Nuclide identification method based on CNN-RNN

Abstract

The invention discloses a CNN-RNN-based nuclide identification method. The method comprises the steps of S1, gamma energy spectrum data acquisition and expansion; s2, gamma energy spectrum data preprocessing: performing smoothing and normalization processing on the energy spectrum data obtained in the step S1, mapping the processed energy spectrum data to 0-255, then transforming the processed energy spectrum data to a two-dimensional space through a Hilbert curve, generating a nuclide energy spectrum, and dividing the obtained nuclide energy spectrum data into a nuclide training set and a nuclide test set; s3, a nuclide identification model training and testing process: building a CNN-RNN model, decoding the nuclide training set obtained in the S2, and completing model training of the CNN-RNN model based on decoded data to obtain a trained CNN-RNN model; s2, testing the trained CNN-RNN model based on the nuclide test set obtained in S2, and when corresponding test indexes are met, exporting the corresponding CNN-RNN model as a nuclide identification model; and S4, completing the identification of the nuclide to be detected based on the obtained nuclide identification model. Therefore, the problem of the accuracy of nuclide identification is solved.

Description

Technical Field

[0001] The present application belongs to the field of environmental detection, and in particular relates to a CNN-RNN based nuclide identification method for effectively identifying nuclides contained in the gamma energy spectrum collected by a nuclear radiation detector. Background Art

[0002] Over the past few decades, nuclear technology has been deeply integrated with various industries, giving rise to numerous new models and formats. The 2024 "Three-Year Action Plan for the High-Quality Development of the Nuclear Technology Application Industry" aims to promote the deep integration of nuclear technology into the national economy and foster high-quality development of the nuclear technology application industry. At the same time, nuclear safety has also garnered significant attention. As a key technology in nuclear safety, nuclide identification technology has been widely applied in a variety of areas, including nuclear facility safety monitoring, nuclear material testing and analysis, and nuclear accident emergency response.

[0003] Traditional nuclide identification methods primarily rely on searching for full-energy peaks in the nuclide energy spectrum and then performing simple comparisons with a library of standard nuclides. This method focuses specifically on the peak information of characteristic peaks in the energy spectrum, including their position, peak area, and half-width. However, due to factors such as measurement time, radionuclide intensity, and statistical fluctuations, accurately capturing these peaks is difficult, leading to discrepancies in identification results. Furthermore, the nuclide identification process is complex, time-consuming, and labor-intensive, making rapid nuclide identification impossible.

[0004] In recent years, researchers have proposed novel nuclide identification methods based on full-spectrum analysis. The application of artificial intelligence technologies such as artificial neural networks, backpropagation neural networks, and convolutional neural networks in the field of nuclide identification has garnered widespread attention. These neural network models possess powerful nonlinear fitting and adaptive learning capabilities, capable of handling complex functional relationships and data distributions, and automatically adjusting parameters based on training data to accommodate diverse task requirements. However, neural network training often requires large datasets, and collecting nuclide energy spectrum data solely through detectors consumes significant time, increasing researchers' risk of exposure to radioactive materials. Furthermore, the accuracy of nuclide identification is affected by the detector's energy resolution, background noise, and interference between multiple nuclides. Summary of the Invention

[0005] The purpose of this application is to overcome the problems of the prior art and disclose a nuclide identification method based on CNN-RNN to solve the problem of the accuracy of nuclide identification.

[0006] The purpose of this application is achieved through the following technical solutions:

[0007] A nuclide identification method based on CNN-RNN, the nuclide identification method based on CNN-RNN comprising the following steps:

[0008] S1: Gamma spectrum data acquisition and expansion,

[0009] Select several radioactive sources to measure the corresponding nuclide gamma spectrum, and use random sampling to expand the spectrum data;

[0010] S2: Gamma spectrum data preprocessing: smoothing and normalizing the spectrum data obtained in step S1, and mapping the normalized one-dimensional nuclide gamma spectrum data to 0 to 255. Then, transforming it into a two-dimensional space through the Hilbert curve to generate a nuclide spectrum. The obtained nuclide spectrum data are divided into a nuclide training set and a nuclide test set.

[0011] S3: Nuclide identification model training and testing process,

[0012] Build a CNN-RNN model, decode the nuclide training set obtained by S2, and complete the model training of the CNN-RNN model based on the decoded data to obtain a trained CNN-RNN model;

[0013] Based on the nuclide test set obtained in S2, the trained CNN-RNN model is tested. When the corresponding test indicators are met, the corresponding CNN-RNN model is derived as the nuclide recognition model;

[0014] S4: Complete the identification of the nuclide to be tested based on the obtained nuclide identification model.

[0015] According to a preferred embodiment, the gamma spectrum data acquisition process in step S1 includes: selecting several radioactive sources to perform nuclide gamma spectrum measurements under single source, mixed source, different measurement distances, different measurement times and different measurement angles.

[0016] According to a preferred embodiment, the gamma spectrum data expansion process in step S1 includes: performing integral normalization processing on the original energy spectrum of each nuclide obtained by measurement to obtain the probability density of the nuclide gamma photons at different energies, and then extracting the gamma spectrum of the nuclide with different counts by random sampling.

[0017] According to a preferred embodiment, in step S2, the five-point centroid method is used to smooth the radionuclide gamma spectrum data; and the smoothed radionuclide gamma spectrum data is normalized by the maximum-minimum method.

[0018] According to a preferred embodiment, in step S3, based on the requirements of the nuclide identification task and the data characteristics, the pytorch framework is selected to build the CNN-RNN model.

[0019] According to a preferred embodiment, the CNN-RNN model includes two identical convolutional modules, a gated recurrent unit GRU, a fully connected layer and an output layer;

[0020] The convolution module is used to extract the features of the energy spectrum; the gated recurrent unit (GRU) is used to extract the sequence information of the energy spectrum features and avoid gradient vanishing and gradient explosion during model training; the fully connected layer is used to reduce the dimensionality of the features, and Dropout is introduced with the parameter set to 0.1 to prevent overfitting; the output layer activation function is Sigmoid, and the output categories are 8.

[0021] According to a preferred embodiment, the convolution module is provided with a convolution layer, an activation function and a maximum pooling layer in sequence, the convolution kernel size in the convolution layer is 5x5, the activation function adopts relu, and the maximum pooling layer is used to reduce the spatial dimension of the feature map.

[0022] According to a preferred embodiment, during the training process of the CNN-RNN model, the hyperparameter training rounds and the initial learning rate are first set, and then the parameters are optimized by the gradient descent method to find the optimal solution;

[0023] After each round of training, first calculate the total time consumed by model training, then calculate the model performance F1 score of the validation set, and compare the current model performance F1 score with the current optimal model performance F1 score to find the optimal model in the training process and save the weight file of the optimal model.

[0024] The aforementioned main solution of this application and its further options can be freely combined to form multiple solutions, all of which can be adopted and protected by this application. After understanding the solution of this application, those skilled in the art will understand that there are many combinations based on existing technology and common knowledge, all of which are technical solutions to be protected by this application, and these are not exhaustive here.

[0025] Beneficial effects of this application:

[0026] (1) The one-dimensional nuclide gamma spectrum data is converted into a two-dimensional nuclide gamma spectrum, which utilizes all the information of the spectrum. This solves the shortcoming of traditional nuclide identification methods that only utilize part of the spectrum information. It can effectively suppress the impact of temperature drift on nuclide spectrum identification and improve the robustness of nuclide identification. Moreover, the Hilbert curve uses a unique mapping method to ensure that adjacent points in the spectrum sequence can also maintain adjacent or close positional relationships in the two-dimensional spectrum. This can better preserve the spatial locality of the spectrum sequence and help deep learning models better identify and distinguish different nuclides.

[0027] (2) The CNN-RNN model is applied to the field of nuclide identification. The convolutional neural network (CNN) automatically extracts local features in the energy spectrum data, and the fusion recurrent neural network (RNN) further extracts the sequence information of the energy spectrum features, which can more completely represent the correlation in the gamma energy spectrum data and improve the accuracy of nuclide identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flowchart of the nuclide identification method based on CNN-RNN in this application;

[0029] Figure 2 This is a schematic diagram of the CNN-RNN model of this application;

[0030] Figure 3 It is the original one-dimensional gamma spectrum of nuclide Co60;

[0031] Figure 4 This is the gamma spectrum of the nuclide Co60 obtained by probability sampling in this application;

[0032] Figure 5 This is the gamma spectrum of the nuclide Co60 obtained by the five-point smoothing method in this application;

[0033] Figure 6 This is the gamma energy spectrum of the nuclide Co60 obtained by Hilbert transform in this application. DETAILED DESCRIPTION

[0034] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0035] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0036] Example 1

[0037] refer to Figures 1 to 6As shown, this embodiment discloses a nuclide identification method based on CNN-RNN, which includes the following four stages: (1) gamma spectrum data acquisition and expansion, selecting several radioactive sources to measure the corresponding nuclide gamma spectrum, and using random sampling to expand the spectrum data; (2) gamma spectrum data preprocessing, smoothing and normalizing the spectrum data obtained in stage (1), and mapping the normalized one-dimensional nuclide gamma spectrum data to 0 to 255, and then transforming it into a two-dimensional space through the Hilbert curve to generate a nuclide spectrum diagram, and the obtained nuclide spectrum diagrams are converted into the nuclide spectrum diagram. The data is divided into a nuclide training set and a nuclide test set; (3) the nuclide identification model is trained and tested, a CNN-RNN model is built, the nuclide training set obtained in stage (2) is decoded, and the model training of the CNN-RNN model is completed based on the decoded data to obtain a trained CNN-RNN model; based on the nuclide test set obtained in (2), the trained CNN-RNN model is tested, and when the corresponding test indicators are met, the corresponding CNN-RNN model is derived as a nuclide identification model; (4) the identification of the nuclide to be tested is completed based on the obtained nuclide identification model.

[0038] Specifically, (1) the gamma spectrum data acquisition and expansion process includes:

[0039] The first step is to measure the original nuclide gamma spectrum. Using the NaI detector provided by the laboratory, the gamma spectrum data of Cs137, Co60, Eu152, Cs137+Co60, Cs137+Eu152, Co60+Eu152, Cs137+Co60+Eu152, and the background environment are measured at different measurement times, distances, and angles.

[0040] In the second step, x1 is randomly divided into x2 and x3 according to different nuclides, and the proportions of the two in x1 are 4 / 5 and 1 / 5 respectively.

[0041] The third step is to construct a nuclide gamma spectrum training dataset x4 based on x2. First, all gamma spectra in x2 are normalized by class weights, so that the probability density of gamma photons at different energies can be calculated. The formula is as follows:

[0042]

[0043] Where: Ei represents a specific energy value; C(Ei) represents the count value under the energy value Ei, and λ_i represents the desired probability density.

[0044] Then, based on the calculated probability densities of gamma photons at different energies, a random sampling method was used, with the ratio of the total count of the sampled gamma spectrum to the total count of the original spectrum set to [0.5, 5]. This method acquired four gamma spectrum data sets with different counts. This expanded the gamma spectrum data and addressed the issue of insufficient actual data collection.

[0045] The fourth step is to group the energy spectrum data set by nuclide type. The gamma energy spectra of the same nuclides in x3 and x4 are grouped into the same category, resulting in x5 and x6.

[0046] (2) The gamma spectrum data preprocessing process includes:

[0047] In the first step, the gamma spectrum data sets x5 and x6 are obtained through the gamma spectrum data acquisition process. All gamma spectra in the data sets are smoothed to obtain smoothed gamma spectrum data sets x7 and x8. The smoothing method is the five-point smoothing method, and the formula is as follows: ～ _i=1 / 16(y_(i-2)+4y_(i-1)+6y_i+4y_(i+1)+y_(i+2))

[0048] Where: y ～ _i represents the smoothed count value.

[0049] The second step is to normalize x7 and x8 to obtain x9 and x10. The normalization method is the maximum and minimum method. The formula is as follows:

[0050] X_norm=(X-X_min) / (X_max-X_min)

[0051] Where: X represents the original data, X_min represents the minimum count, X_max represents the maximum count, and X_norm represents the normalized data.

[0052] The third step is to convert the normalized one-dimensional energy spectrum data into a two-dimensional energy spectrum. First, map each energy spectrum data in x9 and x10 to 0 to 255 to obtain x11 and x12. The formula is as follows:

[0053] X ～ =(X-X_min) / (X_max-X_min)*255

[0054] Where: X ～ Represents the one-dimensional energy spectrum data after mapping.

[0055] Then, x11 and x12 are converted into two-dimensional energy spectrum grayscale images x13 and x14 through Hilbert transform. x13 is used as the test set data for the evaluation of the subsequent model; x14 is used as the training set data for the training of the subsequent model. Figure 6 shown.

[0056] The one-dimensional nuclide gamma energy spectrum data is converted into a two-dimensional nuclide gamma energy spectrum diagram, and all the information of the energy spectrum is applied. This solves the shortcoming of the traditional nuclide identification method that only uses part of the energy spectrum information. It can effectively suppress the influence of temperature drift on nuclide energy spectrum identification and improve the robustness of nuclide identification.

[0057] At the same time, the Hilbert curve uses a unique mapping method to ensure that adjacent points in the energy spectrum sequence can maintain adjacent or close positional relationships in the two-dimensional energy spectrum diagram. This can better preserve the spatial locality of the energy spectrum sequence and help deep learning models better identify and distinguish different nuclides.

[0058] (3) The training and testing process of the nuclide identification model includes:

[0059] In the first step, based on the requirements of the radionuclide identification task and data characteristics, the pytorch framework was used to build the CNN-RNN model.

[0060] The second step is to define the function mean_std, which is used to calculate the mean and variance of a single-channel grayscale image.

[0061] The third step is to create an image data processing method object c1, which is used to convert the image data into a tensor and perform normalization. The mean and variance required for normalization are obtained by calling the mean_std function.

[0062] The fourth step is to create a dataset object for loading and parsing the image dataset. The ImageFolder class is called, the path is set to the path of the dataset x14, and c1 is called as the data processing method. Labels are automatically assigned to each image based on the names of the subfolders containing the different nuclide energy spectra in x14. Finally, a processed image dataset object s1 is returned.

[0063] The fifth step is to create a training set and a validation set object for training and validating the neural network model. S1 is randomly divided into training set S2 and validation set S3, and the proportions of the two in S1 are 4 / 5 and 1 / 5 respectively.

[0064] The sixth step is to build a CNN-RNN model. The model consists of two identical convolutional modules, a gated recurrent unit (GRU), a fully connected layer, and an output layer. The purpose of the convolutional module is to extract the features of the energy spectrum map. The first layer is a convolutional layer (Conv2d) with a convolution kernel size of 5x5 and an activation function of ReLU, followed by a maximum pooling layer to reduce the spatial dimension of the feature map. The purpose of the GRU is to further extract the sequence information of the energy spectrum map features while avoiding the problems of gradient vanishing and gradient exploding during model training. The purpose of the fully connected layer is to reduce the dimensionality of the features. To prevent overfitting, Dropout is introduced and the parameter is set to 0.1. The activation function of the output layer is Sigmoid, and there are 8 output categories.

[0065] In the seventh step, define two functions: time_use and f1_score, which are used to calculate the time consumed by model training and the F1 score of judging model performance respectively.

[0066] Step 8: Set the device used for model training to the computer's GPU.

[0067] In the ninth step, Adam is set as the optimizer for model training and cross entropy is set as the loss function for model training.

[0068] In the tenth step, we set the number of hyperparameter training rounds and the initial learning rate. During model training, we optimize the parameters using gradient descent to find the optimal solution. After each round of training, we first calculate the total time consumed by model training. We then calculate the F1 score of the validation set and compare the current F1 score with the current best F1 score. This method is used to find the optimal model during training and save the weight file of the optimal model.

[0069] In the eleventh step, load the weight file exported from the training model, test the single-source and mixed-source data in the test set data x13 respectively, and analyze the model effect through precision, recall rate and F1 score.

[0070] (4) Complete the identification of the nuclide to be tested based on the obtained nuclide identification model.

[0071] This application is based on the CNN-RNN nuclide identification method, which integrates convolutional neural networks and recurrent neural networks. It uses the translation invariance of CNN in time and space to automatically extract nuclide energy spectrum features. Combined with the memory capacity of RNN, it further captures the sequence information in the features and improves the accuracy of nuclide identification.

[0072] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A nuclide identification method based on CNN-RNN, characterized in that: The nuclide identification method based on CNN-RNN includes the following steps: S1: Gamma spectrum data acquisition and expansion, Select several radioactive sources to measure the corresponding nuclide gamma spectrum, and use random sampling to expand the spectrum data; S2: Gamma spectrum data preprocessing: smoothing and normalizing the spectrum data obtained in step S1, and mapping the normalized one-dimensional nuclide gamma spectrum data to 0 to 255. Then, transforming it into a two-dimensional space through the Hilbert curve to generate a nuclide spectrum. The obtained nuclide spectrum data are divided into a nuclide training set and a nuclide test set. S3: Nuclide identification model training and testing process, Build a CNN-RNN model, decode the nuclide training set obtained by S2, and complete the model training of the CNN-RNN model based on the decoded data to obtain a trained CNN-RNN model; Based on the nuclide test set obtained in S2, the trained CNN-RNN model is tested. When the corresponding test indicators are met, the corresponding CNN-RNN model is derived as the nuclide recognition model; S4: Complete the identification of the nuclide to be tested based on the obtained nuclide identification model.

2. The nuclide identification method of CNN-RNN according to claim 1, wherein The gamma spectrum data acquisition process in step S1 includes: selecting several radioactive sources to perform nuclide gamma spectrum measurements under single source, mixed source, different measurement distances, different measurement times and different measurement angles.

3. The nuclide identification method of CNN-RNN according to claim 2, characterized in that: The gamma spectrum data expansion process in step S1 includes: performing integral normalization processing on the original energy spectrum of each nuclide obtained by measurement to obtain the probability density of the nuclide gamma photons at different energies, and then extracting the gamma spectrum of the nuclide with different counts by random sampling.

4. The nuclide identification method of CNN-RNN according to claim 1, wherein In step S2, the five-point centroid method is used to smooth the radionuclide gamma spectrum data; and the smoothed radionuclide gamma spectrum data is normalized by the maximum-minimum method.

5. The nuclide identification method of CNN-RNN according to claim 1, wherein: In step S3, based on the requirements of the nuclide identification task and the data characteristics, the pytorch framework is selected to build the CNN-RNN model.

6. The nuclide identification method of CNN-RNN according to claim 5, characterized in that: The CNN-RNN model includes two identical convolutional modules, a gated recurrent unit (GRU), a fully connected layer, and an output layer. The convolution module is used to extract the features of the energy spectrum; the gated recurrent unit (GRU) is used to extract the sequence information of the energy spectrum features and avoid gradient vanishing and gradient explosion during model training; the fully connected layer is used to reduce the dimensionality of the features, and Dropout is introduced with the parameter set to 0.1 to prevent overfitting; the output layer activation function is Sigmoid, and the output categories are 8.

7. The nuclide identification method of CNN-RNN according to claim 6, characterized in that: The convolution module is sequentially provided with a convolution layer, an activation function and a maximum pooling layer. The convolution kernel size in the convolution layer is 5x5, the activation function adopts relu, and the maximum pooling layer is used to reduce the spatial dimension of the feature map.

8. The nuclide identification method of CNN-RNN according to claim 6, wherein: During the training of the CNN-RNN model, the hyperparameter training rounds and initial learning rate are first set, and then the parameters are optimized by gradient descent to find the optimal solution; After each round of training, first calculate the total time consumed by model training, then calculate the model performance F1 score of the validation set, and compare the current model performance F1 score with the current optimal model performance F1 score to find the optimal model in the training process and save the weight file of the optimal model.

Images