core neutron probe is challenged by extreme operating environments, including high temperature, high irradiation, corrosion, and strong magnetic field environments. These factors seriously restrict the arrangement and long-term operation reliability of the traditional in-core neutron probe, thereby affecting the efficiency and accuracy of the core safety monitoring and control. Especially in special application scenarios such as space nuclear power propulsion systems, the core space is limited, which further aggravates the implementation difficulty of in-core measurement.

[0003] In order to overcome the above limitations, in recent years, a large amount of research work on inversion methods for full-core power (neutron) distribution based on non-in-situ discrete probe signals has been carried out at home and abroad. However, the existing inversion techniques mainly rely on two paths: one is a physical mechanism model, and the other is a data-driven model. The physical mechanism model such as harmonic synthesis, eigenorthogonal decomposition, principal component analysis, etc. has certain interpretability, but it seriously depends on fixed prior knowledge base, and the preset parameters are difficult to dynamically adjust for the rapid change of power distribution under transient conditions, which seriously affects the real-time monitoring efficiency of future advanced reactors. On the other hand, with the rapid development of artificial intelligence technology, the method of replacing traditional physical mechanism model with data-driven intelligent model has become an effective solution to complex engineering problems in reactors, showing stronger modeling flexibility and data adaptability. However, pure data-driven black-box modeling is difficult to break through the precision bottleneck caused by the lack of physical laws, and the introduction of complex network structure to improve the prediction ability will reduce the real-time performance of the prediction. Therefore, a non-in-situ inversion technique that can explicitly express the core structure information and has good generalization ability and calculation efficiency is needed to improve the prediction stability and accuracy under complex conditions. SUMMARY

[0004] Under the high-resolution complex core change scenario, pure data-driven black-box modeling is difficult to break through the precision bottleneck caused by the lack of physical laws, and the introduction of complex network structure to improve the prediction ability will reduce the real-time performance of the prediction. Therefore, in the micro reactor scene where the observation information is sparse and the neutron transport and thermal feedback are highly coupled, how to clearly define the multi-physical coupling evolution characteristics of the core power distribution, and construct an explicit and lightweight inversion model that deeply integrates physical laws and data-driven, based on this, the application proposes a non-in-situ explicit inversion method for a reactor core neutron field based on a graph structure, which can significantly improve the inversion accuracy under high-dimensional resolution, and can provide reliable technical support for the safe operation and intelligent monitoring of advanced reactors.

[0005] In order to achieve the above object, the technical scheme adopted by the present application is:

[0006] A non-in-situ explicit inversion method of core neutron field based on graph structure, the steps are as follows:

[0007] Step one: non-in-situ observation signal acquisition and up-sampling

[0008] In order to solve the problem of incomplete graph structure information caused by sparse observation data, the non-in-situ detection signal outside the core is simulated and up-sampled;

[0009] Step two: core graph structure modeling

[0010] The structural information of the core is explicitly represented as a graph data structure to support the modeling of physical coupling relationship and the propagation of neutron information between nodes;

[0011] Step three: graph neural network feature diffusion

[0012] In order to realize the multi-region information fusion of neutron behavior between core blocks, a graph neural network (GNN) is introduced as a feature propagation mechanism based on graph structure modeling; the input of the network is the initial feature vector of each node, including local physical properties and up-sampled observation signals, as well as the adjacency matrix constructed to reflect the direct and indirect coupling structure between nodes;

[0013] Step four: three-dimensional convolution feature transformation

[0014] In order to further extract the structural continuity features of the core neutron field in three-dimensional space, the node features output by the graph neural network are rearranged into a tensor form, and a three-dimensional voxel data is constructed to input into a three-dimensional convolutional neural network (3D-CNN);

[0015] Step five: loss function and training optimization

[0016] In order to consider the global inversion accuracy of the core and the physical consistency of the boundary observation constraint, a double loss mechanism is designed, which is used to predict and supervise the outermost component region and the whole core region respectively, which makes the model optimize the whole field prediction, and uses the high reliability observation information of the outer region as auxiliary guidance, so as to improve the inference accuracy and stability under boundary control.

[0017] As a preferred technical scheme of the present application, in step one, the Monte Carlo transport method is used to simulate the neutron behavior of the lead-based core, a plurality of observation points are set outside the reactor core, and the neutron detection response at each measurement point is counted; through enough particle history sampling, a statistically stable observation signal data set is obtained, denoted as:

[0018] (1)

[0019] where M is the number of non-in-situ detectors actually deployed or virtually set;

[0020] Subsequently, the sparse observation signal is constructed into a voxel-level tensor:

[0021] (2)

[0022] Each is mapped to the voxel index position of the corresponding observation point, and the remaining voxels are filled with 0 or NaN as a missing marker;

[0023] Next, the spatial signal reconstruction is performed by using a designed up-sampling network UpsampleCNN:

[0024] (3)

[0025] The dense signal tensor is reshaped into a matrix form for splicing with the graph node features.

[0026] As a preferred technical solution of the present application, in step two, the core space is first divided into K segments, each segment corresponds to a graph node; the node feature vector is:

[0027] (4)

[0028] where represents the neutron flux of the i-th segment, represents the material code of the i-th segment, represents the local temperature of the i-th segment, represents the fuel enrichment and other physical information of the i-th segment;

[0029] Subsequently, according to the component geometric connection and physical coupling, a first-order adjacency relationship matrix between nodes is constructed, if node and node are directly adjacent in space, then is defined; the connection criteria between nodes can be based on geometric proximity, physical proximity, and material continuity or symmetry distribution;

[0030] In order to represent the cross-region propagation of neutron distribution changes, a high-order adjacency relationship is further introduced; the second-order adjacency is defined as:

[0031] (5)

[0032] wherein, representing nodes and nodes There is a two-hop path between; it can be extended to three, four-order adjacency; to integrate multi-order information, construct a comprehensive adjacency matrix:

[0033] + +... (6)

[0034] wherein, is the attenuation factor of the k-order adjacency information, satisfying , control the contribution weight of the far neighbor block of neutron distribution to the overall structure propagation; the edge weight is defined as:

[0035] (7)

[0036] wherein, is the sum of the geometric distance on the k-order path, is the multi-hop physical propagation factor.

[0037] As a preferred technical solution of the present application, in step three, in the graph neural network, graph convolution (GCN) is used as the basic information transmission unit; the update of each layer of graph convolution follows the neighborhood feature aggregation mechanism, which is specifically expressed as:

[0038] (8)

[0039] wherein, representing the adjacency matrix with self-loop, is its degree matrix, is a trainable parameter, is an activation function;

[0040] The network structure further introduces multi-layer stacking, and combines residual connection and Dropout regularization;

[0041] The finally output node feature matrix contains the spatial expression information of each block after multiple rounds of feature propagation.

[0042] As a preferred technical solution of the present application, in step four, the graph output feature matrix is reshaped into a tensor according to the core spatial grid layout, wherein are the discrete block numbers of the core in three spatial dimensions, respectively, is the number of output channels of the graph neural network;

[0043] The tensor is input into a designed three-dimensional convolution network, and a multi-layer convolution operation is performed to identify the spatial smoothness, gradual structure and non-uniform disturbance mode in the local neutron field distribution; the size of the convolution kernel is generally taken , the step is 1 or 2, and a ReLU activation function is used; the final output of the network is a single-channel tensor , corresponding to the neutron flux inversion value of each block of the core.

[0044] As a preferred technical scheme of the application, in step five, the loss function is composed of two parts: one is the prediction loss of the outer layer node , which is used to strengthen the response consistency between the boundary region and the non-in-situ observation data; the second is the prediction loss of the whole core range , which reflects the accuracy of the global neutron flux distribution; the joint loss function is expressed as:

[0045] (9)

[0046] (10)

[0047] (11)

[0048] Wherein, N is the set of all nodes of the core, is the set of nodes of the outermost component, is the adjustment coefficient of the auxiliary loss term.

[0049] The application proposes a graph structure-based core neutron field non-in-situ explicit inversion (GS-EINF) method, which fuses the core spatial coordinates, energy spectrum grouping, material properties and prior information through an explicit graph structure, and uses a lightweight graph convolution network architecture to realize accurate inversion of high-dimensional core neutron fields. At the same time, it can significantly improve the inversion accuracy under high-dimensional resolution, and can provide reliable technical support for the safe operation and intelligent monitoring of advanced reactor types. Compared with the existing deep learning black box model (MCRNet), the advantages are:

[0050] 1. Under different core change positions, the graph structure-based core neutron field non-in-situ explicit inversion (GS-EINF) method proposed by the application can significantly reduce the probability of local high bias, and realize more uniform and stable prediction performance in the overall space. At the same time, a better balance is achieved between global feature learning and local detail expression.

[0051] 2. In the case of multiple regional changes in the core, the graph structure-based core neutron field non-in-situ explicit inversion (GS-EINF) method proposed by the application still shows better inversion performance. In the case of medium and low complexity changes, it can effectively and stably capture the subtle multi-region features of the core, and has better overall error suppression ability. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the explicit inversion process of the core field based on the graph structure.

[0053] Figure 2 This is a comparison of the inversion results of MCRNet and GS-EINF under a single region variation in the reactor core.

[0054] Figure 3 This is a comparison of the inversion results of MCRNet and GS-EINF under multi-region variations in the reactor core. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0056] Please see Figure 1 As shown, this invention proposes a graph-based non-in-situ explicit inversion method for core neutron fields (GS-EINF), the detailed technical solution of which is as follows:

[0057] Step 1: Acquisition and Upsampling of Non-In-Situ Observation Signals

[0058] To address the issue of incomplete graph structure information caused by sparse observation data, it is necessary to simulate and upsample the non-in-situ detection signals outside the reactor core.

[0059] First, a Monte Carlo transport method (such as MCNP or OpenMC) is used to physically simulate the neutron behavior of the lead-based reactor core. Multiple observation points are set up outside the reactor core, and the neutron detection response (such as count rate or reaction rate) at each measurement point is statistically analyzed. Through a sufficient number of particle history samplings, a statistically stable observation signal dataset is obtained, denoted as:

[0060] (1)

[0061] Where M represents the number of non-in-situ detectors that are actually deployed or virtually set up.

[0062] Subsequently, this sparse observation signal was constructed into a voxel-level tensor:

[0063] (2)

[0064] Each Map to the voxel index position of the corresponding observation point, and fill the remaining voxels with 0 or NaN as missing measurement markers.

[0065] Next, spatial signal reconstruction is performed using the designed upsampling network UpsampleCNN (which can be composed of deconvolution, interpolation + convolution modules):

[0066] (3)

[0067] Obtain dense signal tensor , reshape into matrix form for concatenation with graph node features.

[0068] Step 2: Core graph structure modeling

[0069] The goal of this step is to explicitly represent the structural information of the core as a graph data structure to support modeling of physical coupling relationships and propagation of neutron information between nodes.

[0070] First, the core space is divided into individual bins, each bin corresponding to a graph node. The node feature vector is:

[0071] (4)

[0072] where denotes the neutron flux of the i-th bin, denotes the material code of the i-th bin, denotes the local temperature of the i-th bin, denotes the fuel enrichment of the i-th bin, and other physical information.

[0073] Subsequently, based on component geometric connections and physical coupling, a first-order adjacency relationship matrix between nodes is constructed . If node and node are directly adjacent in space, then is defined. The connection criteria between nodes can be based on geometric proximity (such as sharing edges or vertices), physical proximity (such as thermally coupled regions or neutron coupling regions within the diffusion length), and material continuity or symmetry distribution.

[0074] To represent the cross-region propagation of neutron distribution changes, higher-order adjacency relationships are further introduced. Taking the second-order adjacency as an example, it is defined as:

[0075] (5)

[0076] where denotes that there is a two-hop path between node and node . Similarly, it can be extended to three-order, four-order adjacency. To integrate multi-order information, a comprehensive adjacency matrix is constructed:

[0077] + +... (6)

[0078] where, is the decay factor of the k-th order adjacency information, satisfying , which controls the contribution weight of the neutron distribution of the far neighbor bin to the overall structure propagation. The edge weight is defined as:

[0079] (7)

[0080] where, is the geometric distance sum on the k-th order path, is the multi-hop physical propagation factor (such as thermal neutron scattering probability, power correlation degree, etc.).

[0081] Through this high-order adjacency modeling mechanism, the core graph structure not only retains the local topology and physical structure information, but also captures the neutron and energy influence between remote regions, providing more complete physical connection support for subsequent information diffusion and global modeling of graph neural networks.

[0082] Step three: feature diffusion of graph neural network

[0083] On the basis of step one and step two, in order to realize the multi-region information fusion of neutron behavior between core bins, the present application introduces a graph neural network (Graph Neural Network, GNN) as a feature propagation mechanism on the basis of graph structure modeling. The input of the network is the initial feature vector of each node, which includes local physical properties and up-sampled observation signals, as well as the constructed adjacency matrix , which integrates first-order and high-order adjacency relationships to reflect the direct and indirect coupling structure between nodes.

[0084] In the graph neural network, graph convolution (GCN) is used as the basic information transmission unit. The update of each layer of graph convolution follows the neighborhood feature aggregation mechanism, which is specifically expressed as:

[0085] (8)

[0086] where, denotes the adjacency matrix with self-loop, is its degree matrix, is a trainable parameter, is an activation function. This mechanism enables each node to fuse the information of its adjacent nodes in the feature space, realizing the abstract modeling of neutron diffusion effect.

[0087] In order to enhance the expression ability and convergence stability of the model, a multi-layer stack (usually 2 to 4 layers) is further introduced in the network structure, combined with residual connection and Dropout regularization. The residual structure helps to alleviate the problem of over-smoothing in deep network, while Dropout improves the generalization ability of the model, adapting to multiple working conditions of the core state. The final output node feature matrix , which contains the spatial representation information of each block after multiple rounds of feature propagation.

[0088] The graph neural network module simulates the multi-step leakage path and non-local interaction of neutrons inside the core from an algorithmic perspective, and realizes "information diffusion under physical constraints" through the graph structure. This data-driven topology modeling method not only explicitly utilizes the relationship between the core structure and the detection signal, but also provides a structurally consistent and semantically rich feature basis for subsequent three-dimensional field reconstruction.

[0089] Step four: three-dimensional convolution feature transformation

[0090] To further extract the structural continuity features of the neutron field in the three-dimensional space of the core, the invention rearranges the node features output by the graph neural network into a tensor form, and constructs a three-dimensional voxel data input into a three-dimensional convolutional neural network (3D-CNN).

[0091] Specifically, the graph output feature matrix is reshaped into a tensor according to the spatial grid layout of the core , where are the discrete block numbers of the core in three spatial dimensions, respectively, is the number of output channels of the graph neural network.

[0092] The tensor is input into the designed three-dimensional convolutional network, and multi-layer convolution operations are performed to identify the spatial smoothness, gradual structure and non-uniform disturbance pattern in the local neutron field distribution. The convolution kernel size is generally , the step is 1 or 2, and the ReLU activation function is used. The final output of the network is a single-channel tensor , corresponding to the neutron flux inversion value of each block of the core. This module models the local consistency of node features from a spatial perspective, providing structural enhancement of physical continuity for fine neutron field reconstruction.

[0093] Step five: loss function and training optimization

[0094] To balance the global inversion accuracy of the core and the physical consistency of the boundary observation constraint, the invention designs a dual loss mechanism, which respectively predicts and supervises the outermost component area and the overall core area.

[0095] The loss function consists of two parts: one is the prediction loss of the outer nodes , which is used to strengthen the response consistency between the boundary area and the non-in-situ observation data; the second is the prediction loss of the whole core range , which reflects the accuracy of the global neutron flux distribution. The joint loss function is expressed as:

[0096] (9)

[0097] (10)

[0098] (11)

[0099] where N is the set of all nodes in the core, is the set of nodes in the outermost assembly, is the adjustment coefficient of the auxiliary loss term.

[0100] This design enables the model to optimize the overall field prediction while using the high-credibility observation information in the outer region as auxiliary guidance, thereby improving the inference accuracy and stability under boundary control.

[0101] During training, the Adam optimizer is used, with an initial learning rate of 0.0015, combined with Dropout (anti-overfitting), Batch Normalization (accelerate convergence) and Early Stopping strategy (avoid over-training) mechanisms to improve the robustness and generalization ability of the overall training.

[0102] The present application fuses the core spatial coordinates, energy spectrum grouping, material properties and prior information through explicit graph structure, and adopts a lightweight graph convolution network architecture to realize accurate inversion of high-dimensional core neutron field. Compared with existing deep learning black box models, in order to comprehensively evaluate the performance of the GS-EINF method, it is compared with the latest MCRNet method (paper name: Multi-zone cooperative reconstruction network for off-situ monitoring of the core neutron field; paper address: https: / / www.sciencedirect.com / science / article / abs / pii / S0306454924006984) under different core change positions, as shown in Figure 2 .

[0103] From the proportion R RD≥10% of the inversion units with a relative deviation greater than or equal to 10%, it can be seen that the inversion deviation of GS-EINF can be kept within 10% in most cases, and only a small number of units exceed the standard near the core at the stage m≥11, and the proportion is always controlled within 0.7%; while the MCRNet method has nodes with a relative deviation exceeding 10% in the whole range, R RD≥10% fluctuates between 0.22%-1.74%. This result shows that GS-EINF can significantly reduce the probability of local high deviation, and achieve more uniform and stable prediction performance in the overall space.

[0104] In terms of the maximum relative deviation RD max In terms of the maximum relative deviation RD max In terms of the average relative deviation ARD, the GS-EINF is superior to the MCRNet at all complexity stages, and the average deviation is maintained in the interval of 0.19%-0.69%, with an overall reduction of more than 50%. This result shows that the proposed model achieves a better balance between global feature learning and local detail expression.

[0105] In the case of multiple regional changes in the core, the GS-EINF proposed by the application still exhibits better inversion performance, as shown in Figure 3

[0106] In terms of the proportion R RD≥10% of the inversion units with a relative deviation greater than 10%, the GS-EINF can basically be maintained at 0% or a very low level when n=2-6, and only a small increase of less than 1% occurs when n=4, n=7 and n=8, indicating that when the complexity of changes is low, the new method can effectively and stably capture the subtle multi-regional features of the core. In comparison with the MCRNet method, as the complexity of changes increases, the proportion of R RD≥10% rapidly increases, and starts to jump to more than 1% from n=4, and further climbs to 5%-6% when the number of core change regions increases (n>9), showing the problem of insufficient adaptability to multi-regional changes.

[0107] In terms of the maximum relative deviation RD max and the average relative deviation ARD, the GS-EINF method maintains a relatively low error level. In terms of the maximum relative deviation RD max , the GS-EINF is always controlled within 16% when n=2-6, which is much lower than the level of more than 20% of the MCRNet, and even when the number of core change regions increases (n>9), the RD max is 2%-7% lower than the comparative method. In terms of the average relative deviation ARD, the GS-EINF method is stably maintained between 0.15%-1.45%, and even at the stage of severe changes (n=14-18), the growth rate of ARD is still significantly smaller than that of the MCRNet method, which reflects better overall error suppression capability.

[0108] In summary, the application can significantly improve the inversion accuracy under high-dimensional resolution, and can provide reliable technical support for the safe operation and intelligent monitoring of advanced reactor types.​

Claims

1. A method for in-core neutron field non-in situ explicit inversion based on graph structure, characterized in that, The steps are as follows: Step one: non-in-situ observation signal acquisition and up-sampling To solve the problem of incomplete graph structure information caused by sparse observation data, non-in-situ detection signals outside the core are simulated and up-sampled; Step two: modeling of the core graph structure The structural information of the core is explicitly represented as a graph data structure to support modeling of physical coupling relationships and propagation of neutron information between nodes; Step three: graph neural network feature diffusion To realize multi-region information fusion of neutron behavior between core blocks, a graph neural network (GNN) is introduced as a feature propagation mechanism based on graph structure modeling; the input of the network is the initial feature vector of each node, including local physical properties and up-sampled observation signals, as well as the constructed adjacency matrix, which reflects the direct and indirect coupling structure between nodes; Step four: three-dimensional convolution feature transformation To further extract the structural continuity features of the core neutron field in three-dimensional space, the node features output by the graph neural network are rearranged into a tensor form, and a three-dimensional voxel data is constructed to input into a three-dimensional convolutional neural network (3D-CNN); Step five: loss function and training optimization To balance the global inversion accuracy of the core and the physical consistency of the boundary observation constraint, a dual loss mechanism is designed to predict and supervise the outermost component region and the overall core region, respectively. This design enables the model to optimize the overall prediction while using the high-accuracy observation information of the outer region as an auxiliary guide to improve the inference accuracy and stability under boundary control.

2. The graph structure based incore neutron field non-in situ explicit inversion method of claim 1, wherein, In step one, the Monte Carlo transport method is used to simulate the neutron behavior of the lead-based core, and multiple observation points are set outside the reactor core to statistically record the neutron detection response at each point. By sampling a sufficient number of particle histories, a statistically stable observation signal dataset is obtained, denoted as: （1） where M is the number of non-in-situ detectors actually deployed or virtually set; Then, the sparse observation signal is constructed into a voxel-level tensor: （2） Each of the voxels is assigned a value of 1 if the corresponding observation point is within the voxel, and a value of 0 if the observation point is not within the voxel. voxel index locations to the corresponding observation points, with the remaining voxels filled with 0 or NaN as missing marker; Next, the designed up-sampling network UpsampleCNN is used for spatial signal reconstruction: （3） dense signal tensor reshaped into matrix form for concatenation with graph node features.

3. The graph structure based incore neutron field non-in situ explicit inversion method of claim 2, wherein, In the second step, the core space is first divided For each segment, a graph node is created The node feature vector is （4） wherein, represents a neutron flux of the i-th bin, represents a material code of the i-th bin, represents a local temperature of the i-th bin, represents a fuel enrichment of the i-th bin, and the like. Subsequently, according to the component geometry connection and physical coupling, a first-order adjacency relationship matrix between nodes is constructed If node and node are directly adjacent in space, then define ; The connection criteria between nodes can be based on geometric proximity, physical proximity, and material continuity or symmetry distribution; To represent the cross-region propagation of neutron distribution changes, higher-order adjacency relationships are further introduced; the second-order adjacency is defined as: （5） wherein, representing nodes there is a two-hop path between the nodes ; this can be extended to three-order, four-order adjacency; to integrate multi-order information, a comprehensive adjacency matrix is constructed: + +... （6） wherein, is the attenuation factor of the kth-order adjacency information, satisfying , controls the contribution weight of the neutron distribution of the far-neighbor bin to the overall structure propagation; the edge weight is defined as: （7） wherein, is the sum of geometric distances over k-hop paths, is a multi-hop physical propagation factor.

4. The graph structure based incore neutron field non-in situ explicit inversion method of claim 3, wherein, In step three, in the graph neural network, graph convolution (GCN) is used as the basic information transmission unit; the update of each layer of graph convolution follows the neighborhood feature aggregation mechanism, which is specifically expressed as: （8） wherein, denotes the adjacency matrix with self-loops added, is its degree matrix, are trainable parameters, is an activation function; The network structure further introduces multi-layer stacking, combined with residual connection and Dropout regularization; The final output node feature matrix , contains the spatial representation information of each section block after multiple rounds of feature propagation.

5. The graph structure based incore neutron field non-in situ explicit inversion method of claim 4, wherein, In the fourth step, the graph output feature matrix reshape to a tensor according to the core spatial grid layout wherein are the discrete numbers of partitions in three spatial dimensions of the core, respectively, is the number of output channels of the graph neural network. The tensor is input into a designed three-dimensional convolution network, and multi-layer convolution operation is performed to identify the spatial smoothness, gradual structure and non-uniform disturbance pattern in the local neutron field distribution; the size of the convolution kernel is generally taken , the step is 1 or 2, and a ReLU activation function is used; the final output of the network is a single-channel tensor , corresponding to the neutron flux inversion value of each block of the core.

6. The graph structure based incore neutron field non-in situ explicit inversion method of claim 5, wherein, In the fifth step, the loss function is composed of two parts: one is the prediction loss of the outer nodes , which is used to enhance the response consistency between the boundary region and the out-of-core observation data; the other is the prediction loss of the whole core range , which reflects the accuracy of the global neutron flux distribution; the joint loss function is expressed as: （9） （10） （11） where N is a set of all nodes in the core, is a set of nodes of the outermost assembly, is a coefficient of adjustment for the auxiliary loss term.

Images