A method for real-time correction of core neutron probe signals based on a hybrid graph model

Abstract

The application provides a kind of based on mixed graph model's core neutron probe signal real-time correction method, based on existing nuclear power plant instrument control facility, utilize graph aggregation network and long short time memory network (GSA-LSTM) to build core neutron probe analysis model. First, the historical monitoring data and spatial layout information of core neutron probe are input into mixed graph model (GSA-LSTM) to complete training and deploy in industrial computer;Second, the multi-dimensional current signal generated by core neutron probe is collected, and the multi-dimensional current signal is transmitted to the signal processing cabinet through the signal core line, and the data is filtered and sent to the industrial computer;Finally, the multi-dimensional core neutron probe signal is input into the model for real-time reconstruction, and the deviation signal is corrected by using the model output to maintain the high-precision output of the signal, and more accurate reactor state monitoring information is provided for the monitoring personnel.

Description

Technical Field

[0001] This invention relates to the fields of data mining, nuclear power plants, and reactor core neutron detectors, and in particular to a real-time correction method for reactor core neutron detector signals based on a hybrid graph model. Background Technology

[0002] Core neutron detectors, as crucial nuclear measurement equipment in nuclear power plant core neutron measurement systems, are used to monitor changes in neutron flux and three-dimensional power distribution within the reactor core, playing a vital role in maintaining reactor power stability. However, during actual core operation, core neutron detectors are inevitably affected by uncertainties such as high temperature, high humidity, and high radiation in the pressure vessel environment. Prolonged operation at full power can also impact the stability of the measurement process, leading to detector malfunctions and causing the generated current signals to deviate significantly from the normal range, resulting in erroneous judgments in subsequent analysis of the overall reactor power. Therefore, real-time signal correction of core neutron detectors is a key research focus for ensuring the safety of core parameter monitoring equipment. Domestic and international research institutions have conducted fruitful research on core neutron detector status analysis and achieved some important progress, but most studies focus on the signal processing of individual core neutron detectors. However, core neutron flux measurement systems monitor multi-dimensional current signals generated by hundreds of neutron detectors within the reactor. Different reactor core neutron detectors are affected by the non-uniform distribution of neutron flux, resulting in numerical differences in the generated current signals. State analysis focusing only on a single reactor core neutron detector lacks consideration of the correlation information inherent in the multidimensional current signals obtained by the detectors at different spatial locations. This makes it impossible to simultaneously correct the signals from multiple reactor core neutron detectors online, and is unsuitable for complex, nonlinear reactor core neutron flux measurement systems.

[0003] The harsh internal environment of the reactor core often causes abnormal signals from the neutron detectors in the core, making it impossible to reflect the correct key parameters such as neutron flux. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a real-time correction method for reactor core neutron detector signals based on a hybrid graph model. Utilizing existing instrumentation and control facilities within a nuclear power plant, and using a core signal processing cabinet and an industrial control computer as carriers, this method analyzes the spatiotemporal fusion information contained in multidimensional core neutron detector signals, effectively capturing the spatial neighborhood information and time-varying dynamic characteristics of core neutron detectors distributed at different locations. First, historical monitoring data and spatial layout information of the core neutron detectors are input into a hybrid graph model (GSA-LSTM) for training and deployment on the industrial control computer. Second, multidimensional current signals generated by the core neutron detectors are collected and transmitted to the signal processing cabinet via signal core wires. After data filtering, the signals are sent to the industrial control computer. Finally, the multidimensional core neutron detector signals are input into the model for real-time reconstruction. The model output is used to correct deviations in the signal, maintaining high-precision output and providing monitoring personnel with more accurate reactor status monitoring information.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a real-time correction method for reactor core neutron detector signals based on a hybrid graph model. First, historical monitoring data and spatial layout information of the reactor core neutron detector are input into a hybrid graph model (GSA-LSTM) to complete training and deployment on an industrial control computer. Second, multidimensional current signals generated by the reactor core neutron detector are collected and transmitted to a signal processing cabinet via signal core wires. After data filtering, the signals are sent to the industrial control computer. Finally, the multidimensional reactor core neutron detector signals are input into the model for real-time reconstruction. The model output is used to correct deviations in the signal to maintain high-precision signal output, providing monitoring personnel with more accurate reactor status monitoring information.

[0006] In a preferred embodiment, the steps include:

[0007] Step 1: Obtain historical current data of the entire reactor core neutron detector;

[0008] Step 2: Construct a core neutron detector relationship diagram G = (D, V, F) based on the spatial layout information of the core neutron detector assembly;

[0009] Step 3: Treat the current signals continuously recorded by the core neutron detectors distributed at different locations as features of the corresponding nodes in the relationship graph G, and construct the core neutron detector feature matrix X∈R. N×L ;

[0010] Where N and L represent the number of neutron detectors in the reactor core and the length of the signal, respectively, and R is the set of real numbers;

[0011] Step 4: Construct spatiotemporal graph data of all core neutron detectors by combining the core neutron detector relationship graph G and the feature matrix X, and realize the representation learning of deep graph data through graph aggregation GSA network;

[0012] Step 5: The GSA network obtains the target node in the relationship graph G through random walk, and extracts k-order samples of the target node to set a fixed number of nearest neighbor nodes; where k represents the search depth of the target node;

[0013] Step 6: Use aggregation functions to aggregate nearest neighbor nodes and update the node information itself;

[0014] Step 7: Through spatial feature aggregation, node features are used for information propagation between different layers;

[0015] Step 8: Use the GSA network to establish a graph aggregation model for extracting the spatial features of the neutron detector signal in the reactor core, train the model, and obtain a converged model;

[0016] Step 9: Input the features output by the graph aggregation model into the LSTM network, and use the LSTM network to extract the time series features of the core neutron detector signal;

[0017] Step 10: The LSTM network consists of an input gate, a forget gate, and an output gate;

[0018] Step 11: The forget gate in the LSTM network eliminates c from the neurons of the previous time step using the Sigmoid activation function. t-1 Information to obtain neuron state

[0019] Step 12: Use the Sigmoid activation function to control the information g input to the neuron at time t. t Thus obtaining processed

[0020] Step 13: Discard part of the neuron information at time t-1 through the forget gate, and introduce the input information at time t through the input gate. Then the neuron completes the update. The specific calculation process is as follows:

[0021]

[0022] Where c t The output gate uses the Sigmoid activation function to output the updated neuron information;

[0023] Step 14: Obtain the output gate;

[0024] Step 15: Further combining the above GSA and LSTM network calculation process, a GSA-LSTM hybrid graph model is obtained. The GSA-LSTM hybrid graph model is considered as a mapping function F. By fusing the spatial information on graph G and the temporal features of the feature matrix X, the core neutron detector current signal at time t is reconstructed. The calculation expression is simplified to:

[0025] Y t =F[G;(X t-w ,...,X t-1 ,X t )]

[0026] Where w represents the time window width of the input signal sequence, Y t ∈R N×L This represents the reconstruction matrix of the neutron detector current signal in the reactor core;

[0027] Step 16: After steps 5 to 15, the hybrid graph model based on the core neutron detector signal is established, and the model is trained using the dataset obtained in step 1.

[0028] Step 17: Visualize model training, evaluate model performance, and perform parameter tuning;

[0029] Step 18: Evaluate model performance based on metrics such as response time and reconstruction accuracy requirements. Reconstruction accuracy is measured using the mean absolute percentage error δ. MAPE and root mean square error δ RMSE As an indicator;

[0030] Step 19: Export the trained model as an executable file and package it into the industrial control computer's application or system;

[0031] The above completes the preliminary model training and deployment work. The next step is the real-time correction stage of the neutron detector signal in the reactor core.

[0032] Step 20: Signal acquisition. The signal processing cabinet will send the acquired and filtered current data to the industrial control computer.

[0033] Step 21: The industrial control computer inputs real-time data streams into the model and outputs the corresponding signal reconstruction values;

[0034] Step 22: Calculate the residual e(t) between the model reconstructed value and the actual value at time t, and judge the signal;

[0035] Step 22: Determine the magnitude of the residual using the 3-sigma criterion to assess the degree of deviation of the core neutron detector signal;

[0036] Step 23: If the residual is less than the residual threshold limit, the core neutron detector will normally display the core neutron flux monitoring information;

[0037] Step 24: If the residual is greater than the residual threshold limit, an alarm is sent to the industrial control computer, and the reconstruction value obtained in step 21 is used to correct the core neutron detector current signal.

[0038] Step 25: Transmit the corrected information to the remote display terminal to provide monitoring personnel with more accurate core operation monitoring information.

[0039] In a preferred embodiment, in step 2, each node in the relation graph G corresponds to a single core neutron detector, D and F represent the set of nodes and edges in graph G, respectively, and V is an adjacency matrix, representing the connection relationship between each core neutron detector in graph G. V contains only two elements, 1 and 0, where 1 indicates that there is a connection relationship between core neutron detectors and 0 indicates that there is no connection relationship.

[0040] In a preferred embodiment, step 6 is specifically implemented as follows:

[0041]

[0042] in Let ψ represent the features of all nearest neighbors of node v in the k-th layer, ψ represent the aggregated information of nearest neighbors, and N(v) represent the nearest neighbors of node v. This represents the embedding information of any nearest neighbor node u of node v in layer k-1.

[0043] In a preferred embodiment, step 7 is specifically carried out as follows:

[0044]

[0045] Where σ * W represents a non-linear activation function. k Represents the learning parameters of the k-th layer. and This represents the features obtained after learning from nodes v at layers k and k-1. CON() represents the concatenation function, which connects the features of the current node with the features of the aggregated nodes.

[0046] In a preferred embodiment, the specific calculation process in step 11 is as follows:

[0047] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0048]

[0049] Where x t and f t h is the input and output of the forget gate. t-1 For the output of the hidden layer at time t-1, [h t-1 ,x t ] represents h t-1 and xt splicing, W f Let b be the weight matrix of the forget gate. f Let c be the bias matrix of the forget gate, σ be the sigmoid activation function, and c be the bias matrix of the forget gate. t-1 It represents the neuron state at time t-1, and is the Hadamard product operator.

[0050] In a preferred embodiment, the specific calculation process in step 12 is as follows:

[0051] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0052] g t =T(W g ·[h t-1 ,x t ]+b g )

[0053]

[0054] Its i t It is the output of the input gate, g t W is the candidate vector at time t. i and W g Both are the weight matrices of the input gate, b i and b g Both are the bias matrices of the input gates, and T is the tanh activation function.

[0055] In a preferred embodiment, the calculation process for obtaining the output gate in step 14 is as follows:

[0056] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0057] h t =T(c t )o t

[0058] Where h t Let o be the hidden state at time t. t It is the output of the output gate, W o It is the weight matrix of the output gate, b o It is the bias matrix of the output gate.

[0059] In a preferred embodiment, in step 18, the calculation formula is as follows:

[0060]

[0061]

[0062] Where y truth For the i-th real signal from the reactor core neutron detector, y rec Let be the i-th reconstructed value output by the model, and M be the number of samples for all signals.

[0063] In a preferred embodiment, in step 22, the formula for calculating e(t) is:

[0064]

[0065] Where y truth (t) and y re (t) represents the actual value and the reconstructed value at time t, respectively.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. This invention exemplifies a real-time correction method for neutron detector signals in reactor cores based on a hybrid graph model. This method addresses the harsh environment within a reactor, where severe deviations in neutron detector signal measurements can potentially trigger systemic accidents, posing safety risks to the equipment, personnel, and the environment. First, a Gaussian Array (GSA) network is employed to extract the spatial characteristics of the distributed neutron detector signals in the reactor core. Second, an LSTM network is used to extract the temporal characteristics of the signals. Finally, real-time monitoring of the entire reactor's neutron detector operating status is implemented, allowing for timely correction of errors when detector monitoring data deviates, thereby improving reactor operational safety.

[0068] 2. The real-time correction method for reactor core neutron detector signals based on GSA-LSTM network exemplified in this invention utilizes existing nuclear power plant instrumentation and control system devices to construct a hybrid graph model with reactor core neutron detector components in actual scenarios as the object. This effectively captures the spatiotemporal variation patterns of the overall reactor core neutron detector, enhances the complementarity of spatiotemporal modeling within the reactor, and provides technical reserves and application support for the status monitoring of nuclear power plant instrumentation systems.

[0069] 3. The real-time correction method for core neutron detector signals based on a hybrid graph model, as exemplified in this invention, overcomes the shortcomings of conventional models that only utilize time-domain data. It constructs spatiotemporal graph data for core neutron detector signals and integrates spatiotemporal joint information of multi-dimensional signals to correct the current signals of individual core neutron detectors, thus having a greater advantage in real-time tracking of the changing trends of core neutron detector signals. Attached Figure Description

[0070] Figure 1This is a flowchart illustrating the real-time correction of the neutron detector signal in the reactor core according to a preferred embodiment of the present invention.

[0071] Figure 2 This is a flowchart illustrating the computation process of the GSA-LSTM hybrid graph model according to a preferred embodiment of the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0073] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0074] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0075] This embodiment analyzes the spectral signals of multidimensional core neutron detectors in a real-world scenario. A spatiotemporal graph is constructed using the topology of all core neutron detectors within the reactor as the object, deeply exploring the spatiotemporal variation characteristics inherent in the overall neutron flux, and building a real-time signal correction model for the core neutron detectors. The signal correction model is implemented by combining graph aggregation and long short-time memory (GSA-LSTM) networks. GSA captures spatial neighborhood information of core neutron detectors at different locations, and the extracted spatiotemporal information is input into the LSTM network to reconstruct the multidimensional current signal and correct deviation signals in real time.

[0076] like Figure 1-2 As shown, this invention provides a real-time correction method for core neutron detector signals based on a hybrid graph model, comprising:

[0077] S1: Acquire the current signals of all core neutron detectors transmitted via the signal core wire through the signal processing cabinet.

[0078] S2: Based on the spatial layout of the detector components within the reactor, construct the core neutron detector relationship diagram G=(D,V,F) and the neutron detector characteristic matrix X∈R. N×L This is used to characterize the spatiotemporal information between neutron detectors in the reactor core.

[0079] S3: Initialize the parameters of the GSA-LSTM hybrid graph model.

[0080] S4: Input the dataset into the model for training, perform parameter tuning, save the optimal model, and package it.

[0081] S5: Deploy the optimal hybrid graph model to the industrial control computer.

[0082] S6: Input all core neutron detector signals to the industrial control computer in real time.

[0083] S7: The hybrid graph model in the industrial control computer performs in-depth data analysis on the signal and realizes nonlinear reconstruction of the current signal to obtain the residual between the actual value and the model reconstruction value.

[0084] S8: The residual magnitude is determined by the 3-sigma criterion to assess the degree of deviation of the neutron detector signal in the reactor core.

[0085] S9: If there is no signal deviation, the feedback signals from the core neutron detector are all normal, and the core neutron flux monitoring information is valid; if there is a deviation, the current signal is corrected in real time using the hybrid graph model reconstruction value, and accurate core operation monitoring information is provided to the remote display terminal.

[0086] In the relational graph G constructed in S2, each node corresponds to a core neutron detector. D and F represent the sets of nodes and edges in graph G, respectively. V is an adjacency matrix, representing the connection relationships between the core neutron detectors in graph G. V contains only two elements: 1 and 0. 1 indicates that there is a connection between core neutron detectors, and 0 indicates that there is no connection. X∈R N×L In this context, N and L represent the number of neutron detectors in the reactor core and the length of the signal sequence, respectively, and R is the set of real numbers.

[0087] The computation process of the GSA-LSTM hybrid graph model in S3 and S4 is as follows: Figure 2 As shown, the GSA network is responsible for representation learning of deep graph data. It obtains target nodes in the relationship graph G through random walks and extracts K-order samples of the target nodes to set a fixed number of nearest neighbor nodes. Here, K represents the search depth of the target node.

[0088] In GSA networks, aggregation functions are used to aggregate nearest neighbor nodes and update the node's own information. The specific implementation is as follows:

[0089]

[0090] In formula (1) ψ represents the features of all nearest neighbors of node v in the k-th layer; ψ represents the aggregated information of nearest neighbors; N(v) represents the nearest neighbors of node v. This represents the embedding information of any nearest neighbor node u of node v in layer k-1. The aggregated node features are then used for information propagation between different layers, as follows:

[0091]

[0092] In equation (2), σ * W represents a non-linear activation function. k Represents the learning parameters of the k-th layer. The features obtained after learning from node v in layer k-1 are represented by CON(), which represents the concatenation function.

[0093] A graph aggregation model for extracting spatial features of neutron detector signals from the reactor core was established using a GSA network. The model was then trained to obtain the optimal spatial feature extraction model.

[0094] In S4, the features output from the graph aggregation model are input into an LSTM time series network to extract the time series features of the neutron detector signal in the reactor core. The LSTM network consists of an input gate, a forget gate, and an output gate. The forget gate uses a sigmoid activation function to eliminate c from the neurons of the previous time step. t-1 Information, and then the neuron state after forgetting some information. The specific calculation process is as follows:

[0095] f t =σ(W f ·[h t-1 ,x t ]+b f (3)

[0096]

[0097] In equation (3), x t and f t h is the input and output of the forget gate. t-1 For the output of the hidden layer at time t-1, [h t-1 ,x t ] represents h t-1 and x t splicing, W f Let b be the weight matrix of the forget gate. f Let c be the bias matrix of the forget gate, σ be the Sigmoid activation function, and c be the bias matrix of the forget gate in equation (4). t-1 It represents the neuron state at time t-1, and is the Hadamard product operator.

[0098] Furthermore, the sigmoid activation function is used to control the information g input to the neuron at time t. t Thus obtaining processed The specific calculation process is as follows:

[0099] i t =σ(Wi ·[h t-1 ,x t ]+b i (5)

[0100] g t =T(W g ·[h t-1 ,x t ]+b g (6)

[0101]

[0102] In equations (5) to (7), i t It is the output of the input gate, g t W is the candidate vector at time t. i and W g Both are the weight matrices of the input gate, b i and b g Both are the bias matrices of the input gates, and T is the tanh activation function.

[0103] The LSTM network discards part of the neuron information at time t-1 through the forget gate and introduces the input information at time t through the input gate, thus completing the neuron update. The specific calculation process is as follows:

[0104]

[0105] In equation (8), c t The output gate uses the Sigmoid activation function to output the updated neuron information.

[0106] Furthermore, the calculation process for the output gate within the LSTM network is as follows:

[0107] o t =σ(W o ·[h t-1 ,x t ]+b o (9)

[0108] h t =T(c t o t (10)

[0109] In equations (9) to (10), h t Let o be the hidden state at time t. t It is the output of the output gate, W o It is the weight matrix of the output gate, b o It is the bias matrix of the output gate.

[0110] Further combining the above GSA and LSTM network calculation processes, a GSA-LSTM hybrid graph model is obtained. Treating this model as a mapping function F, and fusing the spatial information on graph G with the temporal characteristics of the feature matrix X, the core neutron detector current signal at time t is corrected. The calculation expression simplifies to:

[0111] Y t =F[G;(X t-w ,...,X t-1 ,X t (11)

[0112] In equation (11), w represents the time window width of the input signal sequence, and Y t ∈R N×L This represents the reconstruction matrix of the current signal from the neutron detector in the reactor core.

Claims

1. A method for real-time correction of core neutron detector signals based on a hybrid graphical model, characterized in that, Includes the following steps: Step 1: Obtain historical current data of the entire reactor core neutron detector; Step 2: Construct a core neutron detector relationship diagram G=(D, V, F) based on the spatial layout information of the core neutron detector assembly; Step 3: Treat the current signals continuously recorded by the core neutron detectors distributed at different locations as features of the corresponding nodes in the relationship graph G, and construct the core neutron detector feature matrix. ; Where N and L represent the number of neutron detectors in the reactor core and the length of the signal, respectively, and R is the set of real numbers; Step 4: Construct spatiotemporal graph data of all core neutron detectors by combining the core neutron detector relationship graph G and the feature matrix X, and realize the representation learning of deep graph data through graph aggregation GSA network; Step 5: The GSA network obtains the target node in the relationship graph G through random walk, and extracts k-order samples of the target node to set a fixed number of nearest neighbor nodes; where k represents the search depth of the target node; Step 6: Use aggregation functions to aggregate nearest neighbor nodes and update the node information itself; Step 7: Through spatial feature aggregation, node features are used for information propagation between different layers; Step 8: Use the GSA network to establish a graph aggregation model for extracting the spatial features of the neutron detector signal in the reactor core, train the model, and obtain a converged model; Step 9: Input the features output by the graph aggregation model into the LSTM network, and use the LSTM network to extract the time series features of the core neutron detector signal; Step 10: The LSTM network consists of an input gate, a forget gate, and an output gate; Step 11: The forget gate in the LSTM network eliminates the information from neurons in the previous time step using the Sigmoid activation function. Information to obtain neuron state ; This refers to the neuron state at time t-1; Step 12: Use the Sigmoid activation function to control the information input to the neuron at time t. Thus obtaining processed ; This indicates the processed neuron state. It is the candidate vector at time t; Step 13: Discard part of the neuron information at time t-1 through the forget gate, and introduce the input information at time t through the input gate. Then the neuron completes the update. The specific calculation process is as follows: ,in The output gate uses the Sigmoid activation function to output the updated neuron information; Step 14: Obtain the output gate; Step 15: Further combining the above GSA and LSTM network calculation process, a GSA-LSTM hybrid graph model is obtained. The GSA-LSTM hybrid graph model is considered as a mapping function F. By fusing the spatial information on graph G and the temporal features of the feature matrix X, the core neutron detector current signal at time t is reconstructed. The calculation expression is simplified to: Where w represents the time window width of the input signal sequence, This represents the reconstruction matrix of the neutron detector current signal in the reactor core; Step 16: After steps 5 to 15, the hybrid graph model based on the core neutron detector signal is established, and the model is trained using the dataset obtained in step 1. Step 17: Visualize model training, evaluate model performance, and perform parameter tuning; Step 18: Evaluate model performance based on response time and reconstruction accuracy requirements. Reconstruction accuracy is measured using the mean absolute percentage error δ. MAPE and root mean square error δ RMSE As an indicator; Step 19: Export the trained model as an executable file and package it into the industrial control computer's application or system; The above completes the preliminary model training and deployment work. The next step is the real-time correction stage of the neutron detector signal in the reactor core. Step 20: Signal acquisition. The signal processing cabinet will send the acquired and filtered current data to the industrial control computer. Step 21: The industrial control computer inputs real-time data streams into the model and outputs the corresponding signal reconstruction values; Step 22: Calculate the residual e(t) between the model reconstructed value and the actual value at time t, and judge the signal; Step 22: Determine the magnitude of the residual using the 3-sigma criterion to assess the degree of deviation of the core neutron detector signal; Step 23: If the residual is less than the residual threshold limit, the core neutron detector will normally display the core neutron flux monitoring information; Step 24: If the residual is greater than the residual threshold limit, an alarm is sent to the industrial control computer, and the reconstruction value obtained in step 21 is used to correct the core neutron detector current signal. Step 25: Transmit the corrected information to the remote display terminal to provide monitoring personnel with more accurate core operation monitoring information; In step 2, each node in the relation graph G corresponds to a single core neutron detector. D and F represent the set of nodes and edges in graph G, respectively. V is an adjacency matrix, representing the connection relationship between each core neutron detector in graph G. V contains only two elements: 1 and 0. 1 indicates that there is a connection relationship between core neutron detectors, and 0 indicates that there is no connection relationship. Step 6 is specifically implemented as follows: ,in This represents the characteristics of all nearest neighbor nodes of node v in the k-th layer. This represents the information of the aggregated nearest neighbors, where N(v) represents the nearest neighbors of node v. The embedding information of any nearest neighbor node u of node v in layer k-1; In step 7, the specific method is as follows: ,in Represents a non-linear activation function. Represents the learning parameters of the k-th layer. and This represents the features obtained after learning from nodes v at layers k and k-1. CON() represents the concatenation function, which connects the features of the current node with the features of the aggregated nodes.

2. The method for real-time correction of core neutron detector signals based on a hybrid graph model according to claim 1, characterized in that, In step 11, the specific calculation process is as follows: ,in and For the input and output of the forget gate, This represents the output of the hidden layer at time t-1. represent and splicing Here is the weight matrix for the forget gate. Let be the bias matrix of the forget gate. It is the Sigmoid activation function. It is the neuron state at time t-1. It is the Hadamard product operator.

3. The method for real-time correction of core neutron detector signals based on a hybrid graph model according to claim 1, characterized in that, In step 12, the specific calculation process is as follows: ,That It is the output of the input gate. It is the candidate vector at time t. and These are all weight matrices of the input gates. and Both are the bias matrices of the input gates, and T is the tanh activation function. It is the Sigmoid activation function. This represents the output of the hidden layer at time t-1. For the input of the forget gate, It is the Hadamard product operator.

4. The method for real-time correction of core neutron detector signals based on a hybrid graph model according to claim 1, characterized in that, In step 14, the calculation process for obtaining the output gate is as follows: ,in Let be the hidden state at time t. It is the output of the output gate. It is the weight matrix of the output gate. It is the bias matrix of the output gate. It is the Sigmoid activation function. This represents the output of the hidden layer at time t-1. For the input of the forget gate, It is the Hadamard product operator, and T is the tanh activation function.

5. The method for real-time correction of core neutron detector signals based on a hybrid graph model according to claim 1, characterized in that, In step 18, the calculation formula is as follows: in (i) represents the i-th real signal from the core neutron detector. (i) represents the i-th reconstructed value of the model output, and n represents the number of samples of all signals.

6. The method for real-time correction of core neutron detector signals based on a hybrid graph model according to claim 1, characterized in that, In step 22, the formula for calculating e(t) is: , where y truth (t) and y re (t) represents the actual value and the reconstructed value at time t, respectively.

Images