Neutron-gamma discrimination method based on step gradient method

Abstract

The application discloses a neutron-gamma discrimination method based on a ladder gradient method, comprising filtering a radiation mixed field pulse signal; introducing the filtered neutron-gamma pulse signal into a quasi-continuous pulse emission cortex model to extract a firing map containing dynamic information; calculating a ladder gradient value R according to the neutron-gamma firing map; and discriminating neutron-gamma pulse data by taking the ladder gradient value R as a discrimination factor. The application generates a firing map required in a ladder gradient calculation process by proposing a quasi-continuous pulse emission cortex model, and obtains a ladder gradient value by calculation, which is taken as a discrimination factor for neutron-gamma discrimination. Compared with an existing pulse coupled neural network, the application can realize better detail information extraction performance and noise processing capability, has less firing and artificial parameter adjustment requirement, and has better performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radiation mixed field pulse signal processing, in particular to a neutron-gamma discrimination method based on a ladder gradient method. BACKGROUND

[0002] Since the discovery of neutrons by Chadwick in 1932, neutron detection technology has made great progress in many fields. Such as nuclear reactors, meteorology, national defense, aerospace, biology and radiopharmaceuticals, etc. Due to the existence of non-elastic scattering and slow neutron radiation capture interactions between neutrons and the surrounding environment, gamma rays always exist with neutrons, and detectors sensitive to neutrons are also very sensitive to gamma rays, making it difficult to distinguish between neutrons and gamma rays. The detector detects neutrons while also detecting gamma signals, making it particularly difficult to calculate the number of neutrons per unit time. In order to overcome this difficulty, many scholars have attempted to distinguish between neutrons and gamma rays by their different sensitivity characteristics to the two particles, which can be reflected by the difference in pulse shape, thereby giving birth to pulse shape discrimination technology (PSD). At present, this technology has been widely used in various scientific fields to meet the requirements of neutron detection. The most important thing in PSD technology is the discrimination algorithm, which can achieve information acquisition for each signal obtained through the algorithm, and achieve neutron-gamma discrimination through the calculation of the discrimination factor.

[0003] In the past few decades, many discrimination methods have emerged, such as the most commonly used charge comparison method, the frequency gradient method based on the frequency domain, the fractal spectrum method, and the zero-crossing comparison method that can quickly discriminate, etc. In 2021, Liu Haoran et al. first proposed applying the pulse coupled neural network method to neutron-gamma discrimination and achieved significant discrimination effect, which was significantly better than the traditional charge comparison method and zero-crossing comparison method. They attributed the outstanding discrimination effect of the pulse coupled neural network to its dynamic information extraction capability. As a kind of neuron network based on biological neural research, the pulse coupled neural network was originally conceived to mimic the working way of the biological neuron cortex to obtain information processing inside the picture or video. In the biological visual system, when the external light source stimulates the photoreceptor cells in the retina, these cells generate electrical signals (spikes) and transmit these spikes to the optic nerve adjacent to them. These electrical signals will stimulate neurons in the cortex, leading to further spike generation and transmission between cell assemblies. Biological neural research has confirmed that this spike behavior between cell assemblies can identify the information contained in the original stimulus received by the photoreceptor cells and enable the brain to understand the features, details and other information of the image or video. According to this working way, the pulse coupled neural network can also extract the dynamic information of the image. Since it was proposed by Johnson et al. in 1994, it has been widely applied in the field of image processing. For example, image recognition, image shadow removal and image feature extraction, etc.

[0004] Although Liu Haoran et al. have proved the discrimination effect of this pulse coupled neural network method, the high computational complexity limits its rapid discrimination application. This computational difficulty comes from two parts: the high iteration number of the pulse coupled neural network and the integral process of the firing map. Therefore, how to optimize the existing method to reduce the computational pressure and obtain better detail information extraction performance and noise processing ability is a problem that needs to be further studied and solved by those skilled in the art. SUMMARY

[0005] In view of the above problems in the prior art, the present application provides a neutron-gamma discrimination method based on the ladder gradient method, which generates the firing map required by the ladder gradient calculation process by proposing a quasi-continuous pulse firing cortex model, and replaces the integral process of the previous pulse coupled neural network method with the gradient calculation of the ladder gradient method, thereby reducing the computational pressure of the discrimination method, achieving better detail information extraction performance and noise processing ability, and having less firing and artificial parameter adjustment requirements.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0007] A neutron-gamma discrimination method based on the ladder gradient method, comprising the following steps:

[0008] S10, obtaining a radiation mixed field pulse signal for filtering processing to obtain a neutron-gamma pulse signal;

[0009] S20, introducing the filtered neutron-gamma pulse signal into a quasi-continuous pulse firing cortex model to extract a firing map containing dynamic information, wherein each neutron-gamma pulse signal corresponds to a firing map, and has the same vector as the original neutron-gamma pulse signal, and the quasi-continuous pulse firing cortex model is constructed by introducing a continuous structure into the pulse firing cortex model and for one-dimensional signal data;

[0010] S30, calculating the ladder gradient value R according to the neutron-gamma firing map, wherein the ladder gradient is defined as the slope of the line between the peak point and the mth mode point after the peak point in the firing map, and m is an empirical parameter related to the neutron-gamma pulse shape characteristics;

[0011] S40, using the ladder gradient value R as a discrimination factor to discriminate the neutron-gamma pulse data.

[0012] Specifically, the quasi-continuous pulse firing cortex model is represented as:

[0013]

[0014] θ i (t+Δt)=gΔt θ i (t)+hY i (t+Δt)

[0015]

[0016] In the formula, U i represents the membrane potential of the neuron located at the i th sampling point of the neutron-gamma pulse signal; Δt is a parameter for determining the time continuity characteristic of the quasi-continuous pulse firing cortex model, and the value range of Δt is between 0 and 1, and the closer Δt is to 0, the closer the quasi-continuous pulse firing cortex model is to a continuous time system; f △t is the attenuation coefficient of U i ; S i is the external stimulus received by the neuron, that is, each neutron-gamma pulse signal; W ij is a synaptic weight matrix, which controls the connection between the central neuron located at the i th sampling point of the neutron-gamma pulse signal and the peripheral neuron located at the j th sampling point of the neutron-gamma pulse signal; Y i and Y j are the neuron pulse outputs located at the i th sampling point of the neutron-gamma pulse signal and the j th sampling point of the neutron-gamma pulse signal, respectively; θ i is a dynamic threshold; g Δt is the attenuation coefficient of θ i ; h is the attenuation coefficient of Y i .

[0017] Specifically, the calculation formula of the ladder gradient value is as follows:

[0018]

[0019] In the formula, (x A , y A ) and (x B , y B ) are the coordinate values of the position of the peak point and the position of the m th mode after the peak point in the ignition mapping respectively.

[0020] Compared with the prior art, the present application has the following beneficial effects:

[0021] The present application generates the ignition mapping required in the ladder gradient calculation process by proposing the quasi-continuous pulse firing cortex model, and obtains the ladder gradient value by calculation, which is used as a discrimination factor for neutron-gamma discrimination. Compared with the existing pulse coupled neural network, the present application can realize better detail information extraction performance and noise processing capability, has fewer ignition and artificial parameter adjustment times, and has better performance.

[0022] The quasi-continuous pulse firing cortex model improves the conventional pulse firing cortex model SCM, first uses the SCM to extract the dynamic information contained in the radiation pulse signal, not only considers the pulse amplitude at each point of the pulse signal, but also considers the attenuation trend of the pulse signal, effectively includes the difference between the falling edge and the afterglow effect positions of the neutron gamma signal from the overall and dynamic point of view, so that the difference between the two types of radiation pulse signals is obtained, thereby achieving the purpose of efficiently discriminating the two types of particles; secondly, the quasi-continuous pulse firing cortex model QC-SCM also uses quasi-continuous iteration, which is different from the discrete iteration of the traditional SCM, realizes the collection of more detailed information in the signal to be processed, further enhances the anti-noise performance, so that the QC-SCM model can achieve similar effects to the traditional SCM with fewer iteration times, improves the data processing efficiency, and the enhancement of the anti-noise ability further suppresses the noise caused by the random fluctuation of the current and voltage of the radiation detection system in the signal to be processed, so that the final obtained ignition map is more stable, especially the fluctuation between 100-200ns is eliminated.

[0023] Because the QC-SCM obtains a stable ignition map, the present application further uses the ladder gradient method LG to calculate the discrimination factor directly on this basis, which can effectively extract the information in the falling edge and afterglow effect parts of the neutron gamma pulse ignition map, thereby realizing effective discrimination. In contrast, if the ladder gradient method is used to calculate the discrimination factor on the ignition map produced by the traditional SCM, the discrimination result will be particularly poor, because the ignition map produced by the SCM has obvious fluctuations in the interval from the falling edge to the afterglow effect, and the ladder gradient is particularly sensitive to the fluctuations caused by such noise, directly leading to poor discrimination effect, so the traditional SCM discrimination can only use the integral method to extract the information in the neutron gamma pulse ignition map to calculate the discrimination factor and perform discrimination. And compared with the integral calculation of the discrimination factor, the calculation amount of the ladder gradient method is obviously reduced. In addition, the QC-SCM used in the present application has higher processing efficiency than the SCM, and the calculation burden of the ladder gradient method in generating the ignition map is also reduced, so the present application method shows that under the condition of the same order of time consumption as the traditional neutron gamma discrimination method, the discrimination effect (strong anti-noise, high FoM value) is much better than the traditional method. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The flowchart of the embodiment of the present application.

[0025] Figure 2 The structure diagram of the ladder gradient method in the embodiment of the present application.

[0026] Fig. 3 is a neutron-gamma pulse signal and its firing map in the embodiment of the present application, Figure 3a is a pulse signal diagram of neutrons and gamma rays, Figure 3b is a firing map generated based on the pulse-coupled neural network method, Figure 3c is a firing map generated based on the quasi-continuous pulse firing cortex model.

[0027] Fig. 4 is a scatter plot of the effects of five comparative neutron-gamma discrimination methods in the embodiment of the present application, wherein Figure 4a is a scatter plot of the effect of the zero-crossing method, Figure 4b is a scatter plot of the effect of the charge comparison method, Figure 4c is a scatter plot of the effect of the frequency step analysis method, Figure 4d is a scatter plot of the effect of the pulse-coupled neural network, Figure 4e is a scatter plot of the effect of the step gradient method.

[0028] Figure 5 is a scatter plot of the effects of five comparative neutron-gamma discrimination methods in the embodiment of the present application.

[0029] Fig. 6 is a scatter plot of the effects of five methods considering optimal filtering, wherein Figure 6a is the zero-crossing method, Figure 6b is the charge comparison method, Figure 6c is the frequency step analysis method, Figure 6d is the pulse-coupled neural network, Figure 6e is the step gradient method.

[0030] Figure 7 is a scatter plot of the effects of five methods considering optimal filtering. DETAILED DESCRIPTION

[0031] The present application will be further described below in conjunction with the drawings and embodiments, and the embodiments of the present application include but are not limited to the following embodiments.

[0032] EMBODIMENT

[0033] As Figure 1 shown, the neutron-gamma discrimination method based on the step gradient method includes the following steps:

[0034] S10, filtering the pulse signal of the mixed radiation field to reduce the interference of noise on the discrimination effect, and obtaining a neutron-gamma pulse signal;

[0035] S20, the filtered neutron-gamma pulse signal is introduced into a quasi-continuous pulse firing cortex model to extract a firing map containing dynamic information, wherein each neutron-gamma pulse signal corresponds to a firing map, and has the same vector as the original neutron-gamma pulse signal, and the quasi-continuous pulse firing cortex model is constructed by introducing a continuous structure into a pulse firing cortex model and for one-dimensional signal data;

[0036] S30, according to the neutron-gamma firing map, the step gradient value R in each firing map is calculated by using a step gradient method, and the step gradient is defined as the slope of the line between the peak point and the mth mode point after the peak point in the firing map, wherein m is an empirical parameter related to the neutron-gamma pulse shape characteristics;

[0037] S40, according to the final calculation result, the final step gradient value is taken as a discrimination factor to discriminate the neutron-gamma pulse data.

[0038] The calculation formula of the step gradient value is:

[0039]

[0040] In the formula, (x A , y A ) and (x B , y B ) are the coordinate values of the position of the peak point and the position of the mth mode point after the peak point in the firing map, and the slope between the two points is the step gradient value, as shown in Figure 2 The previous pulse coupled neural network discrimination method calculates the discrimination factor by calculating the integral process of the falling edge of the neutron-gamma pulse signal and the delayed fluorescence part. The discrimination factor calculation based on the step gradient method is relatively simple, however, the calculation based on the step gradient method is simple, but it has the disadvantages of poor noise sensitivity and low information extraction performance. Since the influence of noise on the coordinates of the two points is obviously greater than the influence of the firing map integration, and the information of the firing map integration is much greater than the information of the two-point slope. Therefore, for the step gradient method, an optimized model needs to be obtained to achieve better noise resistance and stronger information extraction ability.

[0041] The method generates a neutron-gamma pulse ignition mapping by a quasi-continuous pulse firing cortex model. The quasi-continuous pulse firing cortex model is newly proposed according to an existing pulse coupled neural network developed in the field of image processing. The existing pulse coupled neural network model is a non-integer step pulse coupled neural network model, which is used to simulate the continuous structure of biological neurons, so as to better identify the detailed information inside the image and have a stronger noise processing capability, but has a higher calculation burden. Therefore, the inventor introduces the continuous structure into the firing cortex model to simplify and optimize the existing pulse coupled neural network model, so as to realize the processing capability of the detailed information of the pulse signal and the ignition resolution, while keeping a lower calculation burden. The mathematical expression of the quasi-continuous pulse firing cortex model in the application is:

[0042]

[0043] θ i (t+Δt)=g Δt θ i (t)+hY i (t+Δt)

[0044]

[0045] In the formula, U i represents the membrane potential of the neuron at the i-th sampling point of the neutron-gamma pulse signal; Δt is a parameter for determining the time continuity of the quasi-continuous pulse firing cortex model, and the value range is between 0 and 1, and the closer Δt is to 0, the closer the quasi-continuous pulse firing cortex model is to the continuous time system; f △t is the attenuation coefficient of U i ; S i is the external stimulus received by the neuron, that is, each neutron-gamma pulse signal; W ij is a synaptic weight matrix, which controls the connection between the central neuron at the i-th sampling point of the neutron-gamma pulse signal and the surrounding neuron at the j-th sampling point of the neutron-gamma pulse signal; Y i and Y j are the neuron pulse outputs at the i-th sampling point of the neutron-gamma pulse signal and the j-th sampling point of the neutron-gamma pulse signal, respectively; θ i is a dynamic threshold; g Δt is the attenuation coefficient of θ i ; h is the attenuation coefficient of Y i .

[0046] As shown in FIG. 3, the neutron-gamma pulse signal and its ignition mapping are shown. Figure 3a is the pulse signal diagram of the neutron and gamma rays;Figure 3b This is an ignition map generated based on the pulse-coupled neural network method; Figure 3c This is an ignition mapping diagram generated based on a quasi-continuous pulse firing cortical model. For example... Figure 3a As shown, the signal difference between neutrons and gamma rays is mainly reflected in the pulse falling edge (approximately 90 ns) to the delayed fluorescence portion (approximately 180 ns). Figure 3b As shown, this difference was successfully captured and amplified using the pulse-coupled neural network method, where the ignition times of these two parts are typically higher. However, the ignition pattern based on the pulse-coupled neural network method is not stable, fluctuating significantly between 100 and 200 ns. This fluctuation is mainly caused by noise introduced by the radiation detection system. Although the noise-induced fluctuations can be compensated for by integrating the falling edge and the delayed fluorescence interval, the increase in computational complexity is unavoidable.

[0047] In comparison, such as Figure 3c As shown, the ignition map obtained by this invention based on a quasi-continuous pulse emission cortical model does not exhibit step-trapezoidal fluctuations. This stable step-trapezoidal shape is due to the superior noise handling capability of the quasi-continuous pulse emission cortical model compared to the pulse-coupled neural network model. Therefore, based on the ignition map obtained by this quasi-continuous pulse emission cortical model, the step gradient value is calculated using the step gradient method as a discrimination factor to evaluate neutron-gamma discrimination, which is more convenient than integral calculation. Furthermore, because the quasi-continuous pulse emission cortical model has better information extraction capabilities, it can achieve difference amplification performance similar to that of the pulse-coupled neural network, but with fewer iterations. The number of manually tuned parameters for the quasi-continuous pulse emission cortical model is far less than the number of parameters for the pulse-coupled neural network. Similar to the pulse-coupled neural network, the quasi-continuous pulse emission cortical model does not require any training process before discrimination.

[0048] Introduction to the filtering method in step S10

[0049] Neutron-gamma pulse data is a one-dimensional time series signal. A one-dimensional signal can be considered as a useful signal superimposed with white Gaussian noise, represented as:

[0050] s(n) = j(n) + ke(n)

[0051] In the formula, n represents time, s represents a one-dimensional signal, j represents the useful signal, and ke represents Gaussian noise. In practical applications, the sampled signal is a discrete-time signal with equal time steps. Therefore, it can be represented by an N-dimensional random vector s(n):

[0052]

[0053] The noise reduction process is to extract useful signal j(n) from original signal s(n). It is worth noting that there is a significant difference between the traditional noise reduction method and the neutron-gamma discrimination noise reduction method. In the field of neutron-gamma discrimination, as long as the pulse shape difference between neutron and gamma rays is amplified, even if there is slight distortion in the noise reduction process, it is acceptable. Of course, under the same discrimination performance, the filter with less signal distortion is selected first. Therefore, the inventors determine that the best filter suitable for the staircase gradient method is an elliptical filter, a moving average filter or a wavelet filter by adopting 11 filtering methods combined with 17 filtering conditions. These filtering methods include Butterworth filter, Chebyshev filter, elliptical filter, median filter, moving average filter, Fourier filter, wavelet filter, Wiener filter, least mean square adaptive filter, morphological filter and windowed sine filter.

[0054] Further explanation of the application compared with the conventional pulse cortex model, in the image and video processing work, each neuron in the conventional pulse cortex model SCM network is closely connected with the surrounding 8 neurons, because the number and position of neurons in SCM are one-to-one corresponding to the pixel points in the image to be processed, so SCM considers the influence of the gray value of the 8 pixel points around each pixel point in the image when processing each pixel point. In the neutron-gamma discrimination of the application, the signal to be processed is a one-dimensional signal, and each neuron in the QC-SCM network corresponds to each sampling point of the radiation pulse signal, so each neuron only needs to be coupled with the previous and the next neuron, that is, when QC-SCM processes the amplitude of each sampling point of the pulse signal, it only needs to consider the influence of the amplitudes of the previous and the next sampling points, and the amount of data to be processed is greatly reduced. The effect of the firing map is mainly that in the working process of QC-SCM, the signal to be processed is repeatedly fired into the network, and after stimulating the neurons for many times, QC-SCM can analyze the importance of the dynamic information contained in each sampling point to obtain the firing map; in the firing map, the information of the radiation pulse signal is extracted and summarized, and the noise is suppressed to a certain extent, so that the difference between the neutron and gamma pulses is amplified.

[0055] For the evaluation of the neutron-gamma discrimination result, the evaluation standard involves quality factor, noise reduction signal similarity measurement, time-consuming time, discrimination accuracy.

[0056] The figure of merit FoM is a standard measure for evaluating the discrimination effect of neutrons and gamma rays. First, the calculation of FoM requires drawing a discrimination result histogram, mainly including a neutron counting spectrum and a gamma ray counting spectrum. Then, a Gaussian fitting function is used to fit the two sets of data to form a fitting curve, which can be further used to calculate the distance S between the two peaks and the half-height width of the two peaks. Finally, FoM is defined as:

[0057]

[0058] where S represents the interval between the neutron peak and the gamma peak, FWHM γ represents the full width at half maximum of the gamma peak (i.e. the full width at half the peak height), FWHM n is the full width at half maximum of the neutron peak. If the discrimination performance is good, S should be larger, the FoM value should be larger, and the discrimination performance should be better; it is worth pointing out that in the application of neutron gamma discrimination based on plastic scintillators, if the FoM value is higher than 1.0, it can be considered as effective discrimination; if the FoM value is higher than 1.6, it can be considered as excellent discrimination.

[0059] Noise reduction signal similarity measure. In the field of image noise reduction, there are several objective measurement methods. Using this characteristic, it can be calculated under one-dimensional conditions and applied in the noise reduction of radiation pulse signals to realize the evaluation of the noise reduction performance.

[0060] The peak signal-to-noise ratio PSNR can evaluate the similarity between the denoised signal y(n) and the original signal s(n), and is specifically defined as:

[0061]

[0062] where N is the length of the original signal;

[0063] The mean square error MSE is defined as:

[0064]

[0065] The root mean square error RMSE can be calculated as follows:

[0066]

[0067] DIV can be used to measure the noise reduction performance, and is defined as:

[0068]

[0069] where σ y and σ s represent the variances of the denoised signal and the original signal, respectively.

[0070] Shannon entropy is a concept in information theory, which can represent the average information contained in a signal. It can be estimated as:

[0071] SE = -∑ n S 2 (n)log(s 2 (n))

[0072] By calculating SE of s(n) and y(n), the following entropy difference ED is obtained:

[0073] ED = |SE(s) - SE(y)|

[0074] When the values of MSE, RMSE and DIV are close to 0, the value of PSNR is more considerable, the denoised signal is more similar to the original signal, indicating that the filtering effect is better and the signal distortion is smaller. When the value of ED is smaller, the damaged information in the filtering process is less, so that the filtering result is better.

[0075] Regarding the time-consuming time, in the present application, the ladder gradient method and other discrimination methods are used for 9414 neutron-gamma pulse signals, and the time-consuming time of each method is calculated. At the same time, the time-consuming time of filtering for each discrimination method is also calculated. The time spent by the filter to process all signals is measured. The time-consuming time fully reflects the calculation complexity of the neutron-gamma discrimination method and the filtering method, and illustrates their real feasibility.

[0076] Regarding the discrimination accuracy, when the neutron-gamma pulse signal is smoothed by the filter, it will affect the performance of neutron-gamma discrimination to some extent. We record the results of the number of neutron and gamma ray particles filtered by each filter. At the same time, since the Fourier filter is widely used in the field of pulse shape discrimination, the discrimination results of the signals processed by the Fourier filter are used as the standard reference in the present application. The number of pulse signals discriminated incorrectly after processing by other filters is recorded as the error rate. The error rate is defined as:

[0077]

[0078] Wherein, the smaller the error value and error rate, the better the result.

[0079] The following verifies the discrimination results and comparison through experiments:

[0080] The experimental equipment and parameter settings are as follows: 241Am-Be neutron source is selected as the radiation source, which can generate neutrons with an average energy of 4.5 MeV. In this experiment, EJ299-33 plastic scintillator is used as the detector, and high-performance digital oscilloscope is used to realize offline neutron-gamma discrimination experiment. The model of the digital filter is TPS2000B, the sampling rate is 1 GS / s, the resolution is 10 bits, and the bandwidth is 200 MHz. The trigger threshold is set to 500 mV, which is equivalent to an energy of 1.6 MeVee. The pulse acquisition duration point is 160 ns, and the pulse signal information collected meets the standard of the Shannon sampling law. After acquiring the mixed field neutron-gamma pulse signal, the discrimination algorithm and filtering process are carried out offline using AMD 5900X CPU on Windows 11. The main parameters of the ladder gradient discrimination method are as follows: Δt = 0.5, W ij = [0.44, 0, 0.44], f △t = 0.38, g Δt = 0.8, h = 8.45.

[0081] The following four discrimination methods are compared with the ladder gradient method LG, including zero-crossing comparison method ZC, charge comparison method CC, frequency gradient analysis method FGA, and pulse coupled neural network PCNN. In the subsequent experiment, the adjustable parameters of the above discrimination methods have been optimized to achieve the best discrimination effect of each method itself.

[0082] As shown in FIG. 4, the scatter plot effect diagram of neutron-gamma discrimination using the above five discrimination methods is shown in FIG. 4, Figure 4a the zero-crossing comparison method ZC, Figure 4b the charge comparison method CC, Figure 4c the frequency gradient analysis method FGA, Figure 4d the pulse coupled neural network PCNN, Figure 4e the ladder gradient method LG; in each scatter plot, if the two clusters in the scatter plot each present a Gaussian distribution with small variance, the gap between the clusters is large, and the number of points scattered between the clusters and on both sides of the clusters is small, then the discrimination effect is excellent. It can be seen that:

[0083] The discrimination effect of the zero-crossing comparison method ZC, ignoring the influence of the pulse delay fluorescence on the method, although the differentiation and integration processes help the method to reduce noise interference, but considering the difference in incomplete pulse shape still has a bad influence on the discrimination effect. Although the Gaussian distribution of its scatter plot is better than that of the falling edge percentage slope method, there are still a large number of pulse signals between the two peaks, resulting in poor discrimination results.

[0084] The discrimination performance of the charge comparison method CC, the frequency gradient analysis method FGA and the pulse coupled neural network PCNN is much better than that of the zero-crossing comparison method, mainly reflected in the obvious separation of the Gaussian distribution diagram of the scatter diagram, the main reason being that the three methods all consider the information contained in the pulse delay fluorescence and use integration to achieve noise immunity. Among them, the charge comparison method integrates the amplitude of the neutron-gamma pulse signal, the frequency gradient analysis method uses the first component of the Fourier transform form of the pulse signal, which corresponds to the average amplitude of the entire pulse signal, and the pulse coupled neural network model integrates the ignition mapping diagram and the ignition time. However, the discrimination method based on the step gradient of the present application can achieve similar discrimination effect as the three methods without the integration process, and in the discrimination process, the influence of noise is pretreated by using the quasi-continuous pulse firing cortex model, so that the discrimination factor can be obtained without the integration process. On the contrary, the step gradient method has a clear advantage in computational complexity; more notably, in the preliminary comparison of various discrimination methods here, in order to control the experimental variables and compare them under the same conditions as much as possible, the zero-crossing comparison method ZC, the charge comparison method CC, the frequency gradient analysis method FGA and the pulse coupled neural network PCNN all use a Fourier filter to filter and pretreat the pulse. Through the pretreatment of the pulse signal by the Fourier filter, the influence of noise in the original pulse signal on the effect is reduced, which improves the effect of the four discrimination methods, while the step gradient method of the present application directly uses the original pulse signal as the basic data, because the pre-noise reduction performance of the Fourier filter cannot help the LG method which has excellent noise immunity, but on the contrary, it destroys part of the effective information in the radiation pulse signal in the filtering process, and reduces the effect of the LG method. That is, the discrimination effect achieved by the method of the present application even directly processing the original pulse signal reaches the level of conventional discrimination methods, and significantly exceeds the effect of ordinary discrimination methods.

[0085] As shown in Table 1, the above five discrimination methods are objectively evaluated. As shown in Fig. 1, the scatter diagram of the zero-crossing comparison method ZC is not very clear, and the discrimination effect is not very good. As shown in Fig. 2, the scatter diagram of the charge comparison method CC is better than that of the zero-crossing comparison method ZC, and the discrimination effect is better than that of the zero-crossing comparison method ZC. As shown in Fig. 3, the scatter diagram of the frequency gradient analysis method FGA is better than that of the charge comparison method CC, and the discrimination effect is better than that of the charge comparison method CC. As shown in Fig. 4, the scatter diagram of the pulse coupled neural network PCNN is better than that of the frequency gradient analysis method FGA, and the discrimination effect is better than that of the frequency gradient analysis method FGA. As shown in Fig. 5, the scatter diagram of the step gradient method of the present application is better than that of the pulse coupled neural network PCNN, and the discrimination effect is better than that of the pulse coupled neural network PCNN. Figure 5The discrimination factor curve shown can calculate the FoM value of each method. Among them, the discrimination effect of the zero-crossing comparison method is the worst, and the FoM value is about 1; the discrimination results of the charge comparison method and the frequency gradient analysis method are equivalent; the discrimination FoM value of the ladder gradient method is 1.54, which is slightly lower than the FoM value 1.75 of the pulse coupled neural network model. At the same time, the time-consuming time of each discrimination method is defined as: the CPU processing time of 9414 pulse signals of experimental data, and the time-consuming time excluding the filtering process. Compared with the pulse coupled neural network model method, the ladder gradient method reduces the time-consuming time by about 37%, close to the level of other traditional discrimination methods. The experimental results verify the efficiency and robustness of the ladder gradient method. It can achieve better performance than traditional methods without pre-processing through filters, and consume less time than pulse coupled neural networks. The calculation complexity of the method makes it possible to realize on the integrated radiation detection system, and thus realize real-time discrimination.

[0086] Table 1 Discrimination results and time-consuming time (CPU processing for 9414 pulses)

[0087] Criteria / Method ZC CC FGA PCNN LG FoM (a.u.) 1.09 1.38 1.47 1.75 1.54 Time consumption (sec) 1.35 1.30 1.31 2.78 1.76

[0088] The foregoing preliminary comparison on the basis of controlling experimental variables has shown the advantages of the LG method of the present application. According to some existing research, each discrimination method will show a certain differential discrimination effect when coupled with different filtering methods, as mentioned in the foregoing experiment that the Fourier filter has a gain effect on the conventional discrimination method, but has a certain negative effect on the ladder gradient method of the present application. That is, each discrimination method should have a corresponding most suitable filtering method to make it ultimately play the best discrimination effect. Therefore, the inventors further experimented on the effect of selecting the most suitable filtering method for each discrimination method on the basis of the comparison of the foregoing various discrimination methods. Other conditions are consistent with the foregoing experiment.

[0089] As shown in FIG. 6, the scatter plot of the foregoing five discrimination methods in considering the most suitable filtering to realize neutron-gamma discrimination is shown in FIG. 6, Figure 6a The zero-crossing comparison method ZC uses a median filter, Figure 6b The charge comparison method CC uses a moving average filter, Figure 6c The frequency gradient analysis method FGA uses a Fourier filter, Figure 6d The pulse coupled neural network PCNN uses an elliptical filter, Figure 6e The ladder gradient method LG uses an elliptical filter. As Figure 7 The discrimination factor curve of the foregoing five discrimination methods in considering the most suitable filtering is shown in FIG. 6, from which the FoM values of each are shown in Table 2. As can be seen,

[0090] In the case of considering the most adaptive filtering, the discrimination effect of each method is improved, and the improvement of the ladder gradient method of the application is the most obvious, which reaches the same level as PCNN, and the time consumption is obviously shorter than PCNN, and the discrimination effect is far superior to the other three traditional discrimination methods.

[0091] Further explanation of data, for the discrimination method itself and the filtering process itself, the time consumption of the same type of method is generally consistent, but since each experimental test process is independent, considering the external factors of the experimental equipment itself, there will be a little fluctuation in the statistical results, which is acceptable.

[0092] Table 2 Discrimination results and time consumption in the case of most adaptive filtering

[0093] Criteria / Method ZC CC FGA PCNN LG FoM (a.u.) 1.15 1.61 1.47 2.19 2.20 Time consumption (sec, without filtering) 1.25 1.31 1.45 2.89 1.71 Time consumption (sec, with filtering) 1.49 1.33 1.84 4.11 2.94

[0094] As can be seen from the above, the pulse coupled neural network PCNN and the ladder gradient method LG of the application are both excellent methods for neutron-gamma discrimination, and the inventors further analyze the effect comparison of the two in the case of considering the adaptability of the filter, and evaluate the performance of the two through four aspects, that is, the aforementioned evaluation standards: (A) discrimination ability (which is crucial in neutron-gamma discrimination application); (B) denoised signal similarity (measure distortion), (C) time (refers to the total CPU processing time of each filter for 9414 pulse signals); (D) discrimination accuracy (using the most commonly used discrimination method and filter method, here the charge comparison method with Fourier filter as reference, compares the discrimination results of other filters with the results of the Fourier filter). It is worth noting that the discrimination ability is not based on because the neutron-gamma pulse signals used in this study are actually measured signals. Although the charge comparison method with Fourier filter has been proved to be robust, its discrimination result still cannot be compared with the basic fact. Therefore, when the discrimination error rate is lower than 2%, it can be considered as successful discrimination.

[0095] According to the experimental results of various types of filters, as shown in Tables 3 and 4 below, three best filters suitable for PCNN and LG are given respectively, it is need to point out that the time Time is only the time consumption of filtering processing.

[0096] The three best filtering performances of the discrimination method based on PCNN are respectively wavelet filter, elliptical filter and median filter. Among them, the discrimination FoM value after filtering by the wavelet filter is the highest (2.5915), which is obviously better than the other two filtering methods, the main advantages are small pulse signal distortion and small discrimination error, and the main disadvantage is long time consumption; the discrimination FoM value after filtering by the elliptical filter is the second (2.1942), the main advantages are small discrimination error and short filtering time, and the main disadvantage is that the parameters need to be reset for different neutron sources and detector measurements, which limits its application range; the discrimination FoM value after filtering by the median filter is the third (2.1811), the main advantage is short filtering time, and the main disadvantage is serious signal distortion.

[0097] Table 3 Filtering performance of the pulse coupled neural network model

[0098]

[0099] The three best filtering performances of the discrimination method based on the ladder gradient method LG are respectively elliptical filter, moving average filter and wavelet filter. Among them, the elliptical filter has the best discrimination performance (FoM value is 2.1942) after filtering, the main advantages are small pulse signal distortion and fast discrimination time, and the main disadvantage is manual parameter adjustment; the discrimination effect of the moving average filter is the second (FoM is 1.6637) after filtering, the main advantage is the shortest time consumption; the wavelet filter has the best discrimination performance and noise reduction performance after filtering, and the main disadvantage is high computational complexity and long time consumption.

[0100] Table 4 Filtering performance of the ladder gradient method

[0101]

[0102] From the above analysis, when the method of the application adopts the most suitable filter, the FoM value can reach the level of 1.66-2.19, which is much higher than the traditional discrimination method and much higher than the excellent discrimination standard of 1.6 FoM value; compared with the FoM value of 2.18-2.59 when PCNN adopts the best filter, the method of the application can also reach the same level (2.19), and the time consumption is shorter, which realizes the final effect of high FoM and low time consumption.

[0103] The above examples are only preferred embodiments of the application, and are not a limitation on the protection scope of the application, as long as the design principle of the application is adopted, and changes made on the basis of non-creative labor should be within the protection scope of the application.

Claims

1. A neutron-gamma discrimination method based on a step gradient method, characterized in that, The method comprises the following steps: S10, obtaining a radiation mixed field pulse signal and performing filtering processing to obtain a neutron-gamma pulse signal; S20, introducing the filtered neutron-gamma pulse signal into a quasi-continuous pulse firing cortex model to extract a firing map containing dynamic information, wherein each neutron-gamma pulse signal corresponds to a firing map, and has the same vector as the original neutron-gamma pulse signal, and the quasi-continuous pulse firing cortex model is constructed by introducing a continuous structure into a pulse firing cortex model and for one-dimensional signal data; S30, calculating a ladder gradient value R according to the neutron-gamma firing map, wherein the ladder gradient is defined as the slope of the line between the peak point and the mth mode point after the peak point in the firing map, and m is an empirical parameter related to the neutron-gamma pulse shape characteristics; S40, using the ladder gradient value R as a discrimination factor to discriminate the neutron-gamma pulse data.

2. The method of claim 1, wherein the step of determining the neutron- gamma discrimination is performed using a step gradient method. The quasi-continuous pulse firing cortex model is represented as: θ i (t) = hX Δt (t+Δt) = g i (t) + hY i (t+Δt) In the formula, U i represents the membrane potential of the neuron at the i-th sampling point of the neutron-gamma pulse signal; Δt is a parameter that determines the time continuous characteristic of the quasi-continuous pulse firing cortical model, and its value range is between 0 and 1, the closer Δt is to 0, the closer the quasi-continuous pulse firing cortical model is to the continuous time system; f △t is the decay coefficient of U i ; S i is the external stimulus received by the neuron, that is, each neutron-gamma pulse signal; W ij is the synaptic weight matrix, which controls the connection between the central neuron located at the i th sampling point of the neutron-gamma pulse signal and the surrounding neuron located at the j th sampling point of the neutron-gamma pulse signal; Y i and Y j are the neuron pulse outputs respectively located at the i th sampling point of the neutron-gamma pulse signal and the j th sampling point of the neutron-gamma pulse signal; θ i is a dynamic threshold; g Δt is a decay coefficient of θ i ; h is a decay coefficient of Y i .

3. The method of claim 2, wherein the step of determining the neutron- gamma discrimination is performed by a method of step gradient. The calculation formula of the ladder gradient value is: where (x A , y A ) and (x B , y B ) are coordinate values of the position of the peak point and the position of the mth mode point after the peak point in the ignition map, respectively.

4. The neutron-gamma discrimination method based on the method of stepwise gradients according to any one of claims 1 to 3, characterized in that, The filtering processing in the step S10 adopts an elliptical filter, a moving average filter or a wavelet filter.

Images