JP7309149B2 - Radiation energy determination method and apparatus using machine learning - Google Patents

Description

The present invention relates to a radiation energy discrimination method and apparatus using machine learning. The present invention relates to a radiation energy determination method and apparatus using machine learning suitable for use in estimating energy.

As a method of measuring the energy of radiation such as protons, neutrons, and gamma rays, there is a method of deriving the energy from the amount of light emitted when the radiation is incident on a scintillator.

Since the mechanism for imparting energy to the scintillator differs depending on the type of radiation, the amount of light emitted from the scintillator differs, but the amount of light emitted is substantially proportional to the energy. That is, as the energy of the radiation increases, the amount of light emitted inside the scintillator increases.

The correlation between the amount of emitted light and the radiation energy can be obtained by introducing radiation of known energy and systematically measuring the relationship between the amount of emitted light and the energy in the scintillator.

However, for example, when high-energy radiation is incident on the scintillator, it passes through the scintillator as radiation whose energy has been attenuated before all of its energy is imparted to the scintillator. In this case, only part of that energy is transferred to the scintillator, and the rest of the energy carried by the transmitted radiation is unknown. Therefore, when the radiation has high energy enough to pass through the scintillator, it becomes difficult to accurately identify the energy.

Therefore, two methods are conceivable for measuring the energy of radiation with the radiation stationary in the scintillator. One is a method of adopting a scintillator large enough to impart sufficient energy, and the other is, as described in Non-Patent Document 1, an energy-absorbing material (absorber) between the radiation source and the detector. It is a method of installing and attenuating energy.

However, the former requires a larger scintillator and a higher manufacturing cost for the detector. On the other hand, the latter does not increase the size of the scintillator itself contained in the detector, but because the absorber is installed in front of the scintillator, it is possible to measure the energy of the radiation that passes through the absorber and is stationary inside the scintillator. , the stationary position in the absorber is unknown, so it can only be identified that the energy is smaller than the energy that can penetrate the absorber. For example, cosmic ray muons have different energies. It is not suitable for accurately measuring the continuous energy distribution of incoming radiation.

Therefore, in Patent Document 1, the applicant proposed a scintillator (for example, a plastic Using a laminated radiation detector (hereinafter simply referred to as a laminated detector or detector) 10 in which a scintillator (hereinafter also referred to as plasticine) 12 and an absorber (for example, iron or lead) 14 are laminated, the incident radiation is detected. A technique of measuring the energy by causing the scintillator 12 to emit light after part of the energy is absorbed by the absorber 14 is proposed. According to this stacked detector 10, by increasing the number of stacked layers, higher energy can be derived, and by sandwiching the absorber 14, the volume of the scintillator 12 can be reduced, resulting in lower cost detection. Crafting becomes possible.

Japanese Patent Application 2018-226493 (unpublished at the time of filing) Japanese Patent Application 2018-36875 (unpublished at the time of filing)

Glenn F. Knoll "Radiation Measurement Handbook 4th Edition" Ohmsha (September 25, 2013)

According to the stacked detector 10 proposed in Patent Document 1, the energy of radiation stationary within the absorber 14 can be estimated within a certain energy range. That is, when identifying the energy of radiation, as illustrated in FIG. 2, when a is stationary within the scintillator 12, its energy b can be uniquely determined. However, when the radiation is stationary within the absorber 14, the width between c1 immediately after entering the absorber 14 and _c2 immediately before penetrating the absorber ₁₄ is greater than or equal to a certain value d1 and _less than or equal to a certain value _d2 . There is a problem that we only know that it had energy d of

This may be insufficient for detecting smaller defects when used in the "Non-Destructive Inspection Method and Apparatus for Structures Using Muography" proposed by the applicant in Patent Document 2.

Furthermore, when deriving the kinetic energy of a muon, the scintillator overlapping the scintillator where the muon is stationary and the energy imparted to the absorber are obtained from the Bethe equation. There is also the problem that the original energy is not uniquely determined from the information on the energy imparted to the scintillator and varies stochastically.

The present invention has been made to solve the above-mentioned conventional problems, and aims to improve the performance of infrastructure inspection by making it possible to estimate the energy of radiation stationary in an absorber and radiation penetrating a detector. and

In the present invention, when discriminating the energy of radiation detected by a detector in which a plurality of scintillators and absorbers are stacked, the energy imparted to each layer of scintillators is used as an input layer, and the energy of incident radiation is used as an output layer. Using a network, the relationship between the energy imparted to the scintillator of each layer and the energy of the incident radiation is machine-learned, and using the results of the machine learning, the energy of the incident radiation is obtained from the energy imparted to the scintillator of each layer. In the radiation energy discriminating method for estimating the incident radiation event, for a radiation event stationary in the absorber, information on the amount of light emitted up to the scintillator in the preceding stage of the stationary absorber is analyzed by machine learning to determine the incident radiation The problem is solved by estimating the energy and estimating the energy of the incident radiation with respect to the radiation event penetrating the detector by analyzing information on the amount of light emitted from all scintillators by machine learning. .

The present invention also includes a detector in which a plurality of scintillators and an absorber are laminated, a neural network whose input layer is the energy given to the scintillators of each layer, and whose output layer is the energy of the incident radiation, and scintillators in each layer. means for performing machine learning on the relationship between the energy imparted to the scintillator and the energy of the incident radiation, and using the results of the machine learning to estimate the energy of the incident radiation from the energy imparted to the scintillators of each layer. In the radiation energy discriminating device, for a radiation event stationary in the absorber, information on the amount of light emitted up to the scintillator at the front stage of the stationary absorber is analyzed by machine learning to determine the energy of the incident radiation. and for radiation events penetrating the detector, information on the amount of light emitted from all scintillators is analyzed by machine learning to estimate the energy of incident radiation. An energy discrimination device is provided.

Here , the scintillator can be a plastic scintillator, and the absorber can be iron or lead.

According to the present invention, by data analysis using a neural network, for a radiation event stationary in the absorber, by analyzing information on the amount of light emitted up to the front stage of the stationary absorber by machine learning, the optimum incidence It becomes possible to derive the energy value. Similarly, for a high-energy radiation event that penetrates the stacked detector, it is possible to derive the optimum incident energy value from the information on the energy imparted to all the scintillators of the stacked detector.

For example, from Patent Document 1, it is known that in infrastructure inspection using muography, by selecting the muon energy according to the size of the defect to be detected, the contrast of the imaging image can be improved and the identification of the defect can be improved. If it becomes possible to select the muon energy in the absorber more finely, the contrast of the imaging image can be improved, and the performance of the infrastructure inspection can be improved.

A diagram schematically showing the basic configuration of a laminated detector proposed by the applicant in Patent Document 1. A diagram showing the problem of the detector with the configuration of FIG. Diagram showing the Bragg curve of muons for water Diagram showing the principle of the present invention A diagram showing an image of muon incidence for explaining the simulation system in the embodiment of the present invention. The figure which shows the structure of the neural network used by the said embodiment. A diagram showing an example of the relationship between the number of iterations and the loss function in the embodiment. Similarly, a diagram showing the learning status of the model used to compare regions with good energy discrimination and regions with poor energy discrimination. Figure showing comparison of estimated accuracy for each incident energy Figure showing an example of limiting the energy before muon injection A diagram showing a comparison of images obtained by similarly narrowing down the energy

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In addition, the present invention is not limited by the contents described in the following embodiments and examples. In addition, the configuration requirements in the embodiments and examples described below include those that can be easily assumed by those skilled in the art, those that are substantially the same, and those that fall within the so-called equivalent range. Furthermore, the constituent elements disclosed in the embodiments and examples described below may be combined as appropriate, or may be selected and used as appropriate.

In the present invention, we propose a method for deriving the accurate energy of radiation stationary in an absorber and radiation penetrating a detector using machine learning. In the laminated detector of an absorber and a scintillator as proposed in Patent Document 1, the fact that the amount of light emitted in each scintillator when radiation passes through differs depending on the energy of the radiation is utilized.

In the muon Bragg curve for water shown in FIG. 3, the horizontal axis is the penetration depth of the radiation, and the vertical axis is the amount of light emission (energy). I know it will be. If a muon of 150 MeV enters the stacked detector, is decelerated to 100 MeV by the first layer absorber, and decelerated to 50 MeV by the second layer absorber, the amount of light emitted by each scintillator sandwiched between should be values such as a, b, and c in FIG.

In this way, depending on the energy of the radiation, the amount of light emitted in the scintillator when it is stationary in the absorber or when it penetrates the detector is created as a data set by simulation calculation, and this is learned by the neural network.

As a result, even when stationary in the absorber 14, by analyzing the information on the amount of light emitted from the radiation in the previous stage and the stage before the previous stage by machine learning, the radiation c stationary in the absorber 14 as shown in FIG. It becomes possible to estimate the energy in a narrow range d'.

In machine learning, it is necessary to create a dataset for learning the model. Therefore, the data set was first created by simulation. FIG. 5 shows an image diagram of muon incidence for explaining the simulation system. This figure shows a system in which three plastic scintillators 16 with a thickness of 1 cm and two iron plates 18 with a thickness of 1 cm are laminated. This simulation takes into account the effects of Coulomb scattering and energy straggling.

Next, by analyzing _the simulation data, a data set of the incident energy E _μ of each muon and the imparted energy {E _d1 , _Ed2 , .

In this embodiment, a fully-connected neural network 20 having the configuration shown in FIG. 6 is used as the neural network for learning the created data set. The input layer 22 is the value of energy applied to each plastic scintillator, and the number of inputs is the number of plastic scintillators 16 .

Although it is part of the output layer 30, the output becomes the muon kinetic energy E _μ . This time, the output part is divided into multiple layers, and the output is in the form of 0 to 1, 1 to 2, .

FIG. 7 shows how hidden layers (also called intermediate layers) 24 and 26 change with respect to the number of iterations (epoch number) of the loss function in the pattern of the number of units. In Case 2, where the number of hidden layers is small, the loss function increases due to over-learning in areas with a large number of iterations. The loss function becomes smaller as the number of units in the hidden layer is increased, as in cases 3 and 4 indicated by ▪ and ◇. do not have. In this embodiment, a two-layer structure having 1500 units and 500 units of Case 1 indicated by ● marks is used.

Note that when the number of input layers increases (when the number of plastic scintillators increases), the optimal number of hidden layers and the number of units may change.

Of the datasets created in the simulation, one million sets of learning data were used for learning, and the expected output results of the model with 100,000 sets of verification data were obtained. The number of iterations in learning was set to 50.

Next, we examined the accuracy for each incident energy. As shown in FIG. 8, a muon with an energy E _μ = 0 to 300 MeV is incident on a laminated detector 10 composed of three plastic scintillators 16 each having a thickness of 1 cm and two iron plates 18, and the imparted energy information { E _d1 , E _d2 , E _d3 } were derived by simulation, and learning data and verification data were created. Then, using the verification data, the prediction accuracy of the incident energy was obtained by dividing it by 10 MeV such as E _μ =0 to 10, 10 to 20, . . . , 290 to 300 MeV.

The prediction results of the model for 40-50 and 80-90 MeV muon kinetic energies are shown in FIG. As shown in FIG. 9(A), the original energy could be accurately estimated at energies of 40 to 50 MeV. Also, when the incident energy is large enough to penetrate the detector, the output energy has a spread as shown in FIG. Therefore, the energy window type muography described in Patent Literature 1 excludes events due to high-energy muons and extracts events due to low-energy muons, thereby improving defect identification, so this model is effective.

Next, we examined muography images in which the present invention was applied and the energy range of incident muons was limited. As illustrated in FIG. 10, a test piece 40 having an internal cavity was simulated on a simulation system, and a muon transmission image was created by simulation while limiting the energy of muons incident on the test piece 40 to Ew _or less.

Simulation results under ideal conditions are shown in FIG. In addition to the normal muography simulation with cosmic ray muons _of total energy shown in FIG. shown in The measurement time is the same in both cases.

It can be seen that the image sharpens by limiting the energy range of the muons that create the transmission image. In this method, it is not always necessary to uniquely determine the energy of the incident muon, and it is sufficient to estimate the possible energy range of the incident muon.

In the above embodiment, the radiation to be measured is muons, but the radiation to be measured is not limited to muons. The scintillator and absorber are not limited to plastic or iron plate, and the output layer is not limited to 1 MeV.

DESCRIPTION OF SYMBOLS 10... Laminated detector 12... Scintillator 14... Absorber 16... Plastic scintillator (Plastin)
18... Iron plate 20... Neural network 22... Input layer 24, 26... Hidden layer 30... Output layer

Claims (4)

When determining the energy of radiation detected by a detector in which multiple scintillators and absorbers are stacked,
Using a neural network with the energy given to the scintillator of each layer as the input layer and the energy of the incident radiation as the output layer,
Machine learning of the relationship between the energy given to the scintillator of each layer and the energy of the incident radiation,
A radiation energy determination method for estimating the energy of incident radiation from the energy given to the scintillator of each layer using the machine learning result,
For a radiation event stationary in the absorber, the energy of the incident radiation is estimated by analyzing information on the amount of light emitted up to the scintillator in front of the stationary absorber by machine learning, and the radiation penetrating the detector. A radiation energy discrimination method using machine learning, characterized in that, for an event, the energy of incident radiation is estimated by analyzing information on the amount of light emitted from all scintillators by machine learning. 2. The radiation energy determination method using machine learning according to claim 1 , wherein said scintillator is a plastic scintillator, and said absorber is iron or lead. a detector in which a plurality of scintillators and absorbers are stacked;
a neural network whose input layer is the energy given to the scintillator of each layer and whose output layer is the energy of the incident radiation;
a means for performing machine learning on the relationship between the energy imparted to the scintillator of each layer and the energy of the incident radiation, and using the results of the machine learning to estimate the energy of the incident radiation from the energy imparted to the scintillator of each layer; ,
A radiation energy determination device comprising
For a radiation event stationary in the absorber, the energy of the incident radiation is estimated by analyzing information on the amount of light emitted up to the scintillator in front of the stationary absorber by machine learning, and the radiation penetrating the detector. A radiation energy discriminating apparatus using machine learning, characterized in that, for an event, information on the amount of light emitted from all scintillators is analyzed by machine learning to estimate the energy of incident radiation. 4. A radiation energy discrimination apparatus using machine learning according to claim 3 , wherein said scintillator is a plastic scintillator and said absorber is iron or lead.