Neutron gamma pulse waveform stacking discrimination method under high counting rate condition

Abstract

The invention relates to a neutron gamma stack waveform discrimination method, in particular to a neutron gamma pulse waveform stack discrimination method under the condition of a high counting rate, and solves the technical problem of low accuracy caused by the fact that neutron gamma waveform trailing edge change is not considered in an existing neutron gamma waveform discrimination method. The neutron gamma pulse waveform stacking discrimination method under the high counting rate condition comprises the following steps: 1) collecting and storing neutron gamma single waveform mixed signals; 2) obtaining a neutron data set and a gamma data set S1; 3) constructing a test set P1 of neutron gamma stack waveforms; the method comprises the steps of (1) obtaining a test set P1, (2) obtaining a neutron data set and a gamma data set S2, (5) constructing a characteristic parameter set T, (6) constructing a training set sample library P2, (7) constructing a training data set P3, and (8) training a neural network, discriminating all waveforms in the test set P1 by using the trained neural network, and outputting prediction labels of the waveforms to complete neutron-gamma pulse waveform stacking discrimination under the condition of high counting rate.

Description

Technical Field

[0001] The invention relates to a neutron gamma stacking waveform identification method, in particular to a neutron gamma pulse waveform stacking identification method under high counting rate conditions. Background Art

[0002] Pulse shape discrimination (PSD) of high-energy emitted neutrons and gamma rays using organic scintillation detectors is a key requirement for applications in nuclear safety, fusion neutron diagnostics, and neutron-induced fission cross section measurement experiments. When the detector operates at a very high count rate, multiple radiation particles arrive at the detector within a pulse window period, resulting in two or more pulses partially or completely overlapping. This greatly changes the structure of the pulse, making traditional single-particle waveform discrimination methods ineffective. Although piled-up pulses can be identified and rejected, the resulting event loss limits the analysis of features against a high gamma background. In order to reduce the number of event losses, various methods have been proposed to extract PSD information from piled-up pulses.

[0003] F. Belli et al. reported a method for identifying the fitting of a double exponential decay function (Belli F, Esposito B, Marocco D, et al. A method for digital processing of pile-up events inorganic scintillators [J]. Nuclear Instruments and Methods in Physics Research, Section A. Accelerators, Spectrometers, Detectors and Associated Equipment, 2008 (2): 595.). This method uses an average decay constant and the original "good" data corresponding to the first pulse for pulse fitting to obtain sample data of the stacked portion of the trailing edge of the first pulse. The average decay constant used for fitting has a standard deviation of 25%, and the fitting process relies on multiple trials and errors, and it is sometimes difficult to converge to the correct solution. XLLuo reported a method for identification using a standard neutron gamma waveform (Pulse pile-up identification and reconstruction for liquid scintillator based neutron detectors [J]. Nuclear Instruments and Methods in Physics Research, Section A. Accelerators, Spectrometers, Detectors and Associated Equipment, 2018, 897: 59-65.), which is used for stacked waveform identification of BC-501A liquid scintillation detector. The standard waveform is obtained by averaging all known neutron gamma waveforms, and then the stacked pulse is compared with the stacked waveform constructed by the standard waveform to achieve identification and reconstruction of the stacked pulse. This method can achieve stacked waveform identification with an interval of 20ns between the two pulses.

[0004] However, the fluorescence generated by neutron gammas of different energies in the scintillator has different decay constants, and the trailing edge of the neutron gamma waveform has different fluctuations. For example, the normalized neutron waveforms will not completely overlap, and the rise time obtained by integrating the entire pulse is different. In the above stacked waveform identification method, the details of the trailing edge changes are not considered. Therefore, when identifying stacked waveforms, instead of simply using standard waveforms or average time constants, the changes in the trailing edge of the pulse waveform are considered, which is expected to further improve the identification accuracy. Summary of the invention

[0005] The purpose of the present invention is to solve the technical problem that the existing neutron gamma waveform identification methods do not consider the change of the neutron gamma waveform trailing edge, resulting in low accuracy, and provide a neutron gamma pulse waveform stacking identification method under high counting rate conditions.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A neutron gamma pulse waveform stacking identification method under high counting rate conditions is special in that it includes the following steps:

[0008] 1] Collect and store neutron gamma single waveform mixed signals;

[0009] 2] Identify the neutron gamma single waveform signal and obtain the neutron data set and gamma data set S1;

[0010] 3] Using the neutron data set and the gamma data set S1, construct the test set P1 of the neutron gamma stack waveform;

[0011] 4] Select the neutron waveforms and gamma waveforms of different amplitudes in the neutron data set and the gamma data set S1 according to the rise time and the count, and obtain the neutron data set and the gamma data set S2;

[0012] 5] Extract the characteristic parameters of all stacked waveforms in the test set P1 and construct a characteristic parameter set T corresponding to the test set P1;

[0013] 6] Based on the test set P1, the neutron data set and the gamma data set S2 and the characteristic parameter set T, the training set sample library P2 is constructed;

[0014] 7] Select the training samples corresponding to the test set P1 from the training set sample library P2 to construct the training data set P3;

[0015] 8] The training data set P3 is used as the training set of the neural network. The neural network is trained and used to identify all waveforms in the test set P1 and output their predicted labels to complete the neutron gamma pulse waveform stacking identification under high counting rate conditions.

[0016] Further, step 1] is specifically as follows:

[0017] A neutron tube is used as the neutron source, and a pulsed neutron detection system is used to collect and store more than or equal to 50,000 neutron gamma single waveform mixed signals.

[0018] Further, step 2] is specifically as follows:

[0019] The rise time method is used to identify the neutron gamma single waveform mixed signal to obtain the neutron waveform and gamma waveform. The peak positions of all neutron waveforms and gamma waveforms are aligned to form the neutron data set and gamma data set S1.

[0020] Further, step 3] is specifically as follows:

[0021] Neutron waveforms and gamma waveforms are randomly extracted from the neutron data set S1 and the gamma data set S1, and a test set P1 of neutron-gamma stacked waveforms is constructed according to four stacking methods: neutron + neutron, neutron + gamma, gamma + neutron, and gamma + gamma.

[0022] Further, step 5] is specifically as follows:

[0023] According to the peak value F1, peak value F2 and peak distance D of the stacked waveform, the characteristic parameters of each waveform in the test set P1 are extracted, and a characteristic parameter set T corresponding to the test set P1 is constructed.

[0024] Further, step 6] is specifically as follows:

[0025] 6.1. Construct a training sample library of neutron + neutron stacking mode for each test waveform in the test set P1;

[0026] 6.2. Extracting a first neutron waveform and a second neutron waveform from the neutron data set and the gamma data set S2;

[0027] 6.3. The peak value F of the i-th test waveform in the characteristic parameter set T i1 The ratio of the peak value f1 of the first neutron waveform is recorded as the peak factor r i1 ; The peak value F of the i-th test waveform in the characteristic parameter set T i2 The ratio of the peak value f2 of the second neutron waveform is recorded as the peak factor r i2 ; i=1,2,...,N, N is the number of test waveforms in the test set P1, N≥2;

[0028] 6.4. Multiply the first neutron waveform by the peak factor r i1 , the second neutron waveform multiplied by the peak factor r i2 , so that the two neutron waveforms are respectively corresponding to the peak value F i1 , peak value F i2 equal;

[0029] 6.5. Shift the second neutron waveform back by the peak spacing D of the i-th test waveform i Then, the first neutron waveform is added to the second neutron waveform, and the peak value f1 of the stacked waveform is extracted. * , peak value f2 * and peak spacing d;

[0030] 6.6. Calculate the new peak factor r1 according to the following formula: * 、r2 * and peak spacing d * :

[0031] r1 * =(F i1 +f1 * ) / 2f1

[0032] r2 * =(F i2 +f2 * ) / 2f2

[0033] d * =D i -(dD i );

[0034] 6.7. Return to step 6.2, and take out the second neutron waveform and the third neutron waveform in the neutron data set and the gamma data set S2 in order, and calculate the peak factor r according to the peak factor r. i1 =r1 * 、r i2 =r2 * and peak spacing D i =d * , reprocess the waveform until the peak value f1 * With peak F i1 and peak value f2 * With peak F i2 The peak error is less than 1%, the peak spacing d and the peak spacing D i The error is less than 2 sampling points, and all neutron waveforms in the neutron data set and the gamma data set S2 are traversed to complete the training set sample library of the neutron + neutron stacking method, which is recorded as the first training sample library;

[0035] 6.8. According to steps 6.1 to 6.7, the training set sample library constructed by the neutron + gamma stacking method is recorded as the second training sample library, the training set sample library constructed by the gamma + neutron stacking method is recorded as the third training sample library, and the training set sample library constructed by the gamma + gamma stacking method is recorded as the fourth training sample library; and the first training sample library, the second training sample library, the third training sample library and the fourth training sample library are taken as the training set sample library P2.

[0036] Further, step 7] is specifically as follows:

[0037] The K-nearest neighbor method is used to calculate the similarity. According to the similarity, the training samples corresponding to the test set P1 are selected from the training set sample library P2 to construct the training data set P3.

[0038] Further, step 8] is specifically as follows:

[0039] 8.1. Use the training data set P3 as the input of the neural network, train the neural network, and define each row as a training sample, and the number of rows is the number of training samples;

[0040] 8.2. Use the trained neural network to identify all test waveforms in the test set P1 and output their predicted labels to complete the stacking identification of neutron gamma pulse waveforms under high counting rate conditions.

[0041] Furthermore, in step 3], the stacked waveform is obtained by shifting any one of the two waveforms backward by 100 to 400 sampling points and then adding the two waveforms together to obtain the stacked waveform.

[0042] Furthermore, in step 8.1, the neural network is a CNN convolutional neural network.

[0043] Beneficial effects of the present invention:

[0044] 1. The present invention provides a neutron gamma pulse waveform stacking identification method under high counting rate conditions, which changes the previous method of randomly extracting samples as data sets. According to the different rise times of each single-particle waveform (i.e., neutron waveform or gamma waveform), the waveform characteristics are refined, and samples are extracted with the rise time as a parameter to make the samples more representative. When processing the training set (i.e., the training set sample library P2), the two peak values ​​and peak spacing of the processed stacked waveform are inconsistent with the waveform to be identified. An iterative method is used to make the stacked waveform continuously approach the characteristic parameters of the stacked waveform to be identified, and the training samples similar to the waveform to be identified are selected as the input of the neural network, thereby avoiding overfitting caused by the input of a large number of irrelevant training sets, and improving the calculation efficiency and accuracy.

[0045] 2. The present invention provides a method for identifying stacked neutron gamma pulse waveforms under high counting rate conditions, which combines rise time, K-nearest neighbor method and artificial neural network. Samples are extracted with rise time as a parameter to make the samples more representative. The K-nearest neighbor method calculates the similarity with the waveform to be identified, and selects training samples similar to the waveform to be identified as the input of the neural network, which provides a new way to identify stacked waveforms.

[0046] 3. The stacked waveform in the neutron gamma pulse waveform stacking identification method under high counting rate conditions of the present invention has three significant characteristic parameters, namely the first peak value, the second peak value and the peak spacing of the stacked waveform. The extracted characteristic parameters can effectively determine the main characteristics of the test waveform.

[0047] 4. The present invention provides a neutron gamma pulse waveform stacking identification method under high counting rate conditions. When the selected neutron waveform and gamma waveform are processed according to four stacking methods, the two peak values ​​and peak spacing of the processed waveform will change, which is different from the waveform characteristics of the test set. After modifying the characteristic parameters, multiple iterations are performed to narrow the gap between the characteristic parameters of the training set (i.e., the training set sample library P2) and the test waveform. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of an embodiment of a neutron gamma pulse waveform stacking discrimination method under high counting rate conditions of the present invention;

[0049] Figure 2 1 is a schematic diagram of using a rise time method to discriminate a neutron gamma single waveform mixed signal in an embodiment of a neutron gamma pulse waveform stacking discrimination method under high counting rate conditions of the present invention, Figure 2 The left side of the red line in the figure represents the gamma single wave signal, and the right side represents the neutron single wave signal;

[0050] Figure 3 Schematic diagram of two neutron waveforms in an embodiment of a neutron gamma pulse waveform stacking discrimination method under high counting rate conditions of the present invention, wherein the horizontal axis represents time and the vertical axis represents amplitude;

[0051] Figure 4 It is a schematic diagram of a neutron waveform being shifted back by 100 sampling points in an embodiment of a neutron gamma pulse waveform stacking discrimination method under high counting rate conditions of the present invention, wherein the horizontal axis represents time and the vertical axis represents amplitude;

[0052] Figure 5 1 is a schematic diagram of a stacked waveform obtained by adding two neutron waveforms in an embodiment of a neutron gamma pulse waveform stacking discrimination method under high counting rate conditions of the present invention, wherein the horizontal axis represents time and the vertical axis represents amplitude;

[0053] Figure 6 1 is a schematic diagram of the relative position comparison of unstacked neutron single waveforms (black and red) and stacked waveforms (green) in an embodiment of a neutron gamma pulse waveform stacking discrimination method under high counting rate conditions of the present invention, wherein the horizontal axis represents time and the vertical axis represents amplitude;

[0054] Figure 7 It is a schematic diagram of a CNN neural network in an embodiment of a neutron gamma pulse waveform stacking discrimination method under high counting rate conditions of the present invention. DETAILED DESCRIPTION

[0055] like Figure 1 As shown, a neutron gamma pulse waveform stacking identification method under high counting rate conditions includes the following steps:

[0056] Step 1: Collect and store neutron gamma single waveform mixed signal:

[0057] Using a neutron tube as the neutron source, a pulsed neutron detection system is used to collect and store more than or equal to 50,000 neutron gamma single waveform mixed signals. The pulsed neutron detection system consists of a stilbene crystal, an ETL-9815 photomultiplier tube, and an oscilloscope;

[0058] Step 2: Identify the neutron gamma single waveform signal:

[0059] like Figure 2 As shown, the rise time method is used to identify the neutron gamma single waveform mixed signal, and the neutron waveform and gamma waveform can be obtained. The peak positions of all neutron waveforms and gamma waveforms are aligned to form the neutron data set and gamma data set S1;

[0060] Step 3: Construct the test set P1 of neutron gamma stack waveform:

[0061] Neutron waveforms and gamma waveforms are randomly extracted from the neutron data set S1 and the gamma data set S1, and neutron-gamma stacked waveforms are constructed according to four stacking methods: neutron + neutron, neutron + gamma, gamma + neutron, and gamma + gamma. The four stacking methods are labeled; in this embodiment, the labels are recorded as 1, 2, 3, and 4, respectively, and 1000 stacked waveforms are processed for each stacking method, so a total of 4000 stacked waveforms are obtained to construct the test set P1.

[0062] When processing the stacked waveform of two neutron waveforms, one of the neutron waveforms is moved back by 100 sampling points (the sampling points vary randomly between 100 and 400, and the sampling points are the distance of the backward movement). Figure 3 As shown, these are the original waveforms of the two neutron waveforms; Figure 4 As shown in , the second neutron waveform is shifted back by 100 sampling points; Figure 5 As shown in , the neutron waveform after 100 sampling points is added to the first neutron waveform to obtain the stacked waveform of the neutron waveform; Figure 6 As shown, the relative position of the unstacked neutron single waveform (black and red) and the stacked waveform (green) is compared. After the waveforms are stacked, the peak value and peak position of the second waveform will change.

[0063] Step 4: Create neutron data set and gamma data set S2:

[0064] like Figure 2As shown in the figure, the rise time and count of the neutron waveform and the gamma waveform conform to the Gaussian distribution. According to the rise time and count, the neutron waveform and the gamma waveform of different amplitudes in the neutron data set S1 are selected, and the distribution of the neutron waveform and the gamma waveform is thinned. The selected neutron waveform and the gamma waveform are more representative. A total of 200 neutron waveforms and 200 gamma waveforms are selected to establish the neutron data set and the gamma data set S2 to prepare for the construction of the training set sample library.

[0065] Step 5: Extract the stacked waveform feature parameters in the test set P1:

[0066] According to the peak values ​​F1, F2 and the peak distance D of the stacked waveforms, the characteristic parameters of each waveform in the test set P1 are used to construct a characteristic parameter set T corresponding to the test set P1; the characteristic parameter set T can effectively determine the main characteristics of the test waveforms;

[0067] Step 6: Build the training set sample library P 2i :

[0068] 6.1. Construct a training sample library of neutron + neutron stacking for the i-th test waveform in the test set P1, i∈4000;

[0069] 6.2. Extracting a first neutron waveform and a second neutron waveform from the neutron data set and the gamma data set S2;

[0070] 6.3. The peak value F of the i-th test waveform in the characteristic parameter set T i1 The ratio of the peak value f1 of the first neutron waveform is recorded as the peak factor r i1 ; The peak value F of the i-th test waveform in the characteristic parameter set T i2 The ratio of the peak value f2 of the second neutron waveform is recorded as the peak factor r i2 ; i = 1, 2, ..., N, N is the number of test waveforms in the test set P1, N = 4000;

[0071] 6.4. Multiply the first neutron waveform by the peak factor r i1 , the second neutron waveform multiplied by the peak factor r i2 , so that the two neutron waveforms are respectively corresponding to the peak value F i1 , peak F i2 equal;

[0072] 6.5. Shift the second neutron waveform back by the peak spacing D of the i-th test waveform i Then, the first neutron waveform is added to the second neutron waveform, and the peak value f1 of the stacked waveform is extracted. * , peak value f2 * and peak spacing d;

[0073] 6.6. Calculate the new peak factor r1 according to the following formula: * 、r2 * and peak spacing d * :

[0074] r1 * =(F i1 +f1 * ) / 2f1

[0075] r2 * =(F i2 +f2 * ) / 2f2

[0076] d * =D i -(dD i )

[0077] The characteristic parameters of the processed waveform will change, resulting in a gap with the characteristic parameters of the test waveform, so further iterative calculation is required to narrow the gap.

[0078] 6.7. Return to step 6.2, and take out the second neutron waveform and the third neutron waveform in the neutron data set and the gamma data set S2 in order, and make the peak factor r i1 =r1 * 、r i2 =r2 * and peak spacing D i =d * , reprocess the waveform until the peak value f1 * With peak F i1 and peak value f2 * With peak F i2 The peak error is less than 1%, the peak spacing d and the peak spacing D i The error is less than 2 sampling points, and all neutron waveforms in the neutron data set and the gamma data set S2 are traversed to complete the training set sample library of the neutron + neutron stacking method, which is recorded as the first training sample library; in this embodiment, the first waveform traverses the neutron waveform once, and the second waveform traverses the neutron waveform once, and 40,000 neutron + neutron stacking waveforms can be constructed.

[0079] 6.8. According to steps 6.1 to 6.7, a training set sample library for neutron + gamma stacking is constructed as the second training sample library, a training set sample library for gamma + neutron stacking is constructed as the third training sample library, and a training set sample library for gamma + gamma stacking is constructed as the fourth training sample library; and the first training sample library, the second training sample library, the third training sample library and the fourth training sample library are taken as the training set sample library P2;

[0080] Similarly, we build a training set sample library of 40,000 neutron + gamma stacking methods, a training set sample library of 40,000 gamma + neutron stacking methods, and a training set sample library of 40,000 gamma + gamma stacking methods, and establish a training set sample library P2 containing four stacking methods. The labels are marked according to the four stacking methods of neutron + neutron, neutron + gamma, gamma + neutron, and gamma + gamma, which are 1, 2, 3, and 4 respectively.

[0081] Step 7: Construct training data set P3:

[0082] The K-nearest neighbor method is used to calculate the similarity, and the training samples corresponding to the test set P1 are selected from the training set sample library P2 according to the similarity to construct the training data set P3; the training data set P3 contains 300 samples of each of the four training set samples, namely neutron + neutron, neutron + gamma, gamma + neutron, and gamma + gamma, which are used as the input of the neural network, avoiding overfitting caused by the input of a large number of irrelevant training sets, and improving the calculation efficiency and accuracy.

[0083] Step 8: Use the training data set P3 as the input of the neural network, train the neural network, use the trained neural network to identify the test set P1, and output its predicted label to complete the neutron gamma pulse waveform stack identification under high counting rate conditions:

[0084] like Figure 7 As shown, the neural network uses a CNN convolutional neural network, and each line of the CNN convolutional neural network is defined as a training sample, and the number of lines is the number of training samples. The one-dimensional vector of the training data set P3 is reconstructed using the function reshape(trai n,M,N,D,SN) in the CNN convolutional neural network. train is the training data set P3, M is the length of a single training sample, N is the width of a single training sample, D is the number of channels, and SN is the number of training samples.

[0085] The CNN convolutional neural network is trained using the corresponding training data set P3. Since the training samples are one-dimensional vectors, N=1 and D=1 are set. After the training set is input into the CNN convolutional neural network, it passes through the convolutional layer, batch normalization layer, activation layer, and pooling layer. Six convolutional layers and pooling layers are set. Finally, it passes through the fully connected layer, sofmax layer, and output layer to output the predicted label. Among them, the activation function is re lu, the gradient descent algorithm is Adam, the regularization parameter is L2, the convolution kernel size is 5×1, and the pooling layer size is 3×1. The CNN convolutional neural network predicts all test waveforms in the test set P1 and outputs its predicted labels to complete the identification of all waveforms in the test set P1.

Claims

1. A neutron gamma pulse waveform stacking identification method under high counting rate conditions, characterized in that: The following steps are involved: 1] Collect and store neutron gamma single waveform mixed signals; 2] Identify the neutron gamma single waveform signal and obtain the neutron data set and gamma data set S1; 3] Using the neutron data set and the gamma data set S1, construct the test set P1 of the neutron gamma stack waveform; 4] Select the neutron waveforms and gamma waveforms of different amplitudes in the neutron data set and the gamma data set S1 according to the rise time and the count, and obtain the neutron data set and the gamma data set S2; 5] Extract the characteristic parameters of all stacked waveforms in the test set P1 and construct a characteristic parameter set T corresponding to the test set P1; 6] Based on the test set P1, the neutron data set and the gamma data set S2 and the characteristic parameter set T, the training set sample library P2 is constructed; 7] Select the training samples corresponding to the test set P1 from the training set sample library P2 to construct the training data set P3; 8] The training data set P3 is used as the training set of the neural network. The neural network is trained and used to identify all waveforms in the test set P1 and output their predicted labels to complete the neutron gamma pulse waveform stacking identification under high counting rate conditions.

2. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 1, characterized in that: Step 1] Specifically: A neutron tube is used as the neutron source, and a pulsed neutron detection system is used to collect and store more than or equal to 50,000 neutron gamma single waveform mixed signals.

3. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 2, characterized in that: Step 2] Specifically: The rise time method is used to identify the neutron gamma single waveform mixed signal to obtain the neutron waveform and gamma waveform. The peak positions of all neutron waveforms and gamma waveforms are aligned to form the neutron data set and gamma data set S1.

4. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 3, characterized in that: Step 3] Specifically: Neutron waveforms and gamma waveforms are randomly extracted from the neutron data set S1 and the gamma data set S1, and a test set P1 of neutron-gamma stacked waveforms is constructed according to four stacking methods: neutron + neutron, neutron + gamma, gamma + neutron, and gamma + gamma.

5. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 4, characterized in that: Step 5] Specifically: According to the peak value F1, peak value F2 and peak distance D of the stacked waveform, the characteristic parameters of each waveform in the test set P1 are extracted, and a characteristic parameter set T corresponding to the test set P1 is constructed.

6. According to the method for distinguishing neutron gamma pulse waveform stacking under high counting rate conditions of claim 5, it is characterized in that step 6] specifically comprises: 6.

1. Construct a training sample library of neutron + neutron stacking mode for each test waveform in the test set P1; 6.

2. Extracting a first neutron waveform and a second neutron waveform from the neutron data set and the gamma data set S2; 6.

3. The peak value F of the i-th test waveform in the characteristic parameter set T i1 The ratio of the peak value f1 of the first neutron waveform is recorded as the peak factor r i1 ; The peak value F of the i-th test waveform in the characteristic parameter set T is i2 The ratio of the peak value f2 of the second neutron waveform is recorded as the peak factor r i2 ; i=1,2,...,N, N is the number of test waveforms in the test set P1, N≥2; 6.

4. Multiply the first neutron waveform by the peak factor r i1 , the second neutron waveform multiplied by the peak factor r i2 , so that the two neutron waveforms are respectively corresponding to the peak value F i1 , peak F i2 equal; 6.

5. Shift the second neutron waveform back by the peak spacing D of the i-th test waveform i Then, the first neutron waveform is added to the second neutron waveform, and the peak value f1 of the stacked waveform is extracted. * , peak value f2 * and peak spacing d; 6.

6. Calculate the new peak factor r1 according to the following formula: * 、r2 * and peak spacing d * : r1 * =(F i1 +f1 * ) / 2f1 <h2 style=";text-align:left;direction:ltr">r2<h2 style=";text-align:left;direction:ltr"> * <h2 style=";text-align:left;direction:ltr"> (F)<h2 style=";text-align:left;direction:ltr"> i2 <h2 style=";text-align:left;direction:ltr"> +f2<h2 style=";text-align:left;direction:ltr"> * <h2 style=";text-align:left;direction:ltr"> ) / 2f2 d * =D i -(d-D i ); 6.

7. Return to step 6.2, and take out the second neutron waveform and the third neutron waveform in the neutron data set and the gamma data set S2 in order, and make the peak factor r i1 =r1 * 、r i2 =r2 * and peak spacing D i =d * , reprocess the waveform until the peak value f1 * With peak F i1 and peak value f2 * With peak F i2 The peak error is less than 1%, the peak spacing d and the peak spacing D i The error is less than 2 sampling points, and all neutron waveforms in the neutron data set and the gamma data set S2 are traversed to complete the training set sample library of the neutron + neutron stacking method, which is recorded as the first training sample library; 6.

8. According to steps 6.1 to 6.7, the training set sample library constructed by the neutron + gamma stacking method is recorded as the second training sample library, the training set sample library constructed by the gamma + neutron stacking method is recorded as the third training sample library, and the training set sample library constructed by the gamma + gamma stacking method is recorded as the fourth training sample library; The first training sample library, the second training sample library, the third training sample library and the fourth training sample library are used as the training set sample library P2.

7. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 6, characterized in that: Step 7] Specifically: The K-nearest neighbor method is used to calculate the similarity. According to the similarity, the training samples corresponding to the test set P1 are selected from the training set sample library P2 to construct the training data set P3.

8. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 7, characterized in that: Step 8] Specifically: 8.

1. Use the training data set P3 as the input of the neural network, train the neural network, and define each row as a training sample, and the number of rows is the number of training samples; 8.

2. Use the trained neural network to identify all test waveforms in the test set P1 and output their predicted labels to complete the stacking identification of neutron gamma pulse waveforms under high counting rate conditions.

9. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 8, characterized in that: In step 3], the stacked waveform is obtained by shifting any one of the two waveforms backward by 100 to 400 sampling points and then adding the two waveforms together to obtain the stacked waveform.

10. The method for distinguishing neutron gamma pulse waveform stacks under high counting rate conditions according to claim 9, characterized in that: In step 8.1, the neural network is a CNN convolutional neural network.

Images