Gamma spectrum nuclide identification method based on convolutional recurrent neural network

Abstract

The invention relates to the field of gamma spectrum analysis, and discloses a gamma spectrum nuclide identification method based on a convolutional recurrent neural network. According to the method, hidden features of low-count and low-resolution complex gamma spectrum data can be efficiently extracted, and meanwhile, low-count peaks and overlapping peaks can be accurately analyzed. The gamma spectrum nuclide identification method based on the convolutional recurrent neural network comprises the steps of data acquisition, construction of the convolutional recurrent neural network (2DCNN-BiLSTM), training of the convolutional recurrent neural network (2DCNN-BiLSTM), and experimental verification. By adopting the gamma spectrum nuclide identification method based on the convolutional recurrent neural network, gamma spectrum nuclide identification can be accurately and effectively realized; not only are gray brightness features of the gamma spectrum gray-scale map extracted, but also time sequence correlation between the gray features is extracted, and nuclide identification is guided by the extracted space-time dependent features; moreover, the method also has good generalization and versatility, and is of great significance to the recognition of radionuclides.

Description

Technical Field

[0001] The invention relates to the field of gamma spectrum analysis, and in particular to a gamma spectrum nuclide identification method based on convolutional recurrent neural network. Background Art

[0002] Nuclide identification is an important technical means for radiation safety supervision, so portable radiation detection equipment is widely used in non-destructive radioactive detection and identification. However, due to the low resolution of portable radiation detection equipment and the complex use environment, the collected gamma spectrum will have low peak counts, spectral peak overlap, spectral peak shift and other phenomena, making the gamma spectrum very complicated. Traditional nuclide identification requires data smoothing, filtering transformation and key peak finding operations on the gamma spectrum measured by the detector to obtain full-energy peak information including peak position, peak area, half-width, and then reverse search of the standard nuclide library to determine the type of target nuclide. However, when faced with the processing of complex γ spectra, traditional methods are not only time-consuming, but also have a high rate of misjudgment.

[0003] In recent years, many neural network methods have been used for nuclide identification to improve the discrimination ability of gamma spectra. For example, some scholars use fully connected neural networks to map gamma spectral data to nuclide categories. However, the feature extraction ability of this method is still limited to the surface of the gamma spectrum, and there is a large room for improvement in nuclide identification performance. With the deepening of research, some scholars began to analyze gamma spectra from the perspective of time series, converting gamma spectral data into time series through sliding windows, and using recurrent neural networks (RNN, such as LSTM: long short-term memory neural network) to analyze the correlation between data to determine the nuclide composition. However, this method ignores many potential spatial features in the gamma spectrum, resulting in low accuracy and stability. In addition, scholars have also proposed a nuclide identification method based on convolutional neural networks. Among them, the one-dimensional convolutional neural network (1DCNN) uses the full spectrum as input, which is simple but computationally intensive. The two-dimensional convolutional neural network (2DCNN) accepts gamma spectral data in the form of images as input to extract high-dimensional spatial features, which improves the recognition performance compared with the one-dimensional convolutional neural network. However, gamma spectra contain both spatial information, such as the shape, size and position of characteristic peaks, and temporal information before and after these peaks. Existing studies, when faced with complex gamma spectra, only analyze gamma spectral data from a single perspective and fail to fully utilize its hidden features, resulting in uncertainty problems in gamma spectrum identification. In addition, most existing methods focus on processing morphologically stable gamma spectral data and are not suitable for the analysis of low-count, low-quality gamma spectra. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a gamma spectral nuclide identification method based on a convolutional recurrent neural network 2DCNN-BiLSTM for complex gamma spectral data with low counts and low resolution, which can efficiently extract its hidden features and accurately analyze low count peaks and overlapping peaks, thereby accurately and effectively realizing gamma spectral nuclide identification.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a gamma spectrum nuclide identification method based on convolutional recurrent neural network, comprising the following steps:

[0006] S1. Acquisition of datasets;

[0007] Modeling and simulating radiation detection equipment using the GEANT4 toolkit 40 K. 57 Co. 60 Co. 137 Cs, 133 8. 67 Ga, 131 I. 75 The energy deposition data of the eight radionuclides Se were randomly extracted and mixed, and the maximum number of mixed data was 4;

[0008] The energy deposition data generated is given by the formula Sampling is performed so that the nuclide count is within 10 3 Up to 10 4 The low count interval of represents the nuclide count, It means that it obeys the uniform distribution function of 0-1. Represents a random floating-point number sampled from a uniform distribution function.

[0009] Then the energy deposition data is analyzed using and Gaussian broadening is performed, where a, b, and c represent the half-width fitting coefficients of the real detector gamma spectrum; represents the half-width of the characteristic peak, E represents the original energy deposition value, a, b, c represent the half-width fitting coefficients of the real detector gamma spectrum, G represents a Gaussian function random variable that obeys a 0-1 distribution, Represents the energy deposition value with resolution;

[0010] S2, performing three-point smoothing and piecewise normalized grayscale mapping processing on the gamma spectrum data obtained in step 1;

[0011] set up Indicates the point to be smoothed. Indicates the dimension number, 2 points are taken on each side. express The previous point, express The next point is to use calculate The arithmetic mean As a correction value for this point;

[0012] The gamma spectrum data is segmented and matrix mapped according to a sliding window of size 64, and is divided into a total of 32×64 data blocks. For each dimension, the 64 eigenvalues ​​are independently mapped to grayscale data between [0,1] using the maximum and minimum normalization method. The normalization formula is: ,in represents the normalized replacement value, Represents the original eigenvalue, max is the maximum value of the data segment, and min is the minimum value of the data segment;

[0013] The processed gamma spectrum grayscale data is divided into training set, test set and validation set in a ratio of 8:1:1;

[0014] S3, construction of convolutional recurrent neural network;

[0015] The convolutional recurrent neural network is used to fuse 2DCNN (two-dimensional convolutional neural network) and BiLSTM (bidirectional long short-term memory neural network) to extract features from gamma spectrum data; the convolutional recurrent neural network consists of an input layer, a 2DCNN layer, a serialization layer, a BiLSTM layer, a fully connected layer, and an output layer;

[0016] The input layer is used to receive the converted 1×32×64 size gamma spectrum grayscale data for processing;

[0017] The 2DCNN layer includes 5 convolution modules, each of which includes feature extraction and feature enhancement processing; the feature extraction refers to performing convolution calculation, normalization, and activation processing on the input gamma spectrum grayscale data to extract the spatial features of the grayscale data and improve the number of feature channel expressions; the feature enhancement refers to performing a maximum pooling and 40% random neuron discarding operation after the feature extraction is performed on the convolution calculation, normalization, and activation, and the feature enhancement does not improve the number of feature channel expressions;

[0018] The results of the first four convolution modules are to increase the feature channel of the data by 1 times and reduce the dimension of the first dimension by 1 times. The fifth convolution module does not change the feature channel. After repeated feature extraction and enhancement of the five convolution modules, a 256×1×64 gamma spectrum feature sequence is finally obtained.

[0019] The serialization layer extracts time series features and serializes the feature sequence obtained by the convolution layer as the input of the BiLSTM layer;

[0020] The BiLSTM layer combines the forward and reverse order of LSTM into BiLSTM, and uses BiLSTM to extract forward and backward features from the convolution feature sequence. The number of hidden layer neurons of BiLSTM is 256. After calculation by the LSTM unit, a 64×256 feature matrix is ​​output;

[0021] The fully connected layer is used to take the last 1×256 time step of the feature matrix for full connection operation, mapping the 256-dimensional features mixed with spatial dependency and temporal dependency into a 1×8-dimensional output;

[0022] The output layer will output 1×8 dimensions using Normalized to classification probabilities, represents the 8-bit output of the fully connected layer, Represents a probability value between 0 and 1;

[0023] S4, training the convolutional recurrent neural network;

[0024] The training platform is PyCharm2021.3.1, the training tool is PyTorch, the number of training times is set to 100, the amount of data processed is 100, and the learning rate is 0.000008;

[0025] S5. Testing and evaluation of convolutional recurrent neural networks;

[0026] After steps S1 to S4 are completed, the convolutional recurrent neural network is evaluated using five indicators: Accuracy, BCE Loss, Precision, Recall, and F1-Score;

[0027] The accuracy is defined as the ratio of the number of correctly identified gamma spectra to the total number, and an accuracy of 95% is considered qualified;

[0028] The BCE Loss is a binary cross entropy loss function that represents the degree of closeness between the predicted value and the true value; if the fluctuation value is less than 0.5 for 10 consecutive trainings, it is qualified;

[0029] The Precision is the accuracy rate, which indicates the frequency of instance 1 in the sample predicted as 1, and is used to describe the false alarm rate of the network for nuclide identification. The larger the value, the lower the false alarm rate. A precision of 95% is considered qualified.

[0030] The Recall is a recall rate used to measure the network's missed recognition rate of nuclides. The larger the value, the lower the missed recognition rate. A Recall of 95% is considered qualified.

[0031] The F1-Score takes into account both precision and recall, and can more comprehensively evaluate the performance of the model. A score of 95% is considered qualified. F1-Score is the harmonic mean of precision and recall, abbreviated as F1. The calculation formula of F1-Score is as follows:

[0032] ;

[0033] In the formula, Precision is the precision and Recall is the recall.

[0034] Furthermore, in step S1, the gamma spectrum data is energy-calibrated using a real radiation source and a radiation detection device, and during simulation, the energy interval value after calibration is used for energy counting statistics;

[0035] The energy scale formula is ,in The address of the characteristic peak, represents the characteristic peak energy, A, B, and C represent the fitting coefficients;

[0036] The gamma spectral data is superimposed with the actual environmental background and the temperature drift of the peak position. The background superposition function used is: ;

[0037] Among them, represents the signal-to-noise ratio, represents the nuclide count, Represents background counts, It means that it obeys the uniform distribution function of 0.5-0.9;

[0038] Furthermore, in step S3, the serialization layer divides the feature sequence using the feature channel as a sliding window and the feature sequence length as a step length to obtain 64×256 data blocks.

[0039] Furthermore, step S5 also includes steps of constructing three common nuclide identification methods, namely 2D-VGG16 (16-layer visual geometry group neural network), 1D-CNN (one-dimensional convolutional neural network), and PCA-BPNN (principal component analysis & back propagation neural network), and comparing them with the convolutional recurrent neural network in terms of total recognition accuracy, mixed source recognition accuracy, and overlapping peak resolution accuracy.

[0040] Furthermore, in step S5, the generalization performance of the convolutional recurrent neural network is experimented and evaluated; when the nuclide counts are 2048, 4096, 6144, 8192, and 10240, 0, 5, and 10 channels of spectral drift and superimposed noise with a signal-to-noise ratio of 0.5 to 0.9 are respectively constructed to test the generalization of the model, so as to test the influence of the three factors of nuclide count, spectral peak drift, and noise on the nuclide recognition performance of the model, and evaluate the model's resistance to interference.

[0041] The beneficial effects of the present invention are as follows: the gamma spectrum nuclide identification method based on convolutional recurrent neural network described in the present invention has the following advantages:

[0042] 1. A nuclide identification model that can more comprehensively utilize the characteristics of gamma spectral data is constructed, which makes up for the defects of emerging methods in feature extraction difficulties and low recognition performance in low-count and low-resolution gamma spectral nuclide identification, provides new ideas for the analysis of gamma spectral data, and improves the nuclide identification accuracy of gamma spectra of radiation detection equipment.

[0043] 2. The convolutional recurrent neural network proposed in the present invention not only has better resolution accuracy for low-count peaks and overlapping peaks, but also has good generalization performance for different nuclide counts, spectral peak drift, and environmental noise. In addition, the model has shown good practicality in the identification of nuclides in short-term measurements of low-count gamma spectra in actual radiation detection equipment experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Flow chart of a gamma spectroscopy nuclide identification method based on a convolutional recurrent neural network in an embodiment of the present invention;

[0045] Figure 2 This is a flow chart of segmented normalized grayscale mapping in an embodiment of the present invention;

[0046] Figure 3 A model structure diagram based on a convolutional recurrent neural network in an embodiment of the present invention;

[0047] Figure 4 It is a schematic diagram showing the comparison of the nuclide recognition of overlapping peaks in the low energy region of the radioactive source by three models including 2D-VGG16, 1D-CNN and PCA-BPNN in the embodiment of the present invention;

[0048] Figure 5 Schematic diagram of the change of drift-free nuclide identification accuracy rate of the gamma spectrum nuclide identification method based on convolutional recurrent neural network in an embodiment of the present invention;

[0049] Figure 6This is a schematic diagram of the change of the nuclide identification accuracy of the gamma spectrum nuclide identification method based on convolutional recurrent neural network in drifting 5 channels in an embodiment of the present invention;

[0050] Figure 7 This is a schematic diagram of the change in the nuclide identification accuracy of the gamma spectrum nuclide identification method based on convolutional recurrent neural network in the embodiment of the present invention when drifting 10 channels;

[0051] Figure 8 Schematic diagram of the change of drift-free nuclide identification accuracy of the 2D-VGG16 model in an embodiment of the present invention;

[0052] Fig. 9 Schematic diagram of the change of the nuclide recognition accuracy of the 2D-VGG16 model in drift 5 channels in an embodiment of the present invention;

[0053] Fig.10 Schematic diagram of the change of the nuclide identification accuracy of the 2D-VGG16 model in the embodiment of the present invention when drifting 10 channels;

[0054] Fig.11 Schematic diagram of the change of the drift-free nuclide identification accuracy of the 1D-CNN model in an embodiment of the present invention;

[0055] Fig.12 Schematic diagram of the change of the nuclide recognition accuracy of the 1D-CNN model in drift 5 channels in an embodiment of the present invention;

[0056] Fig.13 Schematic diagram of the change of the nuclide recognition accuracy of the 1D-CNN model in the embodiment of the present invention when drifting 10 channels;

[0057] Fig.14 Schematic diagram of the change of drift-free nuclide identification accuracy of the PCA-BPNN model in an embodiment of the present invention;

[0058] Fig.15 Schematic diagram of the change of the nuclide identification accuracy of the PCA-BPNN model in drift 5 channels in an embodiment of the present invention;

[0059] Fig.16 Schematic diagram of the change of the nuclide identification accuracy of the PCA-BPNN model in the embodiment of the present invention when drifting 10 channels;

[0060] Fig.17 for Figure 2 The specific schematic diagram of A in the figure represents the schematic diagram of input data;

[0061] Fig.18 for Figure 2 The specific schematic diagram of B in the figure shows the schematic diagram of the serialization segmentation process;

[0062] Fig.19 for Figure 2 The specific schematic diagram of C in the figure represents the maximum and minimum normalization process diagram;

[0063] Fig. 20 for Figure 2 The specific schematic diagram of D in the figure represents the mapped grayscale image. DETAILED DESCRIPTION

[0064] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0065] like Figure 1 The gamma spectrum nuclide identification method based on convolutional recurrent neural network described in the present invention comprises the following steps:

[0066] 1. Data acquisition;

[0067] Specifically, data acquisition includes the following steps:

[0068] 1.1、Acquisition of data set;

[0069] Modeling and simulating radiation detection equipment using the GEANT4 toolkit 40 K. 57 Co. 60 Co. 137 Cs, 133 8. 67 Ga, 131 I. 75 The energy deposition data of the eight radionuclides Se were randomly extracted and mixed, and the maximum number of mixed data was 4;

[0070] In order to adhere to the principle of feature diversity, the generated energy deposition data is applied using the formula Sampling is performed so that the nuclide count is within 10 3 Up to 10 4 The low count interval of represents the nuclide count, It means that it obeys the uniform distribution function of 0-1. Represents a random floating point number sampled from a uniform distribution function. This means that the nuclide characteristic interval counts in the data set range from tens to thousands, which is a low count category;

[0071] Then the energy deposition data is analyzed using and Gaussian broadening is performed, where represents the half-width of the characteristic peak, E represents the original energy deposition value, a, b, c represent the half-width fitting coefficients of the real detector gamma spectrum, and The fitting calculation of CH158-06 NaI(Tl) detector is 137 Cs, 137 Cs+ 60 The fitting coefficients of Co are a=0.002, b=0.01878, c=0.22525, G represents a Gaussian function random variable that obeys a 0-1 distribution. Represents the energy deposition value with resolution.

[0072] In order to enhance the authenticity and credibility of the simulated gamma spectrum, the gamma spectrum data is energy scaled using real radiation sources and radiation detection equipment. During the simulation, the scaled energy interval values ​​are used for energy counting statistics. The energy scale formula is: ,in The address of the characteristic peak, represents the characteristic peak energy, A, B, and C represent the fitting coefficients, and the CH158-06 NaI (Tl) detector is used to detect 137 Cs, 137 Cs+ 60 The fitting of Co yields A = 2.63285e-5, B = b = 0.90461, C = -0.19832;

[0073] The gamma spectral data is superimposed with the actual environmental background and the temperature drift of the peak position. The background superposition function used is: ;

[0074] Among them, represents the signal-to-noise ratio, represents the nuclide count, Represents background counts, It means that it obeys the uniform distribution function of 0.5-0.9;

[0075] After the above work is completed, a total of 68,000 2048-dimensional gamma spectrum data are generated.

[0076] 1.2. Perform three-point smoothing and segmented normalized grayscale mapping on the gamma spectrum data obtained in step 1;

[0077] set up Indicates the point to be smoothed. Indicates the dimension number, 2 points are taken on each side. express The previous point, express The next point is to use calculate The arithmetic mean As a correction value for this point;

[0078] The gamma spectrum data is segmented and matrix mapped according to a sliding window of size 64, and is divided into a total of 32×64 data blocks. For each dimension, the 64 eigenvalues ​​are independently mapped to grayscale data between [0,1] using the maximum and minimum normalization method. The normalization formula is: ,in represents the normalized replacement value, Represents the original eigenvalue, max is the maximum value of the data segment, and min is the minimum value of the data segment; Figure 2 As shown, it is the flow chart of segmented normalized grayscale mapping. Fig.17 for Figure 2 The specific schematic diagram of A in the figure represents the schematic diagram of input data; Fig.18 for Figure 2 The specific schematic diagram of B in the figure shows the schematic diagram of the serialization segmentation process; Fig.19 for Figure 2 The specific schematic diagram of C in the figure represents the maximum and minimum normalization process diagram; Fig. 20 for Figure 2 The specific schematic diagram of D in the figure represents the mapped grayscale image.

[0079] The processed gamma spectrum grayscale data is divided into training set, test set, and validation set in a ratio of 8:1:1.

[0080] 2. Model construction;

[0081] Model building includes the following steps:

[0082] 2.1、Construction of convolutional recurrent neural network;

[0083] The convolutional recurrent neural network is used to fuse 2DCNN and BiLSTM to extract features from gamma spectrum data; Figure 3 As shown in the figure, the convolutional recurrent neural network consists of an input layer, a 2DCNN layer, a serialization layer, a BiLSTM layer, a fully connected layer, and an output layer.

[0084] The input layer is used to receive the converted 1×32×64 size gamma spectrum grayscale data for processing;

[0085] The 2DCNN layer includes 5 convolution modules, each of which includes two sequential tasks of feature extraction and feature enhancement. The feature extraction task refers to performing convolution calculation, normalization, and activation on the input gamma spectrum grayscale data to extract the spatial features of the grayscale data and improve the number of feature channel expressions. Unlike feature extraction, feature enhancement refers to performing a maximum pooling and 40% of neurons randomly discarded after convolution calculation, normalization, and activation in feature extraction, and feature enhancement does not improve the number of feature channel expressions; the results of the first four convolution modules are to increase the feature channels of the data by 1 times and reduce the dimension of the first dimension by 1 times, and the fifth convolution module does not change the feature channels; after repeated feature extraction and enhancement of five convolution modules, a 256×1×64 gamma spectrum feature sequence is finally obtained;

[0086] The serialization layer extracts temporal features and serializes the feature sequence obtained by the convolution layer as the input of the BiLSTM layer;

[0087] The serialization layer uses the feature channel as a sliding window and the feature sequence length as a step size to divide the feature sequence into 64×256 data blocks.

[0088] The BiLSTM layer combines the forward and reverse order of LSTM into BiLSTM, and uses BiLSTM to extract forward and backward features from the convolution feature sequence. The number of hidden layer neurons of BiLSTM is 256. After calculation by the LSTM unit, a 64×256 feature matrix is ​​output;

[0089] The fully connected layer is used to take the last 1×256 time step of the feature matrix for full connection operation, mapping the 256-dimensional features mixed with spatial dependency and temporal dependency into a 1×8-dimensional output;

[0090] The output layer will output 1×8 dimensions using Normalized to classification probabilities, represents the 8-bit output of the fully connected layer, Represents a probability value between 0 and 1;

[0091] 2.2. Train the convolutional recurrent neural network;

[0092] The training platform is PyCharm2021.3.1, the programming language and version are Python3.10, the training tool is PyTorch, the number of model training times is set to 100, the batch data size is set to 100; the learning rate is 0.000008.

[0093] For the network model, the values ​​of weight coefficients and bias items are key parameters for structural optimization. During the training process, the Adam optimizer is used to dynamically adjust the learning rate of each parameter, and adaptively optimize the update of weight coefficients and bias items during the back propagation process, thereby accelerating the convergence speed of model training. The adjustment rules of the algorithm are shown in Table 1:

[0094] Table 1 Adam parameter adjustment rules

[0095]

[0096] in, and denote the first and second order moments of the gradient, respectively. represents the loss function, Indicates the update parameters, represents the random target parameter gradient obtained by back-propagation optimization during this round of training, and are estimates of the first and second moments of the gradient, respectively. and It is the exponential decay rate that controls the two moment estimates, which is generally set to 0.9 and 0.999. is the initial learning rate, set to 0.000008, It is a very small number to prevent division by zero anomalies and is set to 1e-8. k is the current iteration number and is used for deviation correction.

[0097] 3. Generalization analysis,Experimental verification includes the following steps:

[0098] 3.1. Testing and evaluation of convolutional recurrent neural networks;

[0099] After steps S1 to S4 are completed, the convolutional recurrent neural network is evaluated using five indicators: Accuracy, BCE Loss, Precision, Recall, and F1-Score;

[0100] Accuracy is defined as the ratio of the number of correctly identified gamma spectra to the total number, and it is considered that 95% is qualified; the binary cross entropy loss function (BCE Loss) represents the closeness between the predicted value and the true value, and it is considered that the fluctuation value of 10 consecutive training times is less than 0.5, which is qualified; precision represents the frequency of instance 1 in the sample predicted to be 1, which can be used to describe the false alarm rate of the network for nuclide identification. The larger the value, the lower the false alarm rate, and it is considered that 95% is qualified; recall is used to measure the missed recognition rate of the network for nuclides. The larger the value, the lower the missed recognition rate, and it is considered that 95% is qualified; F1-Score comprehensively considers precision and recall, and can more comprehensively evaluate the performance of the model, and it is considered that 95% is qualified.

[0101] The calculation formula of F1-Score is as follows:

[0102] ;

[0103] In the formula, Precision is the precision and Recall is the recall.

[0104] After the model training is completed, the Accuracy is 97.71%, the fluctuation value of BCE Loss is less than 0.4, the Precision is 98.76%, the Recall is 99.43%, and the F1-Score is 99.57%, which means that the model training is completed and meets the qualification standard.

[0105] Three common nuclide identification methods, 2D-VGG16, 1D-CNN, and PCA-BPNN, were constructed and compared with the convolutional recurrent neural network in terms of total recognition accuracy, mixed source recognition accuracy, and overlapping peak resolution accuracy.

[0106] Experiments and evaluations were conducted on the generalization performance of the convolutional recurrent neural network. When the nuclide counts were 2048, 4096, 6144, 8192, and 10240, 0, 5, and 10 channels of spectral drift and noise with a superposition signal-to-noise ratio of 0.5-0.9 were performed to construct new data for the generalization test of the model. The effects of nuclide counts, spectral peak drift, and noise on the nuclide recognition performance of the model were tested, and the model's resistance to interference was evaluated. The test set contained 68 categories of gamma spectral data, with 20 data items in each category, totaling 102,000 gamma spectral test data items.

[0107] Finally, the model is applied to the radiation detection equipment to detect the 137 Cs, 137 Cs+ 60 Gamma spectroscopy nuclide identification task of Co radioactive source.

[0108] Test experiment

[0109] The gamma spectrum recognition accuracy of the gamma spectrum nuclide recognition method based on convolutional recurrent neural network described in the present invention was tested.

[0110] This experiment uses the divided test set gamma spectrum data (the resolution of gamma rays at 662KeV energy is 7.9%) as input for the experiment. The test set includes 8 single radiation sources, 20 random mixed dual radiation sources, 20 triple radiation sources and 20 quadruple radiation sources. The recognition accuracy of the model of the present invention for single source, dual source, triple source and quadruple source gamma spectra is 98.86%, 97.78%, 96.80% and 95.28% respectively.

[0111] In order to verify the recognition performance of the model, this experiment constructed three network models, 2D-VGG16, 1D-CNN, and PCA-BPNN, and compared them with 2DCNN-BiLSTM. Among them, 2D-VGG16 contains 13 convolutional layers, 3 fully connected layers, and 5 pooling layers. It uses the same gamma spectrum grayscale image as the model in this article as the network input, sets the initial input feature channel to 1, the output feature channel to 256, the number of data batches to 80, the learning rate to 0.00001, and trains 100 times in total; 1D-CNN is different from the two-dimensional convolutional neural network in that it uses a one-dimensional convolution operation, which consists of 5 one-dimensional convolution and pooling modules, 1 normalization layer, and 1 fully connected layer, so the input The input end uses the full spectrum for input, the input feature channel is 1, the output feature channel is 128, the data batch number is 480, the learning rate is set to 0.000004, and a total of 80 training times; PCA-BPNN consists of two parts, PCA stands for principal component analysis, which is used to reduce the dimension of the full spectrum data, and BPNN stands for back propagation neural network, which is used for feature extraction and result mapping. This part contains 3 hidden layers, with the number of neurons being 1024, 550, and 350, respectively, and finally connected to an output layer. After PCA dimensionality reduction, the gamma spectrum data of size 1×2048 becomes 1×1024, and then is input into the network for training, the data batch number is 480, the learning rate is set to 0.00008, and a total of 100 training times; the performance of the final four network models on the test set is shown in Table 2.

[0112] Table 2 Comparison of recognition accuracy of four models on the test set (%)

[0113] From the above table, it can be found that 2DCNN-BiLSTM has the best average recognition accuracy of nuclide identification compared with the other three methods, which is 4.74% higher than the PCA-BPNN method, 4.55% higher than the 1D-CNN method, and 1.23% higher than the 2D-VGG16 method. For the recognition effect of single-source nuclide gamma spectra, the gap between the four methods is not large. With the increase in the number of mixed sources, the change in the accuracy of 2DCNN-BiLSTM is about 1%, while the accuracy of other methods is greatly affected. The convolutional recurrent neural network proposed in the present invention has the best recognition performance, and the model has a stronger ability to identify nuclides in mixed source gamma spectra.

[0114] For the convolutional recurrent neural network proposed in the present invention, the recognition accuracy of the gamma spectra with overlapping peaks in the low energy region of the test set is shown in Table 3. Table 3 shows that the average recognition accuracy of the two-source, three-source and four-source mixed gamma spectra is 97.95%, 96.38% and 93.17%, respectively. Figure 4 The comparison curves of the recognition accuracy of different models for overlapping peaks of gamma spectra of different mixed sources are shown. In the comparative experiment, the 2DCNN-BiLSTM model has better nuclide recognition accuracy for different mixed sources than 2D-VGG16 (97.5%, 95.13%, 92%), 1D-CNN (94.75%, 89.5%, 85.5%), and PCA-BPNN (97.5%, 90.13%, 84.92%). Figure 4 As for the curve change trend, the changes in accuracy of 2D-VGG16 and 2DCNN-BiLSTM caused by the increase of mixed sources are basically the same, and the decline is relatively gentle, but the analysis performance of 2DCNN-BiLSTM for overlapping peaks is better than that of 2D-VGG16. The other two methods have poor analysis of overlapping peaks, especially when there are many mixed sources, and the recognition performance is greatly suppressed.

[0115] The above studies show that the 2DCNN-BiLSTM neural network extracts more overlapping peak features and uses these features for predictive classification, so the overlapping peak resolution and nuclide identification performance are stronger than the comparative model.

[0116] Table 3 Gamma spectrum overlapping peak recognition accuracy of the 2DCNN-BiLSTM model of the present invention (%)

[0117]

[0118] In order to evaluate the generalization performance of 2DCNN-BiLSTM, the spectrum drift of 0, 5, and 10 channels was performed respectively when the nuclide counts were 2048, 4096, 6144, 8192, and 10240, and the noise of 0.5, 0.6, 0.7, 0.8, and 0.9 was superimposed respectively to construct a new generalization dataset for comparative testing of 2DCNN-BiLSTM, 2D-VGG16, 1D-CNN, and PCA-BPNN. The generalization test set data contains a total of 102,000 gamma spectral data. Figures 5 to 16 The figure shows the changes in nuclide identification under the influence of signal-to-noise ratio and nuclide count under different models and different drift degrees. Figure 5 Schematic diagram of the change of drift-free nuclide identification accuracy rate of the gamma spectrum nuclide identification method based on convolutional recurrent neural network in an embodiment of the present invention; Figure 6 This is a schematic diagram of the change of the nuclide identification accuracy of the gamma spectrum nuclide identification method based on convolutional recurrent neural network in drifting 5 channels in an embodiment of the present invention; Figure 7 This is a schematic diagram of the change in the nuclide identification accuracy of the gamma spectrum nuclide identification method based on convolutional recurrent neural network in the embodiment of the present invention when drifting 10 channels; Figure 8 Schematic diagram of the change of drift-free nuclide identification accuracy of the 2D-VGG16 model in an embodiment of the present invention; Fig. 9 Schematic diagram of the change of the nuclide recognition accuracy of the 2D-VGG16 model in drift 5 channels in an embodiment of the present invention; Fig.10 Schematic diagram of the change of the nuclide identification accuracy of the 2D-VGG16 model in the embodiment of the present invention when drifting 10 channels; Fig.11 Schematic diagram of the change of the drift-free nuclide identification accuracy of the 1D-CNN model in an embodiment of the present invention; Fig.12 Schematic diagram of the change of the nuclide recognition accuracy of the 1D-CNN model in drift 5 channels in an embodiment of the present invention; Fig.13 Schematic diagram of the change of the nuclide recognition accuracy of the 1D-CNN model in the embodiment of the present invention when drifting 10 channels; Fig.14 Schematic diagram of the change of drift-free nuclide identification accuracy of the PCA-BPNN model in an embodiment of the present invention; Fig.15 Schematic diagram of the change of the nuclide identification accuracy of the PCA-BPNN model in drift 5 channels in an embodiment of the present invention; Fig.16 Schematic diagram of the change of the nuclide identification accuracy of the PCA-BPNN model in the embodiment of the present invention after drifting 10 channels. Figures 5 to 16It can be seen that no matter what kind of drift situation, the accuracy of each model increases with the increase of counts and the increase of signal-to-noise ratio. When the nuclide count is less than 6144, the average count per channel is less than 3, and the nuclide recognition accuracy of each model is greatly affected by the nuclide count, because the gamma spectrum loses more information at this time, and it is difficult for the model to capture enough features for nuclide identification. However, even so, 2DCNN-BiLSTM still maintains an identification accuracy of more than 80% in extreme cases (shift=10, counts=2048, SNR=0.5), which is better than the performance of other models. When the nuclide count is greater than 6144, the recognition accuracy of each model is greatly affected by noise, but for 2DCNN-BiLSTM, the impact of noise is limited, and in most noise conditions, the recognition accuracy of gamma spectra can be maintained above 94%, which is also better than other models. This shows that 2DCNN-BiLSTM has good generalization and anti-interference for low-count, noisy gamma spectra.

[0119] Table 4 shows the gamma spectrum recognition accuracy of different models under the influence of spectrum peak drift and full spectrum count. According to the table information, for the four models of 2DCNN-BiLSTM, 2D-VGG16, 1D-CNN, and PCA-BPNN, the weaker the spectrum peak drift, the higher the model recognition accuracy. In the case of no drift, 2DCNN-BiLSTM has the highest recognition accuracy, with an average value of 95.63%, which is significantly better than other methods. When the drift is 5 channels, although the accuracy of 2DCNN-BiLSTM has decreased, it still maintains a good level, with an average value of 92.53%. In the case of drift of 10 channels, the average accuracy of 2DCNN-BiLSTM is still above 90%, while the other methods are all below 90%. Compared with 2DCNN-BiLSTM, the recognition accuracy of 2D-VGG16 and PCA-BPNN in each case is lower than that of 2DCNN-BiLSTM. Although 1D-CNN performs well in the case of no drift and drift of 5 channels, its accuracy drops significantly when drift is 10 channels, with an average value of only 78.63%. At this time, the generalization performance of 1D-CNN for drift has been greatly reduced. Overall, 2DCNN-BiLSTM performs most stably and superiorly under different drift conditions, and has certain generalization ability and robustness.

[0120] Table 4 Gamma spectrum recognition accuracy under the influence of different models on peak drift and nuclide counts (%)

[0121]

[0122] In order to verify the practical application capability of the model, the CH15806 NaI (Tl) detector was used in the laboratory to collect gamma spectral data of two types of radiation sources for practical application testing of the model: 137Cs (8×10 mm, 373 kBq) and 137 Cs+ 60 Co (8×10 mm, 394 kBq). In general, the central axis of the NaI(Tl) detector was used as the reference angle, and gamma spectral data were collected from the radioactive sources within the 90° symmetric range on the left. Specifically, the radioactive sources were placed between 5 cm and 50 cm along four angle paths of 0°, 30°, 60°, and 90°, with an interval of 5 cm each time. The acquisition time was 1 second, 2 seconds, 3 seconds, 4 seconds, 5 seconds, and 6 seconds, respectively, and a total of 480 data sets were collected for model testing.

[0123] The measured data was input into 2DCNN-BiLSTM for testing, and the recognition results are shown in Table 5. According to the recognition results, it can be found that as the measurement time increases, the more gamma spectrum counting information there is, the more accurate the nuclide identification is. Within the 50cm sector, under the measurement condition of as low as 2s, 137 Cs, 137 Cs+ 60 The Co recognition accuracy rate reached more than 90%, and the model could correctly identify the gamma spectrum when the measurement time was 4s. This shows that the gamma spectrum nuclide identification method based on convolutional recurrent neural network described in the present invention has correctly learned the spatiotemporal dependence characteristics of the characteristic peaks of different radioactive nuclides in the gamma spectrum and performed nuclide identification. The method has certain applicability and can be used for nuclide identification tasks in short-term measurements.

[0124] Table 5 Recognition accuracy of 2DCNN-BiLSTM for gamma spectra at different measurement times (%)

[0125]

Claims

1. A gamma spectroscopy nuclide identification method based on convolutional recurrent neural network, characterized in that: The following steps are involved: S1. Acquisition of datasets; Modeling and simulating radiation detection equipment using the GEANT4 toolkit 40 K. 57 Co. 60 Co. 137 Cs, 133 8. 67 Ga, 131 I. 75 The energy deposition data of the eight radionuclides Se were randomly extracted and mixed, and the maximum number of mixed data was 4; The energy deposition data generated is given by the formula Sampling is performed so that the nuclide count is within 10 3 Up to 10 4 The low count interval of represents the nuclide count, It means that it obeys the uniform distribution function of 0-1. Represents a random floating point number sampled from a uniform distribution function; Then the energy deposition data is analyzed using and Gaussian broadening is performed, where represents the half-width of the characteristic peak, E represents the original energy deposition value, a, b, c represent the half-width fitting coefficients of the real detector gamma spectrum, G represents a Gaussian function random variable that obeys a 0-1 distribution, Represents the energy deposition value with resolution; S2, performing three-point smoothing and piecewise normalized grayscale mapping processing on the gamma spectrum data obtained in step 1; set up Indicates the point to be smoothed. Indicates the dimension number, 2 points are taken on each side. express The previous point, express The next point is to use calculate The arithmetic mean As a correction value for this point; The gamma spectrum data is segmented and matrix mapped according to a sliding window of size 64, and is divided into a total of 32×64 data blocks. For each dimension, the 64 eigenvalues ​​are independently mapped to grayscale data between [0,1] using the maximum and minimum normalization method. The normalization formula is: ,in represents the normalized replacement value, Represents the original eigenvalue, max is the maximum value of the data segment, and min is the minimum value of the data segment; The processed gamma spectrum grayscale data is divided into training set, test set and validation set in a ratio of 8:1:1; S3, construction of convolutional recurrent neural network; The convolutional recurrent neural network is used to fuse 2DCNN and BiLSTM to extract features from gamma spectrum data; the convolutional recurrent neural network consists of an input layer, a 2DCNN layer, a serialization layer, a BiLSTM layer, a fully connected layer, and an output layer; The input layer is used to receive the converted 1×32×64 size gamma spectrum grayscale data for processing; The 2DCNN layer includes 5 convolution modules, each of which is used to realize feature extraction and feature enhancement; Feature extraction refers to convolution calculation, normalization, and activation of the input gamma spectrum grayscale data to extract the spatial features of the grayscale data and increase the number of data feature channel expressions; The feature enhancement refers to performing a maximum pooling and 40% random neuron drop operation after feature extraction, convolution calculation, normalization, and activation, and feature enhancement does not increase the number of channel expressions of the feature; The results of the first four convolution modules are to increase the feature channel of the data by 1 times and reduce the dimension of the first dimension by 1 times. The fifth convolution module does not change the feature channel. After repeated feature extraction and enhancement of the five convolution modules, a 256×1×64 gamma spectrum feature sequence is finally obtained. The serialization layer extracts time series features and serializes the feature sequence obtained by the convolution layer as the input of the BiLSTM layer; The BiLSTM layer combines the forward and reverse order of LSTM into BiLSTM, and uses BiLSTM to extract forward and backward features from the convolution feature sequence. The number of hidden layer neurons of BiLSTM is 256. After calculation by the LSTM unit, a 64×256 feature matrix is ​​output; The fully connected layer is used to take the last 1×256 time step of the feature matrix for full connection operation, mapping the 256-dimensional features mixed with spatial dependency and temporal dependency into a 1×8-dimensional output; The output layer will output 1×8 dimensions using Normalized to classification probabilities, represents the 8-bit output of the fully connected layer, Represents a probability value between 0 and 1; S4, training the convolutional recurrent neural network; S5. Testing and evaluation of convolutional recurrent neural networks; After steps S1 to S4 are completed, the convolutional recurrent neural network is evaluated using five indicators: Accuracy, BCE Loss, Precision, Recall, and F1-Score; The accuracy is defined as the ratio of the number of correctly identified gamma spectra to the total number, and an accuracy of 95% is considered qualified; The BCE Loss is a binary cross entropy loss function that represents the degree of closeness between the predicted value and the true value; if the fluctuation value is less than 0.5 for 10 consecutive trainings, it is qualified; The Precision is the accuracy rate, which indicates the frequency of instance 1 in the sample predicted as 1, and is used to describe the false alarm rate of the network for nuclide identification. The larger the value, the lower the false alarm rate. Precision reaches 95% to be qualified; The Recall is a recall rate used to measure the network's missed recognition rate of nuclides. The larger the value, the lower the missed recognition rate. A Recall of 95% is considered qualified. The F1-Score takes into account both precision and recall, and can more comprehensively evaluate the performance of the model. A score of 95% is considered qualified. The F1-Score is the harmonic mean of precision and recall.

2. The gamma spectrum nuclide identification method based on convolutional recurrent neural network as claimed in claim 1, characterized in that: In step S1, the gamma spectrum data is energy-calibrated using a real radiation source and a radiation detection device, and during simulation, the energy interval value after calibration is used for energy counting statistics; The energy scale formula is: ; in The address of the characteristic peak, represents the characteristic peak energy, A, B, and C represent the fitting coefficients; The gamma spectral data is superimposed with the actual environmental background and the temperature drift of the peak position. The background superposition function used is: ; in represents the signal-to-noise ratio, represents the gamma spectrum coarse counts, Represents background counts, It indicates that it obeys the uniform distribution function of 0.5-0.

9.

3. The gamma spectrum nuclide identification method based on convolutional recurrent neural network as claimed in claim 1, characterized in that: In step S3, the serialization layer uses the feature channel as a sliding window and the feature sequence length as a step length to divide the feature sequence into 64×256 data blocks.

4. The gamma spectrum nuclide identification method based on convolutional recurrent neural network as claimed in claim 3, characterized in that: Step S5 also includes the steps of constructing three common nuclide identification methods, namely, 16-layer visual geometry group neural network 2D-VGG16, one-dimensional convolutional neural network 1D-CNN, and principal component analysis & back propagation neural network PCA-BPNN, and comparing them with the convolutional recurrent neural network in terms of total recognition accuracy, mixed source recognition accuracy, and overlapping peak resolution accuracy.

5. The gamma spectrum nuclide identification method based on convolutional recurrent neural network as claimed in claim 4, characterized in that: In step S5, the generalization performance of the convolutional recurrent neural network is experimented and evaluated; when the nuclide counts are 2048, 4096, 6144, 8192, and 10240, 0, 5, and 10 channels of spectral drift and superimposed noise with a signal-to-noise ratio of 0.5 to 0.9 are respectively constructed to test the generalization of the model, in order to test the influence of the three factors of nuclide count, spectral peak drift, and noise on the nuclide recognition performance of the model, and to evaluate the model's resistance to interference.

Images