3D Reconstruction Method of Radiation Field Based on GA-BP Neural Network

Abstract

The present invention relates to the field of radiation field detection, and the present invention discloses a radiation field three-dimensional reconstruction method based on a GA-BP neural network. The radiation field three-dimensional reconstruction method based on the GA-BP neural network includes the following steps: S1, based on the Geant4 simulation software, a simulation experiment is performed on the radioactive source to obtain three-dimensional radiation field simulation data; at the same time, a real experiment is performed to obtain measured data; S2, a BP neural network model is constructed; based on the GA algorithm, the BP neural network model is optimized to obtain an initial GA-BP model; S3, the training samples in the database are selected to train the initial GA-BP model to obtain an optimized GA-BP model; the remaining untrained data in the database are selected, and the optimized GA-BP model is input to complete the reconstruction of the three-dimensional radiation field. The use of the radiation field three-dimensional reconstruction method based on the GA-BP neural network can significantly improve the accuracy and speed of reconstruction, and provide strong technical support for the accurate grasp of the spatial radioactive distribution of nuclear facilities.

Description

Technical Field

[0001] The invention relates to the field of radiation field detection, in particular to a radiation field three-dimensional reconstruction method based on a GA-BP neural network. Background Art

[0002] Since the three-dimensional radiation field is an important database reflecting the actual external exposure distribution in nuclear facilities, accurately and quickly grasping the radiation dose field distribution is of great significance to the implementation of radiation protection, and radiation field reconstruction has become a research hotspot.

[0003] At present, radiation field reconstruction research mostly uses interpolation methods, such as linear interpolation, inverse distance weighted interpolation, finite element method, Kriging interpolation, etc. However, traditional methods are also subject to many limitations, such as the number of required sampling points should not be too small, the interpolation time is long, and they are mainly for two-dimensional plane reconstruction. Although some scholars have proposed to use the improved Cahn-Hilliard equation to first realize the three-dimensional radiation field inversion, the reconstruction speed is not fast enough.

[0004] Therefore, how to accurately and efficiently achieve three-dimensional radiation field reconstruction is an urgent need in the field of nuclear radiation protection, and related technical bottlenecks need to be broken through. The essence of three-dimensional reconstruction of radiation fields is to find the nonlinear relationship between the influencing factors of source terms and the release rate of source terms. As a method of artificial intelligence research, neural networks have strong nonlinear mapping capabilities by constructing multi-hidden layer learning models and performing feature learning on massive training data, thereby ultimately improving calculation accuracy. Therefore, they perform well in fitting fields where it is difficult to determine functional relationships. Some scholars have applied BP neural networks to two-dimensional reconstruction of radiation fields and achieved good results. However, the initial selection of weights and thresholds of BP neural networks is very random, which will cause the system to fall into local optimal solutions, poor prediction accuracy, instability and other problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a GA-BP neural network-based radiation field three-dimensional reconstruction method which can accurately and efficiently realize three-dimensional radiation field reconstruction and can also realize the reconstruction of three-dimensional radiation fields in a variety of experimental environments.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a radiation field three-dimensional reconstruction method based on a GA-BP neural network, comprising the following steps:

[0007] S1. Conduct simulation experiments on radiation sources based on Geant4 simulation software to obtain three-dimensional radiation field simulation data; conduct real experiments at the same time to obtain measured data; establish a database with the data generated by the simulation experiments using Geant4 simulation software and the data generated by the real experiments;

[0008] S2, constructing a BP neural network model; optimizing the BP neural network model based on the GA algorithm to obtain an initial GA-BP model;

[0009] S3. Select training samples from the database to train the initial GA-BP model to obtain an optimized GA-BP model; select the remaining untrained data from the database, input them into the optimized GA-BP model, and complete the reconstruction of the three-dimensional radiation field.

[0010] Furthermore, the method for obtaining the initial GA-BP model in step S2 includes the following steps:

[0011] S21, initial BP neural network, determine the number of input layer nodes, hidden layer number, hidden layer node number and output layer node number of the BP neural network;

[0012] S22, initial weights and thresholds, preprocessing the input data to obtain initial values, encoding the initial values ​​through the GA algorithm, and establishing a fitness function to optimize the weights and thresholds of the BP neural network;

[0013] S23, performing a crossover operation on the individuals with the highest fitness in the selected group to obtain a number of new individuals;

[0014] S24, performing mutation operation on a number of the new individuals to obtain a new population;

[0015] S25, updating the new population and performing fitness calculation, using individuals with high fitness to replace individuals with low fitness to form a new population, and the method for calculating the fitness function is:

[0016]

[0017] Among them, F fit represents the fitness function, N represents the number of training set samples, d j represents the expected output of the training sample network, O j Represents the actual output of the training sample network;

[0018] S21, repeating steps S22 to S25 until the fitness is greater than a preset threshold or the number of iterations reaches a preset number, then the iteration is stopped, and the optimal weights and thresholds are obtained to obtain an optimized BP network model.

[0019] Specifically, in step S2, a BP neural network model is constructed, and the BP neural network model includes an input layer, a hidden layer and an output layer; the number of nodes in the input layer is 3, and the output layer is a radiation dose value, and the number of nodes in the output layer is 1; the variables of the input layer and the output layer are normalized, and a mapping relationship between the influencing variables and the target variables is established.

[0020] Specifically, in step S2, the hidden layer in the BP neural network model is set to 4 layers, each layer has 40 nodes, and the activation function is the Relu function; the expression of the Relu function is as follows:

[0021] σ(x)=max(x,0);

[0022] Among them, σ(x) is the output of the hidden layer, and x is the input of the hidden layer.

[0023] Specifically, in step S2, the output layer in the BP neural network model selects the Sigmoid function, which is expressed as follows:

[0024]

[0025] in, is the output of the output layer, e is a natural constant, is the output of the last hidden layer.

[0026] Specifically, in step S3, the sampling rate of the training set in the database is selected to be 6%.

[0027] The beneficial effects of the present invention are as follows: the radiation field three-dimensional reconstruction method based on the GA-BP neural network described in the present invention can accurately reconstruct the three-dimensional radiation field under various experimental environment conditions even under low sampling rates; it has a significant reconstruction speed advantage and can complete the reconstruction of nearly 500 points in a very short time (within 0.01 seconds); the radiation field three-dimensional reconstruction method based on the GA-BP neural network described in the present invention can achieve efficient reconstruction of the radiation dose rate at any position in the three-dimensional space, and can significantly improve the accuracy and speed of reconstruction, providing strong technical support for the accurate grasp of the spatial radioactive distribution of nuclear facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is a specific operation flow chart of the radiation field three-dimensional reconstruction method based on the GA-BP neural network in an embodiment of the present invention.

[0029] Figure 2 It is a schematic diagram of the BP neural network structure in an embodiment of the present invention.

[0030] Figure 3 This is the result of 3D reconstruction in a non-shielding environment;

[0031] Indicated in the figure: Figure 3 (a) is a schematic diagram of the simulation results in a non-shielding environment; Figure 3 (b) is a schematic diagram of GA-BP network reconstruction in a non-shielding environment; Figure 3 (c) is a schematic diagram of the relative error of reconstruction in an environment without a shielding object.

[0032] Figure 4 This is the result of 3D reconstruction in a shielding environment.

[0033] Indicated in the figure: Figure 4 (a) is a schematic diagram of simulation results in a shielding environment; Figure 4 (b) is a schematic diagram of GA-BP network reconstruction in a shielding environment; Figure 4 (c) is a schematic diagram of the relative error of reconstruction in a shielding environment.

[0034] Figure 5 This is the result of 3D reconstruction under real experiments.

[0035] Indicated in the figure: Figure 5 (a) is a three-dimensional schematic diagram of a real experiment; Figure 5 (b) is a schematic diagram of GA-BP network reconstruction in a real experiment; Figure 5 (c) is a schematic diagram of the relative error of reconstruction under real experiments. DETAILED DESCRIPTION

[0036] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0037] The radiation field three-dimensional reconstruction method based on the GA-BP neural network of the present invention comprises the following steps:

[0038] S1. Conduct simulation experiments on radiation sources based on Geant4 simulation software to obtain three-dimensional radiation field simulation data; conduct real experiments at the same time to obtain measured data; establish a database with the data generated by the simulation experiments using Geant4 simulation software and the data generated by the real experiments;

[0039] S2, constructing a BP neural network model; optimizing the BP neural network model based on the GA algorithm to obtain an initial GA-BP model;

[0040] The BP neural network model includes an input layer, a hidden layer and an output layer; the number of nodes in the input layer is 3, the output layer is the radiation dose value, and the number of nodes in the output layer is 1; the variables of the input layer and the output layer are normalized to establish a mapping relationship between the influencing variables and the target variables.

[0041] The hidden layer in the BP neural network model is set to 4 layers, each layer has 40 nodes; and the activation function is the Relu function; the expression of the Relu function is as follows:

[0042] σ(x)=max(x,0);

[0043] Where σ(x) is the output of the hidden layer and x is the input of the hidden layer.

[0044] The output layer in the BP neural network model selects the Sigmoid function, which is expressed as follows:

[0045]

[0046] in, is the output of the output layer, e is a natural constant, is the output of the last hidden layer.

[0047] The structure of BP neural network is as follows Figure 2 As shown, the method for obtaining the initial GA-BP model includes the following steps:

[0048] S21, initial BP neural network, determine the number of input layer nodes, hidden layer number, hidden layer node number and output layer node number of the BP neural network;

[0049] S22, initial weights and thresholds, preprocessing the input data to obtain initial values, encoding the initial values ​​through the GA algorithm, and establishing a fitness function to optimize the weights and thresholds of the BP neural network;

[0050] S23, performing a crossover operation on the individuals with the highest fitness in the selected group to obtain a number of new individuals;

[0051] S24, performing mutation operation on a number of the new individuals to obtain a new population;

[0052] S25, updating the new population and performing fitness calculation, using individuals with high fitness to replace individuals with low fitness to form a new population, and the method for calculating the fitness function is:

[0053]

[0054] Where N is the number of samples in the training set, and d j represents the expected output of the training sample network, O j Represents the actual output of the training sample network;

[0055] S26, repeating steps S22 to S25 until the fitness is greater than a preset threshold or the number of iterations reaches a preset number, then the iteration is stopped, and the optimal weights and thresholds are obtained to obtain an optimized BP network model.

[0056] S3, select the training samples in the database to train the initial GA-BP model to obtain the optimized GA-BP model; select the remaining untrained data in the database, input it into the optimized GA-BP model, and complete the reconstruction of the three-dimensional radiation field. The sampling rate of the training set in the database is selected to be 6%.

[0057] Since the GA algorithm is used to optimize the BP neural network model in step S2, the problem that the initial selection of the weights and thresholds of the BP neural network is very random, which will cause the system to fall into a local optimal solution, poor prediction accuracy, and instability can be solved, thereby improving the accuracy of the neural network.

[0058] The radiation field three-dimensional reconstruction method based on GA-BP neural network described in the present application encodes the initial value through the GA algorithm in step S2, and establishes a fitness function to optimize the weights and thresholds of the BP neural network; effectively avoids local optimality, improves the generalization ability of the network, and enables it to accurately model under limited data. At the same time, the nonlinear mapping ability of the BP neural network ensures that sufficient information is extracted from low-sampling data to achieve high-precision three-dimensional radiation field reconstruction. In addition, the efficient implementation and computational optimization of the algorithm significantly improve the processing speed, meeting the demand for fast reconstruction.

[0059] Example

[0060] like Figure 1 As shown, the radiation field three-dimensional reconstruction method based on the GA-BP neural network in this embodiment includes the following steps:

[0061] 1. Based on the Geant4 simulation software, a simulation experiment was conducted on the radiation source to obtain three-dimensional radiation field simulation data; at the same time, a real experiment was conducted to obtain measured data; the simulation results in the non-shielding environment and the shielding environment were as follows: Figure 3 In (a), Figure 4 As shown in (a) in the figure, the measured data obtained in the real experiment is shown in Figure 5 As shown in (a) in .

[0062] In the Geant4 simulation process, the radiation source is Cs-137, whose gamma ray energy is 661.6kev and the number of particle transport is 1e8. Three-dimensional radiation field data is obtained, and real experiments are carried out to obtain measured data. In this example, there are 2947 Geant4 simulation data and 50 measured data.

[0063] 2. Construct a BP neural network model; optimize the BP neural network model based on the GA algorithm to obtain the initial GA-BP model;

[0064] 2.1. Construct a BP neural network. Since the input data is the point coordinates in three-dimensional space, the number of input layer nodes is 3, and the output layer is the radiation dose value, so the number of output layer nodes is 1. Normalize the input layer and output layer variables, and establish a mapping relationship between the influencing variables and the target variables. In this example, the hidden layer is set to 4 layers, each with 40 nodes. The BP neural network model can be seen Figure 2. The activation function introduces nonlinear elements into neurons, allowing the neural network to arbitrarily approximate any nonlinear function. Compared with other activation functions, the Relu function has stronger nonlinear fitting ability and is the default activation function with the best comprehensive performance. Therefore, this example selects the Relu function as the activation function of the hidden layer. Its expression is as follows:

[0065] σ(x)=max(x,0)

[0066] Where σ(x) is the output of the hidden layer and x is the input of the hidden layer.

[0067] Since the data is normalized during network training, and the Sigmiod activation function can convert the input to a range between 0 and 1, the output layer selects the Sigmoid function, whose expression is as follows:

[0068]

[0069] in, is the output of the output layer, e is a natural constant, is the output of the last hidden layer.

[0070] 2.2. Establishment of the initial GA-BP model, introduction of genetic algorithm and calculation of fitness include: calculation of individual fitness values ​​for the initial population, and judging whether the output error meets the termination condition based on the individual fitness value calculation; individual fitness value calculation includes fitness function of the genetic algorithm, selection function calculation of the genetic algorithm, crossover operation and mutation operation. In this embodiment, the initial population is selected as 50, the crossover probability is 0.6, and the mutation probability is 0.005. The optimal values ​​of weights and thresholds of the BP neural network are obtained through the GA genetic algorithm to complete the training; and the optimized BP network model is obtained.

[0071] 3. In this example, the training samples in the database are selected to train the initial GA-BP model to obtain the optimized GA-BP model; the remaining untrained data in the database are selected and input into the optimized GA-BP model to complete the reconstruction of the three-dimensional radiation field.

[0072] The sampling rate of the three-dimensional radiation field is 6%, that is, 6% of the data is used for neural network training, and the trained neural network is used to predict the radiation dose values ​​of the remaining 94% of the points, thereby realizing the reconstruction of the three-dimensional radiation field.

[0073] The reconstruction results under the non-shielding environment, the shielding environment and the real experiment are as follows: Figure 3 In (b), Figure 4 (b) and Figure 5 As shown in (b) in .

[0074] 4. Evaluation of reconstruction effect: Taking relative error as the evaluation index, the relative error is calculated as follows:

[0075]

[0076] Where m is the sample size, and i are the target true value and the predicted value respectively.

[0077] The errors of the reconstructed model under the non-shielding environment, the shielding environment and the real experiment are as follows: Figure 3 (c) in Figure 4 (c) and Figure 5 As shown in (c) in .

[0078] Table 1 Relative error between the data of the radiation field reconstructed by the method of the present invention and the real data

[0079] It can be seen from Table 1 that the radiation field reconstructed by the radiation field three-dimensional reconstruction method based on the GA-BP neural network described in the present invention is basically consistent with the actual three-dimensional radiation field, and a small error is achieved in both the simulation experiment and the real experiment, indicating that the radiation field three-dimensional reconstruction method proposed in the present invention is effective and accurate.

Claims

1. A 3D reconstruction method of radiation field based on GA-BP neural network, characterized in that: The following steps are involved: S1. Conduct simulation experiments on radiation sources based on Geant4 simulation software to obtain three-dimensional radiation field simulation data; conduct real experiments at the same time to obtain measured data; establish a database with the data generated by the simulation experiments using Geant4 simulation software and the data generated by the real experiments; S2, constructing a BP neural network model; optimizing the BP neural network model based on the GA algorithm to obtain an initial GA-BP model; The BP neural network model includes an input layer, a hidden layer and an output layer; the number of nodes in the input layer is 3, the output layer is a radiation dose value, and the number of nodes in the output layer is 1; the variables in the input layer and the output layer are normalized to establish a mapping relationship between the influencing variables and the target variables; The hidden layer in the BP neural network model is set to 4 layers, each layer has 40 nodes; and the activation function is the Relu function; the expression of the Relu function is as follows: σ(x)=max(x,0); Where σ(x) is the output of the hidden layer, and x is the input of the hidden layer; The output layer in the BP neural network model selects the Sigmoid function, which is expressed as follows: in, is the output of the output layer, e is a natural constant, is the output of the last hidden layer; The method for obtaining the initial GA-BP model includes the following steps: S21, initial BP neural network, determine the number of input layer nodes, hidden layer number, hidden layer node number and output layer node number of the BP neural network; S22, initial weights and thresholds, preprocessing the input data to obtain initial values, encoding the initial values ​​through the GA algorithm, and establishing a fitness function to optimize the weights and thresholds of the BP neural network; S23, performing a crossover operation on the individuals with the highest fitness in the selected group to obtain a number of new individuals; S24, performing mutation operation on a number of the new individuals to obtain a new population; S25, updating the new population and performing fitness calculation, using individuals with high fitness to replace individuals with low fitness to form a new population, and the method for calculating the fitness function is: Where N is the number of samples in the training set, and d j represents the expected output of the training sample network, O j Represents the actual output of the training sample network; S26, repeating steps S22 to S25 until the fitness is greater than a preset threshold or the number of iterations reaches a preset number, stopping the iteration, obtaining the optimal weight and threshold, and obtaining an optimized BP network model; S3. Select the training samples in the database obtained in step S1 to train the initial GA-BP model to obtain an optimized GA-BP model; select the remaining untrained data in the database, input it into the optimized GA-BP model, and complete the reconstruction of the three-dimensional radiation field.

2. The method for three-dimensional reconstruction of radiation field based on GA-BP neural network according to claim 1, characterized in that: In step S3, the sampling rate of the training set in the database is selected to be 6%.

Images