KR20220145945A - Method and apparatus for distinguishing radionuclides - Google Patents

Abstract

The present invention relates to a method and an apparatus for distinguishing a radioactive nuclide and, more specifically, to a method and an apparatus for distinguishing a radioactive nuclide which can increase the accuracy of distinguishing a nuclide by combining an energy weighting technique and a machine learning technique. According to one embodiment of the present invention, the apparatus for distinguishing a radioactive nuclide based on a plastic scintillator comprises: a radiation input unit converting energy inputted from one or more radiation sources into light; a spectrum conversion unit collecting the light generated by the radiation input unit to convert the light into an energy spectrum; a peak detection unit comparing an energy value corresponding to the peak value of an energy weighting spectrum resulting from applying an energy weight to the energy spectrum with an energy value of the Compton edge of the radiation sources; a machine learning model unit analyzing features of the energy weighting spectrum, and executing learning on the overall shape of the energy weighting spectrum; and a nuclide distinguishing unit distinguishing a nuclide based on the peak value of the peak detection unit or feature information analyzed by the machine learning model unit.

Description

Radionuclide fractionation apparatus and method {METHOD AND APPARATUS FOR DISTINGUISHING RADIONUCLIDES}

The present invention relates to a radionuclide fractionation apparatus and method, and more particularly, to a radionuclide fractionation apparatus and method capable of increasing the precision of nuclide fractionation by combining an energy weighting technique and a machine learning technique.

A radionuclide detector is a device for discriminating unique radionuclides constituting the material by measuring and analyzing the energy of radiation generated from a material containing an unknown radionuclide.

An unstable atomic nucleus undergoes several nuclear decays to become a stable atomic nucleus, and at this time, alpha rays, beta rays, and gamma rays of specific energy are emitted. The scintillator used in the radionuclide detector has the property of emitting light when it receives energy. By measuring the amount of light generated by the radiation, the energy of the incident radiation can be confirmed.

Radionuclide detectors include a detector using an inorganic scintillator and a detector using a plastic scintillator.

In the case of a scintillation detector using an inorganic scintillator (e.g., NaI(TI) or CsI(TI)), the photoelectric effect absorbing all the energy of gamma rays occurs at a high rate and the energy resolution is excellent. ) based on the energy of radionuclides can be easily analyzed. However, it is expensive and difficult to apply to large-scale detectors.

In a radionuclide detector using a plastic scintillator, the photoelectric effect hardly occurs and the energy resolution is relatively low. Since the plastic scintillator is mainly composed of carbon (C) and hydrogen (H), Compton scattering is more likely to occur than the photoelectric effect when reacting with gamma rays. , a broad form of the Compton continuum appears. Therefore, in radiation screening using a plastic scintillator, an algorithm that compares gross count rate or compares count values between specific energy windows is mainly applied instead of nuclide analysis to detect cargo containing radioactive materials.

In airports and ports, radiation portal monitor (RPM) systems are being operated to monitor radioactive materials for imported container cargo. However, the commercialized large-area plastic scintillator-based radiation monitor is difficult to distinguish between artificial and natural radionuclides due to its low energy resolution, so it frequently shows false alarms and hinders the smooth flow of logistics.

In this regard, Korean Patent No. 10-1680067, which is a prior document, relates to a method and apparatus for fractionating radionuclides using a plastic scintillator, using an energy weighting technique that emphasizes the Compton edge region in a peak form using a plastic scintillator. Disclosed is a configuration for discriminating radionuclides, and Korean Patent No. 10-1904003 relates to a plastic scintillator and a method for manufacturing the same. Disclosed is a radionuclide detector based on a plastic scintillator and a radiation detection method using the same, and a configuration for discriminating radiation using a set algorithm for light collected by a multiplier tube using a multi-segment plastic scintillator is disclosed.

The prior art discloses a structure and a technique for discriminating radiation, but there is a need for a method to increase radiation measurement efficiency in consideration of the measurement environment in which the radiation source is moving in the actual field in which the radiation monitoring (RPM) system is operated.

Korean Patent No. 10-1680067 Korean Patent Registration No. 10-1904003 Korean Patent No. 10-2115618

The present invention was created in view of the above circumstances, and an object of the present invention is to enable nuclide discrimination by emphasizing the Compton edge region of the energy spectrum by applying an energy weighting technique, and applying a machine learning technique to move the radiation source. An object of the present invention is to provide a radionuclide fractionation apparatus and method capable of spectral analysis and nuclide fractionation even under conditions with low measurement efficiency, such as the situation.

A radionuclide fractionation device according to an embodiment of the present invention is a plastic scintillator-based radionuclide fractionation device, a radiation input unit that changes energy input from one or more radiation sources into light, and an energy spectrum by collecting light generated from the radiation input unit A spectrum conversion unit for converting into , a peak detection unit for comparing an energy value corresponding to a peak value of an energy-weighted spectrum to which an energy weight is applied to the energy spectrum with an energy value of a Compton edge of the radiation source, the energy-weighted spectrum A machine learning model unit that analyzes the feature of the energy-weighted spectrum and performs learning on the entire shape of the energy-weighted spectrum, and the peak value of the peak detection unit or a nuclide is determined based on the characteristic information analyzed by the machine learning model unit It includes a nuclide fractionation unit.

Here, the energy-weighted spectrum may have an x-axis as an energy value and a y-axis as a product of a photon coefficient and energy.

Here, the energy-weighted spectrum may have an x-axis as an energy value and a y-axis as a product of a square of a photon coefficient and a cube of energy.

Here, the energy-weighted spectrum may have an x-axis as an energy value and a y-axis as a product of a cube of a photon count and a cube of energy.

Here, the energy value corresponding to the peak value may be an x-axis value of the peak value.

Here, the energy value of the Compton edge may be the energy value of the Compton edge corresponding to the gamma energy of the radiation source.

Here, the machine learning model unit, a convolutional neural network (CNN) model for extracting the characteristic data of the energy-weighted spectrum, and linear and non-linear discriminant functions are learned from the characteristic data input from the convolutional neural network, and in the characteristic data It may include a multi-layer perceptron (MLP) model that outputs a discriminant function value for .

Here, the convolutional neural network model may include a convolution layer that filters the image of the energy-weighted spectrum, and a pooling layer that reduces the size of the image.

Here, the multilayer perceptron model includes an input layer to which the characteristic data of the energy-weighted spectrum is input from the convolutional neural network model, a hidden layer for learning linear and non-linear discriminant functions with respect to the characteristic data input to the input layer, and the It may include an output layer that outputs a discriminant function value for the characteristic data transmitted from the hidden layer.

Here, an activation function may be applied for classification of characteristic data calculated in each layer of the convolutional neural network model or the multilayer perceptron model.

Here, a rectified linear unit (ReLU) function may be applied for nonlinear and complex classification of the feature data in the hidden layer.

The radionuclide fractionation method according to an embodiment of the present invention is a plastic scintillator-based radionuclide fractionation method, comprising the steps of (a) changing energy input from one or more radiation sources into light, (b) capturing the light and energy converting into a spectrum, (c) comparing an energy value corresponding to a peak value of an energy-weighted spectrum to which an energy weight is applied to the energy spectrum with an energy value of a Compton edge of the radiation source, (d) analyzing the features of the energy-weighted spectrum, executing learning on the entire shape of the energy-weighted spectrum, and (e) determining the nuclide based on the peak value or characteristic information of the energy-weighted spectrum include

Here, the energy-weighted spectrum may have an x-axis as an energy value and a y-axis as a product of a photon coefficient and energy.

Here, the energy-weighted spectrum may have an x-axis as an energy value and a y-axis as a product of a square of a photon coefficient and a cube of an energy coefficient.

Here, the energy-weighted spectrum may have an x-axis as an energy value and a y-axis as a product of a cube of a photon coefficient and a cube of an energy coefficient.

Here, the energy value corresponding to the peak value may be an x-axis value of the peak value.

Here, the energy value of the Compton edge may be the energy value of the Compton edge corresponding to the gamma energy of the radiation source.

Here, the step (e) may include extracting characteristic data of the energy-weighted spectrum, learning linear and non-linear discriminant functions from the characteristic data, and outputting a discriminant function value for the characteristic data. can

Here, for the statistical reliability check of the learning in step (d), based on the prediction result for the input data of the energy-weighted spectrum, calculating a matrix of accuracy, precision, recall, and calculating an F-score can do.

The radionuclide fractionation apparatus and method according to the present invention have the effect of enabling spectral analysis and nuclide fractionation in a short time even under conditions of low measurement efficiency, such as a situation in which a radiation source moves.

In addition, the radionuclide fractionation apparatus and method according to the present invention improve the precision of radionuclide fractionation by applying an energy weighting technique and a machine learning technique, thereby having the effect of relatively reducing the rate of false alarms in the radiation monitoring system.

1 is a block diagram of a radionuclide fractionation device according to an embodiment of the present invention.
2 is a diagram illustrating a change in peak sharpness of an energy weighted spectrum according to an energy weighting degree.
Fig. 3 shows various deformations according to the exponential change of the photon count and energy.
4 is a flowchart for classifying nuclides by applying an energy weighting algorithm.
5 is an energy spectrum of ¹³⁷ Cs and ⁶⁰ Co in the shielded state.
6 is an energy-weighted spectrum of ¹³⁷ Cs and ⁶⁰ Co in the shielded state.
7 is an energy spectrum measured for ¹³⁷ Cs, ²²⁶ Ra, ²² Na, and ⁶⁰ Co using the RPM system.
8 is an energy-weighted spectrum measured for ¹³⁷ Cs, ²²⁶ Ra, ²² Na, and ⁶⁰ Co using the RPM system.
9 is an energy-weighted spectrum of ¹³⁷ Cs repeatedly measured in a dynamic state at a speed of 1 m/s.
10 shows energy-weighted spectra of ⁶⁰ Co repeatedly measured in a dynamic state at a speed of 1 m/s.
FIG. 11 shows energy-weighted spectra of static state of ¹³⁷ Cs and ²²⁶ Ra measured for 5 sec, 50 sec, and 500 sec, respectively.
12 shows energy spectra of three types of ¹³⁷ Cs and ⁶⁰ Co in static and dynamic states.
13 shows energy spectra of three types: ²²⁶ Ra, ²³² Th, and ⁴⁰ K in static and dynamic states.
14 shows energy spectra of three types of SNM in static and dynamic states.
15 shows the training phase and the testing phase of the CNN model.
16 shows a machine learning technique for nuclide fractionation using an energy-weighted spectrum.
17 is a diagram illustrating a ReLU function.
18 shows the data recombination process
19 shows a positive or negative binary classification matrix.
20 shows prediction accuracy in a training phase according to an increase in the number of epochs.
21 is a flowchart of a radionuclide fractionation method according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to exemplary drawings. In adding reference numerals to the components of each drawing, it should be noted that the same components are given the same reference numerals as much as possible even though they are indicated on different drawings. In addition, in describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description thereof will be omitted.

1 is a block diagram of a radionuclide fractionation device according to an embodiment of the present invention.

Referring to FIG. 1 , the radionuclide fractionation apparatus 100 of the present invention includes a radiation input unit 110 , a spectrum conversion unit 120 , a peak detection unit 130 , a machine learning model unit 140 , and a nuclide classification unit 150 ). includes

The radiation input unit 110 converts energy input from one or more radiation sources into light, and the spectrum converter 120 collects the light generated from the radiation input unit 110 and converts it into an energy spectrum.

The peak detector 130 compares an energy value corresponding to a peak value of an energy weighted spectrum obtained by applying an energy weight to the energy spectrum with an energy value of a Compton edge of the radiation source. Here, the energy value of the Compton edge may be the energy value of the Compton edge corresponding to the gamma energy of the radiation source.

The machine learning model unit 140 analyzes a feature of the energy-weighted spectrum and performs learning on the entire shape of the energy-weighted spectrum.

The nuclide classifying unit 150 determines the nuclide based on the peak value of the peak detector 130 or the characteristic information analyzed by the machine learning model unit 140 . Specifically, the nuclide classification unit 150 compares the energy value corresponding to the peak value of the energy-weighted spectrum extracted by the peak detection unit 130 with the energy value of the Compton edge of the specific radiation source, and the energy value matches Then, it is identified as the corresponding nuclide.

In addition, the nuclide classification unit 150 analyzes the entire shape of the energy-weighted spectrum based on the characteristic information analyzed by the machine learning model unit 140, and if it matches the energy-weighted spectrum of a specific radiation source, it is determined as the corresponding nuclide.

The energy spectrum of the radiation source's gamma rays is measured with a plastic scintillator with a broad primitive Compton continuum instead of a characteristic photoelectric peak, and an energy weighting algorithm is applied.

The energy weighting algorithm transforms the Compton continuum region into a detectable peak using a linear function. Algorithm multiples measure the coefficients in each energy bin of the energy spectrum using the energy of the channel as counts according to the following equation.

Here, C _EW,i is the energy weighting coefficient (count) of the i-th bin of the energy spectrum, C _i is the photon coefficient of the i-th bin, and E _i is the energy coefficient of the i-th bin.

A linear energy coefficient is applied to weight the spectrum of the coefficients measured in each channel, and the energy weighting coefficient increases linearly as the number of channels increases. Compton edges of the energy spectra are transformed into identifiable peaks according to an algorithm.

By applying the algorithm, high coefficients in the low energy region due to scattered radiation and electromagnetic noise are reduced, while the energy weighting coefficients in the high energy region converge to the ground with zero coefficient. As a result, the Compton maximum region is highlighted as an identifiable peak in the energy weighted spectrum.

As the number of radiation sources of interest increases, the number of peaks in the detectable region also increases. This causes overlap between regions and prevents accurate peak detection for radiation source identification. So taking this into account, the energy weighting algorithm derives different properties of the spectrum. Equation 1 is converted to the following equation having constants a and b.

2 is a diagram illustrating a change in peak sharpness of an energy weighted spectrum according to an energy weighting degree.

The indices a and b are various, and the resulting energy-weighted spectrum of ²¹² Bi according to the change of a and b is shown in FIG. 3 . Fig. 3 shows various deformations according to the exponential change of the photon count and energy.

The peak obtained using the C _i Х E _i equation is highlighted as the exponent of the photon count increases, as shown in Fig. 3(A). In addition, as the index of energy increases as shown in FIG. 3B , a significant increase in the energy weighting coefficient is observed in the high energy region. Furthermore, when the photon count and the energy index are simultaneously increased as shown in FIGS. 3C to 3E , the resolution of the peak is improved.

The degree of spectral resolution and the statistical variation are different depending on which exponent value is applied or not. In an embodiment of the present invention, two energy weighting equations are as follows.

The above formulas may be arbitrarily determined according to the number of radiation sources. Therefore, the above formulas can be modified depending on the radionuclide of interest.

4 is a flowchart for classifying nuclides by applying an energy weighting algorithm.

In FIG. 4 , nuclides are discriminated by comparing count ratios of energy spectra.

For the evaluation of the energy weighting algorithm, the energy spectra of six radionuclides were simulated using Monte Carlo simulation. In the energy spectra of all radiation sources, the counts for each of ¹⁹² Ir, ²¹⁴ Pb, ²¹² Bi, ¹³⁷ Cs, ⁶⁰ Co, and ⁴⁰ K are 8.7Х10 ⁵ , 8.3Х10 ⁵ , 1.9Х10 ⁵ , 2.9Х10 ⁵ , 5.4Х10 ⁵ , and 2.5Х10 ⁵ . The PVT density was 1.032 g/cc, and the interaction potential with high-energy gamma rays was lower than that of low-energy gamma rays. Therefore, the spectral fluctuations of high-energy isotopes with Compton maxima located above 1 MeV, such as ⁶⁰ Co, ⁴⁰ K, and ²⁰⁸ Tl, were increased compared to other nuclides. In addition, it was difficult to find the photoelectric peak in the spectrum due to the low density and the Z-number of the PVT material associated with the low photoelectric effect. However, a clear peak distribution within the Compton continuum region was observed by the energy weighting algorithm. The high coefficient decreased due to incomplete absorption or scattering in the low energy region, and the coefficient in the high energy region also decreased significantly compared to the original coefficient. Consequently, the Compton edge was highlighted as an identifiable peak in the energy-weighted spectrum.

In the energy-weighted spectrum, ⁴⁰ K with a theoretical Compton edge of 1.243 MeV was clearly distinguished from ⁶⁰ CO (1.041 MeV), which differed only by 0.202 MeV from the Compton edge. The position difference between the theoretical Compton edge and the peak of the energy-weighted spectrum for ⁶⁰ CO and ⁴⁰ K, respectively, was evaluated as 0.010 and 0.033 MeV as shown in the table below. Table 1 below shows theoretical Compton edges and peak positions of energy-weighted spectra of eight gamma radiation sources.

nuclide ⁵⁷ Co ¹⁹² Ir ²¹⁴ Pb ¹³³ Ba ²¹² Bi ¹³⁷ Cs ⁶⁰ Co ^40K Theoretical Compton Edge (MeV) 0.039 0.175 0.204 0.207 0.439 0.477 1.041 1.243 Peak position (MeV) 0.042 0.174 0.199 0.205 0.431 0.494 1.031 1.210 Difference (MeV) 0.003 0.001 0.005 0.002 0.008 0.017 0.010 0.033

In the case of ²¹⁴ Pb and ²¹² Bi, which belong to the decay chain of natural radioisotopes of ²³⁸ U and ²³² Th, the possibility of distinguishing them from the artificial radioisotope ¹³⁷ Cs was low due to similar gamma energies. However, it is possible to increase the identification accuracy due to a relatively clear coefficient difference of 1 MeV or more.

5 is an energy spectrum of ¹³⁷ Cs and ⁶⁰ Co in a shielded state, and FIG. 6 is an energy weighted spectrum of ¹³⁷ Cs and ⁶⁰ Co in a shielded state.

For shielding, 0, 1 and 2 cm thick steel was used. Although the height of the Compton edge in the shielded state decreased, the coefficient of radiation sources measured in the shielded state in the low energy region was higher than that of the unshielded radiation sources due to more scattering particles from the shielding material.

On the other hand, in the energy-weighted spectrum, the maximum peak corresponding to the Compton edge was clearly improved. For ¹³⁷ Cs and ⁶⁰ Co, the energy range widths were 0.037 MeV and 0.053 MeV, respectively, with some variation.

A laboratory-scale RPM system was constructed for the application of the energy weighting algorithm in various experimental environments, and the peak positions of the energy weighting spectrum were evaluated.

7 is an energy spectrum measured for ^{137 Cs, 226 Ra, 22 Na, 60 Co using the RPM system, and FIG. 8 is an energy spectrum measured for 137 Cs, 226 Ra, 22 Na, 60 Co using the RPM} system. It is an energy weighted spectrum. Here, each of the radiation sources has specific theoretical Compton edges of 0.477, 0.429, 0.341, 1.041 MeV.

Due to the physical properties of the plastic scintillator, the region containing the Compton edge showed a broad distribution in all energy spectra, which may interfere with the spectral analysis during the primary examination in the RPM system. However, in the energy-weighted spectrum, peaks corresponding to the Compton edges of ¹³⁷ Cs and ⁶⁰ Co were clearly observed. In addition, ²² Na has two peaks at 0.340 and 1.068 MeV due to electron dislocation peaks and photoelectric peaks at 0.511 MeV and 1.275 MeV, respectively. Although the overall distribution of the ²²⁶ Ra energy-weighted spectrum in the entire energy range gradually decreased, a peak in the energy-weighted spectrum was observed near 0.439 MeV.

9 shows energy-weighted spectra of ¹³⁷ Cs repeatedly measured in a dynamic state of 1 m/s speed, and FIG. 10 shows energy-weighted spectra of ⁶⁰ Co repeatedly measured in a dynamic state of 1 m/s speed.

9 and 10, the energy-weighted spectra of ¹³⁷ Cs and ⁶⁰ Co measured in the dynamic state at a speed of 1 m/s show relatively higher statistical fluctuations than in the static state. Therefore, it has been demonstrated that, in the condition of a moving radiation source, a designated energy range for peak detection is required for nuclide fractionation using an energy-weighted spectrum.

FIG. 11 shows energy-weighted spectra of static state of ¹³⁷ Cs and ²²⁶ Ra measured for 5 sec, 50 sec, and 500 sec, respectively.

11, ¹³⁷ Cs and ²²⁶ Ra have similar gamma energies of 0.662 and 0.609 MeV, respectively, and Compton edges were observed in the 0.312-0.591 MeV region for both isotopes. Gamma rays from the daughter isotopes of ²²⁶ Ra were observed as clear peaks near 0.3 MeV, and the high coefficient of ²²⁶ Ra above 0.7 MeV could be distinguished from the coefficient of ¹³⁷ Cs approaching zero in the same energy region. have. Statistical uncertainty decreases with increasing measurement time.

12 shows energy spectra of three types of ¹³⁷ Cs and ⁶⁰ Co in static and dynamic states. The black lines are spectra measured for 300 seconds in a static state, and the colored lines are spectra measured in a dynamic state of 5 km/h and 10 km/h. In the colored lines, peak positions were detected within the designated peak detection range indicated by the dotted lines. As the speed of the radiation source increased from 5 km/h to 10 km/h, the spectrum fluctuations increased.

13 shows energy spectra of three types: ²²⁶ Ra, ²³² Th, and ⁴⁰ K in static and dynamic states. ²²⁶ Ra is one of the natural radionuclides, and its daughter material emits gamma rays at various energy levels.

In the C _i Х E _i energy-weighted spectrum, the measured coefficient increased to about 0.4 MeV and then decreased. As the exponential constant increased, the coefficient increased in the range of 0.8-1.2 MeV, and in the C _i ³ Х E _i ⁶ energy-weighted spectrum, the pit detection area was changed. In all energy-weighted spectra, the peak detection area was assigned a relatively wider energy range compared to other radiation sources due to high fluctuations that could prevent accurate peak detection.

²³² Th is also one of the natural radionuclides, and exhibits a broad coefficient distribution up to 0.8 MeV in the C _i Х E _i energy-weighted spectrum. However, in the C _i ³ Х E _i ⁶ energy-weighted spectrum, a high coefficient distribution is observed due to ²⁰⁸ T1, one of the daughter materials emitting gamma rays at 2.61 MeV. This is the main characteristic that distinguishes ²²⁶ Ra from other radiation sources.

In an RPM system, ⁴⁰ K with a Compton edge at 1.243 MeV would be perceived as ⁶⁰ Co with a theoretical Compton edge at 1.041 MeV, which could cause a malfunction alarm. Even if the energy weighting algorithm is applied, the peak detection area is similar to ⁶⁰ Co, so peaks of ⁴⁰ K may be misidentified.

Therefore, after simultaneous peak detection in the three energy-weighted spectra, an additional algorithm step is applied to compare the coefficient ratios of 0.9573-1.3781 MeV and 0.9573-1.1677 MeV in the C _i ³ Х E _i ⁶ energy-weighted spectrum.

14 shows energy spectra of three types of special nuclear material (SNM) in a static state and a dynamic state. The energy spectrum of DU with a ²³⁵ U concentration of less than 0.72% shows a relatively high modulus below 0.2 MeV. In the three types of energy-weighted spectra, DU has a distribution similar to ²³² Th except for the energy range of 2MeV or more in the C _i ³ Х E _i ⁶ energy-weighted spectrum.

For HEU and WGPu, significant peaks were observed below 0.4 MeV in _{Ci Х E i and Ci 2 Х E i 3} ^energy _weighted ^spectra , and _Ci ³ Х E _i ⁶ energies under migration conditions for three _SNMs . In the weighted spectrum, additional fluctuations were found in HEU outside the expected peak detection region.

Table 2 below shows the radiation source identification results in a dynamic state.

nuclide Identification rate (number of successes/number of attempts) LMS IMCC 5 km/h 10 km/h 5 km/h 10 km/h ¹³⁷ Cs 7.85 µCi 91.7% (11/12) 50% (6/12) - - 17.4 μCi 100% (12/12) 100% (12/12) 66.7% (8/12) 25% (3/12) 58.3 μCi - - 91.7% (11/12) 58.3% (7/12) ⁶⁰ Co 3.6 μCi 100% (12/12) 83.3% (10/12) - - 6.2 μCi 100% (12/12) 100% (12/12) 50% (6/12) 16.7% (2/12) 27.9 μCi - - 91.7% (11/12) 58.3% (7/12) ²²⁶ Ra 8.1 μCi 80% (8/10) 70% (7/10) 33.3% (4/12) 16.7% (2/12) ²³² Th 13.6 μCi 90% (9/10) 90% (9/10) 66.7% (8/12) 25% (3/12) ^40K 4 tons (wgt.) - - 58.3% (7/12) 50% (6/12) DU 0.38 p/cm ² /s 80% (8/10) 80% (8/10) - - HEU 4.37 p/cm ² /s 100% (10/10) 100% (10/10) - - WGPu 7.53 p/cm ² /s 100% (10/10) 100% (10/10) - -

The ¹³⁷ Cs and ⁶⁰ Co radiation sources were 100% identified at 5 km/h and 10 km/h speeds by LMS (linear motion system). ²³² Th with high spectral fluctuation was 90% distinct at the speed of 5 km/h, and was successfully confirmed compared to ²²⁶ Ra. SNMs that should trigger an RPM alarm were identified over 80% at speeds of 5 km/h and 10 km/h, and in some cases were misidentified as natural sources of radiation. At a speed of 10 km/h, the energy-weighted spectra of ²³² Th, ²²⁶ Ra and DU were evaluated as similar spectra and misidentified each other as the relative geometrical efficiency decreased.

For tests conducted with an intermodal cargo container (IMCC), they were misidentified in most cases, and ¹³⁷ Cs at 58.3 μCi and ⁶⁰ Co at 27.9 μCi were successfully discriminated not only at 5 km/h but also at 10 km/h.

A relatively high variation was observed in the energy-weighted spectrum of the misidentified radiation source, which was due to the location of the radiation source in the container. The radiation source was placed on the bottom of the truck and the height from the ground was equal to the center of the PVT panel. Radiation scattering was analyzed to increase in the truck structure.

However, the ⁴⁰ K radiation source was found to be 60 % at both 5 km/h and 10 km/h speeds. Unlike other radiation sources, Compton edges of ⁶⁰ Co and ⁴⁰ K were observed at 1-1.3 MeV in the energy spectrum, and the two nuclides can be distinguished by comparing the count ratio between the designated peak detection regions of the energy-weighted spectrum. . Therefore, in the primary radiological testing site, the rate of false alarms due to ⁴⁰ K radiation sources can be greatly reduced by applying the energy weighting technique.

The classification of nuclides by the peak of the energy-weighted spectrum corresponding to the Compton edge showed a relatively high identification rate in static and dynamic radiation source conditions. However, it is still difficult to discriminate between nuclides with similar Compton edges and very close peaks in energy spectra, such as ⁶⁰ Co and ⁴⁰ K. Therefore, it is necessary to analyze not only the peak positions but also the features of the entire spectrum in order to increase the accuracy of nuclide classification.

The MultiLayer Perceptron (MLP) model is a machine learning technique that can determine the alarm threshold for the simulated energy spectrum of multiple channels and analyze the spectrum distribution. Machine learning provides a computer system with the ability to learn based on statistical methods, and it means that the machine can achieve its own goals without additional programming.

The multilayer perceptron model is used to identify characteristics by biasing and weighting coefficients of all channels in the energy spectrum as input data. The mean square error can be evaluated according to the overall gross coefficient or the gross coefficient of a specific channel bin range.

However, since the multilayer perceptron model is a feed-forward layer artificial neural network, the quality of the features has a very large effect on the output, and back propagation is performed for error correction. As described above, it is very difficult to discriminate between nuclides having a similar Compton edge and very close peaks in an energy spectrum, so an iterative process of input data and training can be used to increase the accuracy of nuclide discrimination.

A convolutional neural network (CNN) is a modified form of a multi-layer perceptron (MLP) model, which effectively extracts features by repeating the convolution step to filter the image and the pooling step to reduce the image size. . The convolutional neural network technique is a structure in which the multi-layer perceptron method is superimposed into multiple layers and multiple layers, and is a technique capable of classifying high-dimensional data. In the same way as in the multi-layer perceptron method, the optimal weight and bias values are derived by repeating mathematical operations during the learning process.

The energy weighting algorithm amplifies a count difference in a high energy region, and a process of comparing the ratio of energy weighting coefficients in a specific energy region as well as a peak position is required. By applying the CNN model, more accurate nuclide classification is possible, and the coefficient difference can be extracted as a characteristic. In the following, the CNN model is applied to the energy-weighted spectra of static and dynamic radiation sources measured with multiple plastic scintillators, and the resulting output is statistically evaluated for nuclide classification.

15 shows the training phase and the testing phase of the CNN model.

The training stage is a stage of training by applying multiple data to the CNN model. When the user artificially labels the characteristics and labels of the data, it is called supervised learning, and when the CNN model recognizes the characteristics by itself, it is called unsupervised learning. If clear labeling is possible, supervised learning is a cheaper and faster learning process than unsupervised learning. However, supervised learning may be inefficient if a complex process is required for the selection of features. In unsupervised learning, the characteristics increase as the training process is repeatedly performed, and the success rate of nuclide discrimination can also be improved.

Since the fluctuation of the energy-weighted spectrum measured for a short period is relatively high, it is difficult to detect an accurate peak. In this case, supervised learning of the CNN model can be a good option to classify nuclides with similar Compton edges, such as ⁶⁰ Co and ⁴⁰ K.

In the training phase, the CNN model outputs predictions and error rates between existing data and new input data based on the characteristics extracted after training. If the error is large, the CNN model runs a feedback process to adjust the weights, and repeats until the error is minimized. After validation, the trained model is tested using unknown data in the testing phase, and the model's performance is evaluated by the prediction's correction rate.

16 shows a machine learning technique for nuclide fractionation using an energy-weighted spectrum.

The CNN model filters the image in the convolutional stage and effectively reduces the computation time by reducing the number of pixels in the filtered image in the pooling stage. After that, the features extracted through the convolution and pooling steps of the CNN model are input to the MLP model, and the prediction is output. The convolution layer of the CNN model and the window size of the filter are set by the user, and higher-level features can be extracted as the distance from the input layer increases.

The CNN model calculates a feature map by striding the filtered window at regular intervals, and passes it to the next layer. In one embodiment of the present invention, 4 convolutional layers of a 3x3 window filter with 2-pixel stride are used, and an activation function is applied for classification in each layer.

A complex type of classifier can be implemented by stacking several perceptrons, and this is called a multi-layer perceptron (MLP) or neural network (NN). The multilayer perceptron consists of three layers: an input layer, a hidden layer, and an output layer. In this case, the perceptrons constituting each layer are represented by circles in the neural network schematic diagram and may be represented by one node. The characteristic of the data to be learned is input to the input layer, and it has an input perceptron in proportion to the type of the feature.

Every layer between the input layer and the output layer is defined as a hidden layer, and the input layer and the output layer are not directly connected, but run through the hidden layer. Through this, in the hidden layer, a number of nonlinear discriminant functions are learned for the data input from the input layer, and finally the output layer outputs discriminant function values corresponding to each group with respect to the data. In this way, the input data is transferred in the order of the input layer, the hidden layer, and the output layer, and is changed into a discriminant function value.

An artificial neural network (ANN) model including MLP and CNN uses an activation function for classification of weighted data computed in each layer. The activation function may be, for example, a rectified linear unit (ReLU) function or a softmax function. The ReLU function is used in the hidden layer, allowing machine learning models to perform nonlinear and complex classification.

17 is a diagram illustrating a ReLU function. The ReLU function is a function having a threshold that outputs only a signal having an intensity greater than or equal to 0 with respect to the input signal. The ReLU function can solve the problem of gradient loss and increase the efficiency of computation time.

The softmax function is used in the output layer and is a normalized exponential function as shown in Equation 9 below. The softmax function estimates each probability of each class according to the input of a k-dimensional vector. That is, the softmax function normalizes the output of the input data to a value between 0 and 1, and the sum of the output values is always 1. It consists of outputs as many as the number of classes to be classified, and the class to which the largest output value is assigned is used as the one with the highest probability.

where p _i is a random vector between (0,1) and z _i is the i-th element in the k-dimensional vector. In the softmax function, the sum of p _i is 1, and classification of the input data is possible compared to the one-hot encoding of the correct class.

In the pooling layer, the size of the convolutional image is reduced. This process makes the trained features less sensitive to changes in the position of the input images. Unlike the convolutional layer, there is no parameter to train, and there is no change in the number of channels for the input data. In general, the pooling method is classified into three types: maximum pooling, average pooling, and minimum pooling, and in one embodiment of the present invention, maximum pooling is used.

Energy-weighted images of ¹³⁷ Cs, ⁶⁰ Co, ²²⁶ Ra, and ⁴⁰ K measured in the static and dynamic states described above may be used as input data.

18 shows the data recombination process. For high accuracy of isotope identification, spectral recombination is performed to obtain abundant spectral data by a combination of arbitrarily selected spectra as shown in Fig. 18, and 1,000 spectral data are obtained from each radiation source. . Among them, 700 images were used as input data in the training stage of the CNN model, and the remaining 300 images were used as input data in the test stage.

Since statistical reliability is very important for machine learning evaluation, data evaluation using some statistical parameters was performed.

Based on the prediction results of the CNN model's input data, a matrix of accuracy, precision, and recall was calculated, and their harmonic average F-score was also calculated.

19 shows a positive or negative binary classification matrix. This matrix gives four results. That is, correct positives for prediction (true positive, TP), incorrect positive predictions (false positive, FP), correct negative predictions (true negative, TN) and incorrect negative predictions (false negative, FN).

Based on these parameters, accuracy, precision, and recall can be derived by the following equations.

Accuracy is the correct prediction divided by the total number of input data sets. This parameter is close to a specific value and intuitively represents the predictive performance of a machine learning model. However, accuracy must take into account the biasing of the data, which is why parameters such as precision and recall are necessary.

Precision refers to the ratio of correct positive predictions to total positive predictions. This parameter shows how precise the model is. In the present invention, precision means the ratio of the number of predicted nuclides to the number of data in which nuclides actually exist.

Recalls are calculated as the number of actual positive predictions divided by the total number of positives and negatives in the actual class. In the present invention, recall means the ratio of the number of specific nuclide data and the number of data predicted by the algorithm to exist among them.

When evaluating a model, accuracy can be a simple and general evaluation indicator, but it has the disadvantage that it is difficult to evaluate quantitatively if the data are biased in either direction. Also, if precision or recall is rated abnormally high, it is not statistically reliable. On the other hand, the F-score, which is a harmonic average of precision and recall, has the advantage of minimizing data bias even if the data set is distorted.

The F-score is expressed by the following equation.

Here, β is a weight that emphasizes the importance of precision for recall. In one embodiment of the present invention, it may be set to 1 so that there is no bias for precision or recall. The closer the F-score is to 1, the closer the machine learning model is to perfection. In the experimental example of the present invention, statistical parameters for TP, FN, FP, and TN were calculated for ¹³⁷ Cs, ⁶⁰ Co, ²²⁶ Ra and ⁴⁰ K at 10, 20, and 30 km/h.

In the training phase of machine learning, prediction accuracy increases as the number of training is repeated, and the entire process including data input to training feedback is called an epoch. Accordingly, the optimal number of training is evaluated according to the increase in the number of epochs, and FIG. 20 shows the prediction accuracy in the training phase according to the increase in the number of epochs. When the number of epochs was 5, the prediction accuracy was about 95%, and the prediction accuracy of more than 99% was possible with 8 or more training.

Therefore, the number of epochs was set to 8, and the test was run using the test spectral data set of ¹³⁷ Cs, ⁶⁰ Co, ²²⁶ Ra, and ⁴⁰ K. The table below shows the test results under the radiation source movement conditions of 10 to 30 km/h.

radiation source speed farsighted
radiation source Number of predicted radiation sources ¹³⁷ Cs ⁶⁰ Co ²²⁶ Ra ^40K 10 km/h ¹³⁷ Cs 100 0 0 0 ⁶⁰ Co 0 100 0 0 ²²⁶ Ra 13 0 87 0 ^40K 0 17 0 83 20 km/h ¹³⁷ Cs 100 0 0 0 ⁶⁰ Co 0 93 0 7 ²²⁶ Ra 17 0 83 0 ^40K 0 27 0 73 30 km/h ¹³⁷ Cs 100 0 0 0 ⁶⁰ Co 0 100 0 0 ²²⁶ Ra 27 0 73 0 ^40K 0 27 0 73

¹³⁷ Cs showed a discernable peak in the actual energy-weighted spectrum above, and was well predicted under all conditions as shown in the table above. ⁶⁰ Co was also identified more than 90% in all speed conditions. However, in the case of ²²⁶ Ra, confusion with ¹³⁷ Cs occurred, and the degree of confusion increased as the speed of the radiation source increased.

In particular, ²²⁶ Ra emits gamma rays at multiple energies, but the daughter ²¹⁴ Bi emits gamma rays at 0.609 MeV with an emission rate of 46%, so there is a theoretical Compton edge at 0.429 MeV. Therefore, ²²⁶ Ra is expected to be recognized as a nuclide similar to ¹³⁷ Cs with a Compton edge of 0.477 MeV.

Nevertheless, at least 73% of ²²⁶ Ra were identified because they show a spectrum different from ¹³⁷ Cs due to the gamma coefficient occurring in daughter nuclides.

⁴⁰ K, one of the natural radionuclides, also showed a similar identification rate to ²²⁶ Ra. Unlike other natural radioactive sources, gamma rays of ⁴⁰ K emit only 1.460 MeV with an intensity of 11%, with a theoretical Compton edge of 1.243 MeV. This is close to 1.041 MeV of ⁶⁰ Co, and the shape of the spectrum is also similar, so it is evaluated that it is very difficult to distinguish between ⁴⁰ K and ⁶⁰ Co nuclides.

Table 4 shows the precision, recall, and F-score of each radiation source calculated to evaluate the reliability of the results obtained by applying the CNN model. The precision and recall of ¹³⁷ Cs, ⁶⁰ Co, and ²²⁶ Ra were evaluated as high values of 0.8 or higher, and ⁴⁰ K showed a relatively low recall value of 0.76 at a moving radiation source of 30 km/h. For all nuclides, the F-score was close to 1.

radiation source ¹³⁷ Cs ⁶⁰ Co ²²⁶ Ra ^40K Precision (p) 0.84 0.8 One 0.94 recall (r) One 0.95 0.81 0.76 F-score 0.91 0.87 0.89 0.84

21 is a flowchart of a radionuclide fractionation method according to an embodiment of the present invention.

The radiation input unit 110 converts energy input from one or more radiation sources into light (S210), and the spectrum converter 120 collects the light generated from the radiation input unit 110 and converts it into an energy spectrum (S220).

Next, the peak detector 130 compares an energy value corresponding to a peak value of an energy weighted spectrum to which an energy weight is applied to the energy spectrum with an energy value of a Compton edge of the radiation source (S230). Here, the energy value of the Compton edge may be the energy value of the Compton edge corresponding to the gamma energy of the radiation source.

Subsequently, the machine learning model unit 140 analyzes a feature of the energy-weighted spectrum and executes learning on the entire shape of the energy-weighted spectrum ( S240 ).

Next, the nuclide classifying unit 150 discriminates the nuclide based on the peak value or characteristic information of the energy-weighted spectrum ( S250 ). Specifically, the nuclide classification unit 150 compares the energy value corresponding to the peak value of the energy-weighted spectrum extracted by the peak detection unit 130 with the energy value of the Compton edge of the specific radiation source, and the energy value is the same Then, it is identified as the corresponding nuclide.

In addition, the nuclide classification unit 150 analyzes the entire shape of the energy-weighted spectrum based on the characteristic information analyzed by the machine learning model unit 140, and if it matches the energy-weighted spectrum of a specific radiation source, it is determined as the corresponding nuclide. To this end, the step S250 may include extracting characteristic data of the energy-weighted spectrum, learning linear and non-linear discriminant functions from the characteristic data, and outputting a discriminant function value for the characteristic data. have.

In an embodiment of the present invention, based on the prediction result for the input data of the energy weighted spectrum for the statistical reliability check of the learning of the step S240, the matrix calculation of accuracy, precision, recall and F-score are calculated. It may include further steps.

Although the present invention has been described in detail with reference to preferred embodiments so far, the present invention is not limited to the above-described embodiments, and without departing from the gist of the present invention as claimed in the following claims, the technical field to which the present invention pertains It will be said that the technical idea of the present invention extends to a range where various modifications or modifications can be made by anyone having ordinary knowledge in the present invention.

100: radionuclide fractionation device
110: radiation input unit
120: spectrum conversion unit
130: peak detection unit
140: machine learning model unit
150: nuclide fractionation unit

Claims (19)

A plastic scintillator-based radionuclide fractionation device, comprising:
a radiation input unit for converting energy input from one or more radiation sources into light;
a spectrum conversion unit that collects the light generated from the radiation input unit and converts it into an energy spectrum;
a peak detector for comparing an energy value corresponding to a peak value of an energy weighted spectrum obtained by applying an energy weight to the energy spectrum with an energy value of a Compton edge of the radiation source;
a machine learning model unit that analyzes a feature of the energy-weighted spectrum and executes learning on the entire shape of the energy-weighted spectrum; and
A radionuclide fractionation device comprising a nuclide fractionator for discriminating a nuclide based on the peak value of the peak detector or the characteristic information analyzed by the machine learning model part. The radionuclide fractionation apparatus according to claim 1, wherein the energy-weighted spectrum has an x-axis as an energy value and a y-axis as a product of a photon count and energy. The radionuclide fractionation apparatus according to claim 1, wherein the energy-weighted spectrum has an x-axis as an energy value and a y-axis as a product of a square of a photon count and a cube of energy. The radionuclide fractionation apparatus according to claim 1, wherein the energy-weighted spectrum has an x-axis as an energy value and a y-axis as a product of a cube of a photon count and a cube of an energy. The radionuclide fractionation device according to any one of claims 2 to 4, wherein the energy value corresponding to the peak value is an x-axis value of the peak value. The radionuclide fractionation apparatus according to claim 1, wherein the energy value of the Compton edge is an energy value of the Compton edge corresponding to the gamma energy of the radiation source. According to claim 1, wherein the machine learning model unit,
a convolutional neural network (CNN) model for extracting characteristic data of the energy-weighted spectrum; and
Linear and non-linear discriminant functions are learned from the characteristic data input from the convolutional neural network, and a multilayer perceptron (MLP) model for outputting a discriminant function value for the characteristic data. The method of claim 7, wherein the convolutional neural network model,
Radionuclide fractionation device, characterized in that it comprises a convolution layer (convolution layer) for filtering the image of the energy-weighted spectrum, and a pooling layer (pooling layer) for reducing the size of the image. According to claim 7, wherein the multi-layer perceptron model,
an input layer to which characteristic data of the energy-weighted spectrum is input from the convolutional neural network model;
a hidden layer for learning linear and non-linear discriminant functions with respect to the characteristic data input to the input layer; and
Radionuclide fractionation device comprising an output layer for outputting a discriminant function value for the characteristic data transmitted from the hidden layer. The radionuclide fractionation device according to claim 8 or 9, wherein an activation function is applied for classification of characteristic data calculated in each layer of the convolutional neural network model or the multi-layer perceptron model. The radionuclide fractionation apparatus according to claim 9, wherein a rectified linear unit (ReLU) function is applied for nonlinear and complex classification of the characteristic data in the hidden layer. As a plastic scintillator-based radionuclide fractionation method,
(a) converting energy input from one or more radiation sources into light;
(b) capturing the light and converting it into an energy spectrum;
(c) comparing an energy value corresponding to a peak value of an energy weighted spectrum obtained by applying an energy weight to the energy spectrum with an energy value of a Compton edge of the radiation source;
(d) analyzing a feature of the energy-weighted spectrum and executing learning on the entire shape of the energy-weighted spectrum; and
(E) Radionuclide fractionation method comprising the step of discriminating a nuclide based on the peak value or characteristic information of the energy-weighted spectrum. The method of claim 12 , wherein the energy-weighted spectrum has an x-axis as an energy value and a y-axis as a product of a photon count and energy. The method of claim 12, wherein the energy-weighted spectrum has an x-axis as an energy value and a y-axis as a product of a square of a photon count and a cube of energy. 13. The method of claim 12, wherein the energy-weighted spectrum has an x-axis as an energy value and a y-axis as a product of a cube of a photon count and a cube of an energy. The method according to any one of claims 13 to 15, wherein the energy value corresponding to the peak value is an x-axis value of the peak value. The method according to claim 12, wherein the energy value of the Compton edge is an energy value of the Compton edge corresponding to the gamma energy of the radiation source. The method of claim 12, wherein step (e) comprises:
extracting characteristic data of the energy-weighted spectrum; and
Learning linear and non-linear discriminant functions from the characteristic data, and outputting a discriminant function value for the characteristic data. 13. The method of claim 12, wherein for the statistical reliability check of the learning in step (d), based on the prediction result for the input data of the energy-weighted spectrum, the matrix calculation of accuracy, precision, recall, and F-score are calculated. Radionuclide fractionation method, characterized in that it further comprises the step.

Images