Efficient nuclide identification method based on multi-label hybrid neural network
Through the nuclide recognition method based on multi-label hybrid neural network, the problem of large and slow computing in complex environments is solved, and fast and accurate nuclide recognition is achieved, which improves the recognition efficiency.
Patent Information
- Application Number
- CN202411626380.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-14
- Publication Date
- 2025-03-18
AI Technical Summary
Traditional nuclide recognition methods have large calculations, slow speed and low efficiency in complex monitoring environments, making it difficult to quickly identify radionuclides in the gamma energy spectrum.
Using an efficient nuclide recognition method based on multi-label hybrid neural network (ML_HNN), the γ energy spectrum data source is obtained through the Monte Carlo method, data preprocessing and feature extraction are carried out, and neural network structures of CNN feature extraction layer, full connection layer and output layer are constructed for training and recognition.
It realizes rapid identification of radionuclides, simplifies the processing process, reduces the complexity of nuclide recognition, and improves recognition efficiency and accuracy.
Smart Images
Figure CN119646479A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of nuclide identification of nuclear power plant radiation monitoring equipment, and in particular relates to an efficient nuclide identification method based on a multi-label hybrid neural network (ML_HNN). Background Art
[0002] Various devices in the radiation monitoring system of a nuclear power plant need to judge the leakage of the nuclear power plant process or the radioactivity level in the area based on performance indicators such as nuclide type and specific activity. Therefore, the nuclide identification method is a technical method widely studied in the field of radiation monitoring research.
[0003] Due to the complexity of the monitoring environment, traditional nuclide identification methods require background testing, energy calibration, and long-term measurement in each experiment to form effective energy spectrum data. The peak search calculation is large and its application has certain limitations.
[0004] Commonly used methods include derivative peak search method and symmetrical zero area peak search method. Among them, the derivative peak search method has advantages in peak boundary determination and peak position determination, and has good ability to screen weak γ peaks and repeated peaks from background noise, and has the ability to distinguish repeated peaks, find weak peaks and suppress false peaks. However, it is greatly affected by the statistical fluctuation of the energy spectrum, and the peak search operation is large, the speed is slow, and the efficiency is low. The symmetrical zero area peak search method has good ability to screen weak γ peaks from background noise, but the recognition effect is poor in the recognition of low energy peaks and overlapping peaks. Summary of the invention
[0005] The purpose of the present invention is to provide an efficient nuclide identification method based on a multi-label hybrid neural network, which can quickly identify radioactive nuclides in a gamma energy spectrum.
[0006] The technical solution of the present invention is as follows: an efficient nuclide identification method based on a multi-label hybrid neural network comprises the following steps:
[0007] Step 1: Obtain the gamma spectrum data source based on the Monte Carlo method;
[0008] Step 2: Data preprocessing based on standardized scheme;
[0009] Step 3: Construct a neural network based on a multi-label hybrid neural network;
[0010] Step 4: Input the preprocessed nuclide energy spectrum data into the multi-label hybrid neural network for nuclide identification.
[0011] In the step 1, the gamma energy spectrum of the radionuclide is obtained as a data source, and the detector model is modeled based on the Monte Carlo method.
[0012] In the step 1, by changing the distance from the radiation source to the detector in an increment of 0.5 cm within the range of 0 to 10 cm, a set of energy spectra is generated, thereby generating multiple sets of nuclide combinations.
[0013] In step 1, select 241 Am, 133 8. 140 8. 139 Ce, 58 Co. 60 Co. 51 Cr, 137 Cs, 152 Eu, 18 F. 59 Fe, 131 I. 40 K. 85 Kr, 54 Mn, 22 Na, 113 Sn, 133 Xe, 88 Y19 single nuclides and random combinations of 19 nuclides are used to obtain the γ energy spectrum data of different radioactive nuclides as the input of the neural network.
[0014] In the step 2, the pre-processed data is standardized, and the energy spectrum is standardized using Normalizer.
[0015] In step 2, the shuffle function is used to shuffle the data so that the training data is randomly sorted, and then the data is divided into a training set, a validation set, and a test set using a random stratified sampling method to ensure the consistency of the data in each set.
[0016] The multi-label hybrid neural network structure in step 3 includes a CNN feature extraction layer, a fully connected layer and an output layer.
[0017] The CNN feature extraction layer consists of three CNN layers and a pooling layer. The number of channels of the three CNN layers is set to 256, 128, and 64 respectively. The Relu function is used for activation and the convolution kernel is initialized by a standard Gaussian distribution. The size of the pooling layer is 2 and the stride is 2. The input energy spectrum data is labeled and one-hot encoded. The data set labels are processed into the label form of a multi-label classification network so that each node of the network output corresponds to a category. The energy spectrum and one-hot encoding constitute the input features of the network.
[0018] Two layers of 5×1 one-dimensional convolution are performed on the input features to extract convolution features, and then the input features and convolution features are merged into the features extracted by the CNN layer and input into the fully connected layer for feature integration. The number of nodes is set to 128, 128, and K, respectively, where K is the number of nuclide categories. At the same time, a linear rectification function is used to activate and initialize the nodes with a standard Gaussian distribution, and the dropout is set to 0.5 to prevent overfitting;
[0019] Finally, the features integrated by the fully connected layer are classified using the activation function, and the classification output is finally obtained to obtain the nuclide identification result.
[0020] In the step 4, the preprocessed nuclide energy spectrum data is input into the multi-label hybrid neural network training result for nuclide identification.
[0021] In step 4, precision, recall, F1 value and receiver operating characteristic curve are used to evaluate the performance of the neural network nuclide recognition; if the data instance is a positive class and is predicted to be a positive class, it is a true positive class; if the data instance is a negative class and is predicted to be a negative class, it is a true negative class; if the data instance is a negative class and is predicted to be a positive class, it is a false positive class; if the data instance is a positive class and is predicted to be a negative class, it is a false negative class. The precision, recall, F1 value and receiver operating characteristic curve are defined, where precision refers to the correct proportion of samples predicted to be positive in the sample, that is, Recall rate refers to the proportion of positive samples in the sample that are predicted correctly, that is, The F1 value represents the weighted harmonic mean of recall and precision, that is, F1Score reaches its best value at 1 and its worst value at 0; ROC curve, with FPR as the horizontal axis With TPR as the vertical axis
[0022] The beneficial effects of the present invention are as follows: the present invention utilizes full spectrum information for training and learning, and realizes rapid nuclide identification through learning and training of spectrum data, specifically including the process of acquiring the gamma energy spectrum data source, data preprocessing to acquire target samples, feature extraction, neural network training and identification, and finally realizes rapid identification of radioactive nuclides. The invention is different from the traditional idea of nuclide identification through peak finding, and no longer requires operations such as peak finding, smoothing, and processing of complex overlapping peaks, thereby simplifying the processing flow, reducing the complexity of nuclide identification, improving identification efficiency, and at the same time improving identification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 A schematic diagram of a process flow of an efficient nuclide identification method based on a multi-label hybrid neural network provided by the present invention;
[0024] Figure 2 This is a schematic diagram of the simulated energy spectrum of mixed nuclides;
[0025] Figure 3 This is a schematic diagram of the ML_HNN neural network architecture;
[0026] Figure 4 This is the ML_HNN ROC curve. DETAILED DESCRIPTION
[0027] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0028] like Figure 1 As shown, the present invention provides an efficient nuclide identification method based on a multi-label hybrid neural network, and the identification method process is divided into four parts: γ spectrum data source acquisition, data preprocessing, feature extraction, neural network training and identification. The first step is to obtain the γ spectrum of radioactive nuclides as the data source based on the Monte Carlo method. The second step is to perform preprocessing operations such as standardization and data set shuffling on the γ spectrum data set to reduce the count value of the spectrum and improve the generalization ability of the network. The third step is to extract the features of the preprocessed spectrum data set as the input of the neural network. The fourth step is to send the feature input to the neural network for training. After the training is completed, the nuclide signal can be recognized. Technical indicators include: nuclide identification types ≥ 8, nuclide identification accuracy ≥ 85%.
[0029] An efficient nuclide identification method based on a multi-label hybrid neural network specifically comprises the following steps:
[0030] Step 1: Obtain gamma spectrum data source based on Monte Carlo method
[0031] like Figure 2 As shown in the figure, in order to obtain the γ spectrum of radioactive nuclides as the data source, the detector model is modeled based on the MCNP (Monte Carlo N-Particle Transport code) Monte Carlo method. Since the same input card and the same conditions of MCNP are used for training, the output data will not change. Therefore, in order to obtain more energy spectrum data to support training, the distance from the radiation source to the detector is changed, and a set of energy spectra is generated by increasing 0.5 cm each time within the range of (0 to 10) cm, thereby generating multiple groups of nuclide combinations. 241 Am, 133 8. 140 8. 139 Ce, 58 Co. 60 Co. 51 Cr, 137 Cs, 152 Eu, 18 F. 59 Fe, 131 I.40 K. 85 Kr, 54 Mn, 22 Na, 113 Sn, 133 Xe, 88 Y19 single nuclides and random combinations of 19 nuclides were used to obtain the γ energy spectrum data of different radioactive nuclides as the input of the neural network. The list of nuclides selected is shown in Table 1.
[0032] Table 1 List of nuclides selected based on MCNP
[0033]
[0034]
[0035] Step 2: Data preprocessing based on a standardized scheme
[0036] Due to factors such as different nuclide activities and different collection times, the average count values of the spectrum range from thousands to tens of thousands. The existence of such data will increase the training time and may not converge. In order to prevent and eliminate the influence of singular sample data, the preprocessed data needs to be standardized before training to ensure the comparability between the data.
[0037] Commonly used normalization schemes include Normalizer, RobustScaler, Min-Max, StandardScaler, MaxabsScaler, QuantileTransformer, etc. By comparing the processing effects of different normalization schemes on nuclide energy spectrum data, it can be seen that the Normalizer normalization method retains characteristic information such as peak position, especially the characteristic peaks of low-activity nuclides, and has a good compression effect. Therefore, Normalizer is used to standardize the energy spectrum.
[0038] In order to increase the generalization ability of the model, the shuffle function is used to shuffle the data so that the training data is randomly sorted, and then the random stratified sampling method is used to divide the data into training set, validation set, and test set to ensure the consistency of the data in each set.
[0039] Step 3: Build a neural network based on multi-label hybrid neural network
[0040] The network structure is as follows Figure 3As shown in the figure, the multi-label hybrid neural network structure mainly includes CNN feature extraction layer, fully connected layer and output layer. The CNN feature extraction layer is mainly composed of three CNN layers and pooling layer. The number of channels of the three CNN layers is set to 256, 128 and 64 respectively. The Relu function is used for activation and the convolution kernel is initialized by the standard Gaussian distribution. The size of the pooling layer is 2 and the stride is 2. The input energy spectrum data is processed with labels and one-hot encoding is performed. The labels of the data set are processed into the label form of the multi-label classification network, so that each node of the network output corresponds to a category. The spectrum and one-hot encoding constitute the input features of the network.
[0041] Then, two layers of 5×1 one-dimensional convolution are performed on the input features to extract convolution features (Convfeatures), and then the input features (Input features) and convolution features (Conv features) are merged into the features extracted by the CNN layer and input into the fully connected layer to integrate the features. The number of nodes is set to 128, 128, and K, respectively, where K is the number of nuclide categories. At the same time, the linear rectified function (Rectified Linear Unit, Relu function) is used to activate and initialize the nodes with a standard Gaussian distribution. The dropout is set to 0.5 to prevent overfitting.
[0042] Finally, the features integrated by the fully connected layer are classified using the activation function (Sigmoid function), and the classification output sigmoid (fcout) is obtained, and finally the nuclide identification result is obtained.
[0043] The source code for building a multi-label hybrid neural network is:
[0044] Step 4: Input the preprocessed nuclide spectrum data into the multi-label hybrid neural network for nuclide identification
[0045] Based on the energy spectrum data of the nuclides listed in Table 1, the preprocessed nuclide energy spectrum data is input into the multi-label hybrid neural network training result for nuclide identification. The present invention uses precision P (precision), recall R (recall), F1 value (F1 score) and receiver operating characteristic (ROC) curve to evaluate the performance of neural network nuclide identification. Before clarifying these indicators, it is necessary to clarify 4 basic concepts: if the data instance is a positive class and is predicted to be a positive class, it is a true positive class (True Positive, TP); if the data instance is a negative class and is predicted to be a negative class, it is a true negative class (True Negative, TN); if the data instance is a negative class and is predicted to be a positive class, it is a false positive class (False Negative, FP); if the data instance is a positive class and is predicted to be a negative class, it is a false negative class (False Negative, FN). After clarifying the four basic concepts, we define precision, recall, F1 score, and receiver operating characteristic (ROC). Precision refers to the correct proportion of samples predicted as positive in the sample, that is, The larger the average value, the better the performance of the algorithm model; recall refers to the proportion of positive samples in the sample that are predicted correctly, that is, The larger the average value, the better the performance of the algorithm model; the F1 value (F1score) represents the weighted harmonic mean of recall rate and precision rate, that is, F1Score reaches its best value at 1 (perfect precision and recall) and its worst value at 0; ROC curve, with FPR as the horizontal axis With TPR as the vertical axis Generally, the better the performance of the network model, the larger the TPR, the better, and the smaller the FPR, the better. This is reflected in the ROC curve as the TPR is higher and the FPR is further to the left.
[0046] Table 2 is a list of multi-label hybrid neural network classifications. Figure 4 It is the ROC curve of multi-label hybrid neural network nuclide identification. As can be seen from Table 2, the average value of precision is 99.63%; the average value of recall is 97.42%; the average value of F1 score is 98.53%. Figure 4It can be seen that the average value reaches 99.81%. The above actual data show that the nuclide identification method based on multi-label hybrid neural network proposed in the present invention has good nuclide identification performance.
[0047] Table 2 Multi-label hybrid neural network classification report
[0048]
[0049]
Claims
1. An efficient nuclide identification method based on multi-label hybrid neural network, characterized in that: The steps include: Step 1: Obtain the gamma spectrum data source based on the Monte Carlo method; Step 2: Data preprocessing based on standardized scheme; Step 3: Construct a neural network based on a multi-label hybrid neural network; Step 4: Input the preprocessed nuclide energy spectrum data into the multi-label hybrid neural network for nuclide identification.
2. The method for efficient nuclide identification based on a multi-label hybrid neural network according to claim 1, characterized in that: In the step 1, the gamma energy spectrum of the radionuclide is obtained as a data source, and the detector model is modeled based on the Monte Carlo method.
3. The method for efficient nuclide identification based on a multi-label hybrid neural network according to claim 2, characterized in that: In the step 1, by changing the distance from the radiation source to the detector in an increment of 0.5 cm within the range of 0 to 10 cm, a set of energy spectra is generated, thereby generating multiple sets of nuclide combinations.
4. The method for efficient nuclide identification based on a multi-label hybrid neural network according to claim 3, characterized in that: In step 1, select 241 Am, 133 8. 140 8. 139 Ce, 58 Co. 60 Co. 51 Cr, 137 Cs, 152 Eu, 18 F. 59 Fe, 131 I. 40 K. 85 Kr, 54 Mn, 22 Na, 113 Sn, 133 Xe, 88 Y19 single nuclides and random combinations of 19 nuclides are used to obtain the γ energy spectrum data of different radioactive nuclides as the input of the neural network.
5. The method for efficient nuclide identification based on a multi-label hybrid neural network according to claim 1, characterized in that: In the step 2, the pre-processed data is standardized, and the energy spectrum is standardized using Normalizer.
6. The method for efficient nuclide identification based on a multi-label hybrid neural network according to claim 5, characterized in that: In step 2, the shuffle function is used to shuffle the data so that the training data is randomly sorted, and then the data is divided into a training set, a validation set, and a test set using a random stratified sampling method to ensure the consistency of the data in each set.
7. The efficient nuclide identification method based on multi-label hybrid neural network according to claim 1, characterized in that: The multi-label hybrid neural network structure in step 3 includes a CNN feature extraction layer, a fully connected layer and an output layer.
8. The method for efficient nuclide identification based on a multi-label hybrid neural network according to claim 7, characterized in that: The CNN feature extraction layer consists of three CNN layers and a pooling layer. The number of channels of the three CNN layers is set to 256, 128, and 64 respectively. The Relu function is used for activation and the convolution kernel is initialized by a standard Gaussian distribution. The size of the pooling layer is 2 and the stride is 2. The input energy spectrum data is labeled and one-hot encoded. The data set labels are processed into the label form of a multi-label classification network so that each node of the network output corresponds to a category. The energy spectrum and one-hot encoding constitute the input features of the network. Two layers of 5×1 one-dimensional convolution are performed on the input features to extract convolution features, and then the input features and convolution features are merged into the features extracted by the CNN layer and input into the fully connected layer for feature integration. The number of nodes is set to 128, 128, and K, respectively, where K is the number of nuclide categories. At the same time, a linear rectification function is used to activate and initialize the nodes with a standard Gaussian distribution, and the dropout is set to 0.5 to prevent overfitting; Finally, the features integrated by the fully connected layer are classified using the activation function, and the classification output is finally obtained to obtain the nuclide identification result.
9. The method for efficient nuclide identification based on a multi-label hybrid neural network according to claim 1, characterized in that: In the step 4, the preprocessed nuclide energy spectrum data is input into the multi-label hybrid neural network training result for nuclide identification.
10. The efficient nuclide identification method based on multi-label hybrid neural network according to claim 9, characterized in that: In step 4, precision, recall, F1 value and receiver operating characteristic curve are used to evaluate the performance of the neural network nuclide recognition; if the data instance is a positive class and is predicted to be a positive class, it is a true positive class; if the data instance is a negative class and is predicted to be a negative class, it is a true negative class; if the data instance is a negative class and is predicted to be a positive class, it is a false positive class; if the data instance is a positive class and is predicted to be a negative class, it is a false negative class. The precision, recall, F1 value and receiver operating characteristic curve are defined, where precision refers to the correct proportion of samples predicted to be positive in the sample, that is, Recall rate refers to the proportion of positive samples in the sample that are predicted correctly, that is, The F1 value represents the weighted harmonic mean of recall and precision, that is, F1Score reaches its best value at 1 and its worst value at 0; ROC curve, with FPR as the horizontal axis With TPR as the vertical axis