Estimation method, estimation device and estimation program

Abstract

To precisely estimate an unknown radiation source distribution from a measured radiation energy spectrum.SOLUTION: An estimation device 10 is configured to: input a radiation energy spectrum measured under an unknown radiation source distribution to a first learning model 1 after machine learning using learning data based upon an input of a measured radiation energy spectrum in a known radiation source distribution and an output of a radiation energy spectrum calculated under the radiation source distribution; input the output of the first learning model 1 to a second learning model 2 after machine learning using learning data based upon an input of a radiation energy spectrum obtained through simulation for the known radiation source distribution and an output of the radiation source distribution; and then output the output of the second learning model 2 as an estimated value of an unknown radiation source distribution.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present disclosure relates to a radiation source distribution estimation method, an estimation device, and an estimation program. [Background technology]

[0002] Patent Document 1 discloses a device that includes a detection unit that outputs a detection signal corresponding to the energy of the incident radiation when radiation emitted from a radioactive material is incident thereon, an analysis unit that derives a first energy distribution, which is an energy distribution showing a count for each energy value of the detection signal, a storage unit that stores a response function of the detection unit, and a calculation unit that performs a signal restoration calculation using the response function on the first energy distribution to calculate a second energy distribution, which is an energy distribution of the radiation of the radioactive material, to discriminate the nuclides of the radioactive material and calculate the radioactivity intensity of the discriminated radioactive material, and the response function is and a second response function for an external radioactive substance that is the radioactive substance not contained in the object to be measured, wherein the calculation unit performs the signal restoration calculation using the second response function for the first energy distribution, thereby performing separation control to separate the radiation energy of the external radioactive substance from the calculated second energy distribution, and discriminates the nuclide of the measured radioactive substance based on the second energy distribution after the separation control, and calculates the radioactivity intensity of the discriminated measured radioactive substance.

[0003] Patent Document 2 discloses a medical image processing device that includes an acquisition unit that acquires a first spectrum including at least a detector response for each detection element that detects radiation, and a judgment unit that judges whether or not to apply to the first spectrum a trained model that outputs a second spectrum obtained by correcting an input spectrum based on the characteristics of the first spectrum.

[0004] Patent Document 3 discloses a radioactive surface contamination inspection device having a radiation detector that detects radiation incident on a planar detection area, an incident position calculator that identifies which of the mesh-like compartments into which the detection area is divided the radiation has been incident, a count holder that stores a compartment radiation integrated amount obtained by integrating the amount of radiation for each compartment into which the radiation has been incident, and a set area integrator that calculates a set area radiation integrated amount by adding up the compartment radiation integrated amounts corresponding to the compartments that belong to each set area formed by combining consecutive compartments. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2021-103153 A [Patent Document 2] JP 2020-22689 A [Patent Document 3] JP 2008-292166 A Summary of the Invention [Problem to be solved by the invention]

[0006] The amount of radiation and radiation energy spectrum in the environment are measured, and the measurement data can be used to estimate the location (distribution) of hotspots, fuel debris in reactors, and radiation sources (hereinafter simply referred to as "radiation sources"), which can be useful for decontamination, etc.

[0007] FIG. 12 is a diagram showing an example of the relationship between the position of a radiation source (e.g., cobalt 60 and cesium 137) and the measured radiation energy spectrum. Cobalt 60 is 60 Co, and cesium 137 137 It is expressed as Cs.

[0008] In a situation where the position of a radiation source and the amount of radiation emitted from the radiation source are known in advance, estimating the radiation energy spectrum at a specific position is a forward problem of estimating the result from the cause, and therefore estimation using a mathematical model is possible. However, measuring the radiation energy spectrum at a specific position and estimating the position of the radiation source from the measurement results is an inverse problem of estimating the cause from the result, and therefore estimation using a mathematical model like a forward problem is difficult.

[0009] Therefore, radiation sources are sought out by a treasure-hunting technique, such as searching for areas with high radiation levels, based on the results of measurements of radiation energy spectra using a radiation detector.

[0010] Meanwhile, research into machine learning as a method of big data analysis is being conducted. With machine learning, even if the mathematical model existing between input and output is unknown, as long as the correspondence between the input and output is known, it is possible to extract features potentially contained between the input and output and generate a learning model that reproduces the correspondence between the input and output. Therefore, by using a learning model in which the radiation energy spectrum measured at each position is used as input and the position of the radiation source at the time of measuring the radiation energy spectrum is used as output, it becomes possible to estimate the position of the radiation source from the measured radiation energy spectrum.

[0011] In this case, learning data to be used for machine learning of the learning model is required, but it is often not realistic to generate the learning data by measuring the radiation energy spectrum in various environments generated while actually changing the position of the radiation source 8. Therefore, a method is used in which the learning data is generated using measurement data at each position calculated from the known position of the radiation source 8 using a Monte Carlo simulation for calculating radiation transport.

[0012] However, the radiation energy spectrum obtained by the Monte Carlo simulation for radiation transport calculation may differ from the radiation energy spectrum measured by a radiation detector in terms of the position and shape of the peak, even if the resolution correction is performed. Although the error between the calculated radiation energy spectrum and the actually measured radiation energy spectrum is often not a problem in nuclide identification, etc., it affects the estimation result when minute changes in the radiation energy spectrum are used as features, such as in the estimation of radiation source distribution. Therefore, when the actually measured radiation energy spectrum is input to a learning model that machine-learns the correspondence between the radiation energy spectrum for each position calculated under a known radiation source distribution and the known radiation source distribution, the estimated radiation source distribution may contain a non-negligible error.

[0013] The present disclosure has been made in consideration of the above circumstances, and aims to provide an estimation method, an estimation device, and an estimation program that can accurately estimate an unknown radiation source distribution from a measured radiation energy spectrum, compared to estimating an unknown radiation source distribution by inputting a radiation energy spectrum for each position measured under an unknown radiation source distribution into a learning model that has machine-learned the correspondence between the radiation energy spectrum for each position calculated under a known radiation source distribution and the known radiation source distribution. [Means for solving the problem]

[0014] In the method for estimating a radiation source distribution according to the first aspect, a computer executes a process of inputting a radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution into a first learning model which has been machine-learned using first learning data, the first learning model having as input a measured radiation energy spectrum which is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution and as output a first calculated radiation energy spectrum which is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation in which the input measured radiation energy spectrum was measured, inputting an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution into a second learning model which has been machine-learned using second learning data, the second learning model having as input a second calculated radiation energy spectrum which is a radiation energy spectrum generated by simulation from a known radiation source distribution and as output a radiation source distribution in a situation in which the second calculated radiation energy spectrum was obtained, and outputting the output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution as an estimate of the unknown radiation source distribution.

[0015] A second aspect of the radiation source distribution estimation method is the method of estimating radiation source distribution according to the first aspect, wherein the ground truth data of the second learning data includes a known radiation source distribution composed of a plurality of types of radiation sources, and the second learning model outputs a distribution of the plurality of types of radiation sources estimated from the output of the first learning model.

[0016] A third aspect of the radiation source distribution estimation method is the radiation source distribution estimation method of the second aspect, wherein the ground truth data of the second learning data further includes a known distribution of shielding objects that block radiation emitted from a radiation source, and the second learning model outputs the distribution of the shielding objects estimated from the output of the first learning model.

[0017] A method for estimating a radiation source distribution according to a fourth aspect includes the method for estimating a radiation source distribution according to the third aspect, in which the radiation source emits gamma rays and the shielding body includes uranium.

[0018] A radiation source distribution estimation method according to a fifth aspect is a radiation source distribution estimation method according to a fourth aspect, in which synthetic basic calculation data obtained by synthesizing each of the basic calculation data, which is a radiation energy spectrum generated by simulation assuming that each of a plurality of radiation sources included in a known radiation source distribution exists alone, is input, and machine learning is performed using learning data in which a radiation energy spectrum generated by simulation assuming that the plurality of radiation sources exist simultaneously is output to a 0th learning model, and the output of the 0th learning model when the synthetic basic calculation data is input is used as an input of the second learning data.

[0019] A sixth aspect of the method for estimating a radiation source distribution is a method for estimating a radiation source distribution according to any one of the first to fifth aspects, in which the output of the first learning model is normalized before being input to the second learning model.

[0020] The radiation source distribution estimation device according to the seventh aspect includes a processor, and the processor inputs a radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution into a first learning model that has been machine-learned using first learning data, the first learning model receiving as input a measured radiation energy spectrum that is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, and outputting a first calculated radiation energy spectrum that is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation in which the input measured radiation energy spectrum was measured; inputs an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution into a second learning model that has been machine-learned using second learning data, the second learning model receiving as input a second calculated radiation energy spectrum that is a radiation energy spectrum generated by simulation from a known radiation source distribution, and outputting a radiation source distribution in a situation in which the second calculated radiation energy spectrum was obtained; and outputs the output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution as an estimate of the unknown radiation source distribution.

[0021] A radiation source distribution estimation program according to an eighth aspect is a program for causing a computer to execute a process of inputting a measured radiation energy spectrum, which is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, into a first learning model which has been subjected to machine learning using first learning data which outputs a first calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation in which the input measured radiation energy spectrum was measured, inputting a radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution into a first learning model which has been subjected to machine learning using first learning data which outputs a first calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, and inputting an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution into a second learning model which has been subjected to machine learning using second learning data which inputs a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, and outputting a radiation source distribution in a situation in which the second calculated radiation energy spectrum was obtained, and outputting the output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution as an estimate of the unknown radiation source distribution. Effect of the Invention

[0022] The first, seventh, and eighth aspects have the advantage that the unknown radiation source distribution can be accurately estimated from the measured radiation energy spectrum, compared to a case in which the unknown radiation source distribution is estimated by inputting the radiation energy spectrum for each position measured under an unknown radiation source distribution into a learning model that has been machine-learned to learn the correspondence between the radiation energy spectrum for each position calculated under a known radiation source distribution and the known radiation source distribution.

[0023] According to the second aspect, there is an effect that it is possible to estimate the radiation source distribution for each type of radiation source.

[0024] According to the third aspect, there is an effect that it is possible to estimate the shielding object distribution in addition to the radiation source distribution.

[0025] According to the fourth aspect, there is an effect that it is possible to estimate the shielding distribution of shielding materials other than lead.

[0026] According to the fifth aspect, there is an effect that it is possible to reduce the calculation load when obtaining the second calculated radiation energy spectrum.

[0027] According to the sixth aspect, there is an effect that the estimation accuracy of the radiation source distribution can be improved compared to the case where the output of the first learning model is input to the second learning model without being normalized. [Brief description of the drawings]

[0028] [Figure 1] FIG. 1 illustrates an example of the configuration of an estimation device. [Diagram 2] FIG. 13 is a diagram illustrating an example of machine learning of a first learning model. [Diagram 3] FIG. 11 is a diagram illustrating an example of acquisition of measurement data. [Figure 4] FIG. 13 is a diagram illustrating an example of machine learning of a second learning model. [Diagram 5] 13 is a flowchart showing an example of a process for estimating a radiation source distribution. [Figure 6] FIG. 13 is a diagram showing an example of a correct distribution of radiation sources. [Figure 7] FIG. 11 is a diagram showing an example of measurement data. [Figure 8] FIG. 11 is a diagram illustrating an example of converted data. [Figure 9] FIG. 13 is a diagram illustrating an example of calculation data. [Figure 10] A figure showing an example of a radiation source distribution estimated by inputting measurement data into the second learning model, and an example of a radiation source distribution estimated by inputting calculation data into the second learning model. [Figure 11] FIG. 13 is a diagram showing an example of a radiation source distribution estimated by inputting converted data into a second learning model. [Figure 12] FIG. 13 is a diagram showing an example of the relationship between the position of a radiation source and a measured radiation energy spectrum. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0029] Hereinafter, the present embodiment will be described with reference to the drawings. Note that the same components and processes are given the same reference numerals throughout the drawings, and duplicated explanations will be omitted.

[0030] 1 is a diagram showing an example of the configuration of an estimation device 10 that estimates a radiation source distribution. The estimation device 10 includes functional units, such as a control unit 11, a communication unit 12, a display unit 13, a reception unit 14, an interface (I / F) unit 15, and a storage unit 16. The control unit 11, the communication unit 12, the display unit 13, the reception unit 14, the interface (I / F) unit 15, and the storage unit 16 are connected by a bus 17, and transmit and receive data between them.

[0031] The control unit 11 controls each functional unit based on instructions from a user who operates the estimation device 10, and estimates the radiation source distribution. The radiation source distribution is a distribution that indicates the positions of the radiation sources 8. The control unit 11 includes a CPU (Central Processing Unit) 11A, which is an example of a processor, a non-volatile memory 11B, and a RAM (Random Access Memory) 11C that is used as a temporary work area for the CPU 11A.

[0032] The nonvolatile memory 11B is an example of a storage device that maintains stored information even if the power supplied to the nonvolatile memory 11B is cut off, and is, for example, a semiconductor memory such as an SSD (Solid State Drive), but a hard disk may also be used. The nonvolatile memory 11B stores in advance an estimation program 11D that the CPU 11A reads in order to perform control related to the estimation of the radiation source distribution, and various parameters that the CPU 11A refers to when performing control related to the estimation of the radiation source distribution.

[0033] The communication unit 12 has a communication protocol for performing data communication, and performs data communication with the external device 4 connected to a communication line. The communication line connected to the communication unit 12 may be either wired or wireless. There are no restrictions on the data communication standard supported by the communication unit 12, and data communication is possible using, for example, Wi-Fi (registered trademark), Bluetooth (registered trademark), and NFC (Near Field Communication), etc.

[0034] The display unit 13 displays the information processed by the control unit 11 to the user through the display device 5. The display unit 13 may display the information processed by the control unit 11 to the user by printing it on a recording medium such as paper through a printer (not shown).

[0035] The reception unit 14 receives instructions and various parameters related to the estimation of the radiation source distribution from the user through the input device 6, and notifies the control unit 11. There are no restrictions on the type of input device 6 that the user uses to input instructions, etc., and examples of the input device 6 that may be used include a switch, a touch panel, a touch pen, a keyboard, and a mouse.

[0036] The I / F unit 15 provides an interface function of notifying the control unit 11 of instructions to a sodium iodide spectrometer 7 (hereinafter referred to as "NaI spectrometer 7"), which is an example of a radiation detector, and notifying the control unit 11 of the radiation energy spectrum measured by the NaI spectrometer 7. The radiation detector connected to the I / F unit 15 is not particularly limited, and may be a germanium detector (also referred to as "Ge detector") or a thallium-activated sodium iodide detector (also referred to as "NaI(TI) spectrometer") as long as it can measure the energy spectrum of the radiation depending on the type of radiation to be detected.

[0037] The storage unit 16 stores, for example, the first learning model 1 and the second learning model 2. The storage unit 16 uses a semiconductor memory like the non-volatile memory 11B, but may also use a hard disk. Note that the storage unit 16 may be used as the non-volatile memory 11B, in which case the estimation program 11D and various parameters are stored in the storage unit 16.

[0038] The first learning model 1 is a learning model in which machine learning is performed using learning data in which a radiation energy spectrum measured using a NaI spectrometer 7 for a known radiation source distribution is input, and a radiation energy spectrum generated by simulation from a radiation source distribution in a situation in which the input radiation energy spectrum is measured is output. That is, the first learning model 1 estimates a radiation energy spectrum obtained when calculated under the same environment from a radiation energy spectrum measured using a NaI spectrometer 7. For example, PHITS (Particle and Heavy Ion Transport code System) is used to calculate a radiation energy spectrum at each position under a specific radiation source distribution, but other Monte Carlo simulations or methods derived therefrom may also be used. Furthermore, finite element method, finite difference method, PINNs (Physics-Informed Neural Networks), and the like other than Monte Carlo simulation may also be used, and the simulation method is not limited to Monte Carlo simulation.

[0039] For convenience of explanation, the radiation energy spectrum measured using the NaI spectrometer 7 will be referred to as "measured data", and the radiation energy spectrum calculated using PHITS under a specific radiation source distribution will be referred to as "calculated data". The measured data according to the present disclosure is an example of a measured radiation energy spectrum, and the calculated data according to the present disclosure is an example of a first calculated radiation energy spectrum. In the first learning model 1, the calculated data becomes the correct answer data a for the measured data.

[0040] 2 is a diagram showing an example of machine learning of the first learning model 1. The first learning model 1 is represented by a three-layer neural network composed of, for example, an input layer, an intermediate layer, and an output layer. Note that the number of layers is not limited to three, and more layers may be used.

[0041] The number of nodes x in the input layer in the first learning model 1 is set to the number of energy band channels in the measurement data, and the number of nodes y in the output layer in the first learning model 1 is set to the number of energy band channels in the calculation data. The number of energy band channels refers to a number assigned to each energy band in ascending order from the lowest to the highest energy range represented by the energy band, after dividing energy into a plurality of consecutive energy bands.

[0042] Although there is no restriction on the number of channels in the energy band, as an example in the present disclosure, the energy is divided into 64 energy bands for handling, so the number of channels in the energy band is "64". Since the intensity of the measured data and the calculated data is measured and calculated for each channel, the measured data and the calculated data are each expressed as a vector having 64 elements. Therefore, the number of nodes x in the input layer of the first learning model 1, which uses the measured data as input and the calculated data as output, and the number of nodes y in the output layer are both 64. The intensity of each channel is expressed by cps (Counts Per Second), which represents the counted amount of radiation.

[0043] Although there is no restriction on the number of nodes z in the intermediate layer of the first learning model 1, in the present disclosure, as an example, the number is set to 64, which is the same as the number of nodes x and y. That is, the first learning model 1 of the present disclosure has nodes x1 to x 64 , Node z1 to Node z in the middle layer 64 , and in the output layer, nodes y1 to y 64If the number of nodes z is made smaller than the number of nodes x and y, the features of the measurement data are compressed in the intermediate layer, and if the number of nodes z is made larger than the number of nodes x and y, the features of the measurement data are projected to a higher dimension than the number of dimensions of the measurement data in the intermediate layer.

[0044] Next, a method of acquiring measurement data used for machine learning of the first learning model 1 will be described. FIG. 3 is a diagram showing an example of acquiring measurement data. When acquiring measurement data, the measurement range is divided into a plurality of regions as shown in FIG. 3(A), and measurement data at the position of each region is acquired using a NaI spectrometer 7. There is no restriction on the number of regions divided in the measurement range, but in this disclosure, as an example, the measurement range is divided into four vertically and four horizontally to generate 16 grid-like regions. There is no restriction on the number of divisions of the measurement range, and it may be less than 16 or more than 17.

[0045] 3A, regions 9A and 9B are regions where radiation sources 8 are placed, and region 9C is a region where a shielding body 20 is placed. In this manner, for example, a shielding body 20 resembling a forest or a building that functions as a radiation shield in an actual environment may be placed in the measurement environment of the measurement data. In this case, the radiation source distribution is affected by the position, shape, size, material, etc. of the shielding body 20.

[0046] For example, cesium 137 is used as the radiation source 8, but cobalt 60 may also be used. In addition, multiple types of radiation sources 8 may be used, or multiple sources of the same type may be used. In the measurement environment of the present disclosure, multiple types of radiation sources 8, namely, cesium 137 and cobalt 60, are used. In FIG. 3(A), region 9A indicates the region where cesium 137 is arranged, and region 9B indicates the region where cobalt 60 is arranged. For example, lead, iron, concrete, etc. are used as the shielding body 20. Note that the type of radiation emitted by the radiation source 8 is not particularly limited as long as the energy spectrum of the radiation can be measured. In addition, when the radiation source 8 emits gamma rays, uranium may be used as the shielding body 20. Uranium has a high density and therefore has the effect of shielding gamma rays, etc. From a different perspective, in the present disclosure, it is also possible to estimate the distribution of uranium present as a shielding body, that is, in the present disclosure, it is also possible to estimate the distribution of uranium, although it is an estimate of the distribution of the shielding body. Of course, the present disclosure also makes it possible to estimate the distribution of uranium by measuring and learning the radiation energy spectrum emitted by uranium using uranium as the radiation source 8. In addition, while the radiation source 8 and the shielding body 20 are arranged on the same plane in Fig. 3, the present invention is not limited to this and may be arranged three-dimensionally.

[0047] Fig. 3(B) is a side view of the measurement environment shown in Fig. 3(A) when viewed along the V-axis direction. As shown in Fig. 3(B), radiation sources 8 of each type are placed on a loading surface 18 such as a table corresponding to areas 9A and 9B, and a shielding body 20 is placed on the loading surface 18 corresponding to area 9C.

[0048] The user places the NaI spectrometer 7 at a position on the detection surface 19 that is a specified distance away from the loading surface 18 in the W-axis direction, and measures the measurement data in each region. There is no restriction on the specified distance that specifies the distance from the loading surface 18 to the NaI spectrometer 7, but in this disclosure, the specified distance is set to 1 cm as an example. Also, like the placement of the radiation source 8, the measurement data is not limited to being measured on the same plane, and may be measured three-dimensionally.

[0049] The estimation device 10 measures the measurement data for each of the 16 regions using the NaI spectrometer 7 under a known radiation source distribution as shown in Fig. 3, and calculates the calculation data for each region for the same positions of the radiation source 8 and shield 20 as when the measurement data was taken, using PHITS. Using the measurement data and calculation data for the same regions thus obtained, learning data is generated in which the measurement data and the calculation data correspond to input data and output for the input data, i.e., correct answer data a. The measurement data is used as input data, and the calculation data is correct answer data a1 to correct answer data a 64 The learning data set as above is an example of first learning data.

[0050] In the present disclosure, as an example, the positions of the radiation source 8 and the shielding body 20 are changed in about 34 patterns, and measurement data and calculation data for each of 16 regions for each pattern are obtained to generate 544 pieces of learning data. The first learning model 1 is a learning model obtained by performing machine learning using the 544 pieces of learning data 10,000 times. Specifically, when each measurement data included in the 544 pieces of learning data is input to node x in the input layer, the first learning model 1 generates the calculation data (i.e., the correct answer data a1 to a2) included in the same learning data as the input measurement data. 64 ) is output from node y in the output layer.

[0051] The number of pieces of learning data used in the machine learning of the first learning model 1 and the number of times of machine learning are just examples, and are not limited to 544 and 10,000, respectively. The user may prepare the number of pieces of learning data that is considered optimal for the machine learning of the first learning model 1, and set the number of times of machine learning that is considered optimal for the machine learning of the first learning model 1.

[0052] On the other hand, the second learning model 2 shown in FIG. 1 is a learning model in which machine learning is performed using learning data in which calculation data at each position in a known radiation source distribution is input and known radiation source distribution in the situation in which the calculation data was obtained is output. In the present disclosure, the calculation data is calculated for each region of the measurement range shown in FIG. 3. That is, the second learning model 2 estimates the radiation source distribution from the calculation data in each of the 16 divided regions in FIG. 3 calculated using PHITS. The calculation data input to the second learning model 2 is an example of the second calculated radiation energy spectrum. Here, calculation data in a three-dimensional radiation source distribution may be used in the second learning data as in the first learning data. The second learning data is learning data used for machine learning of the second learning model 2. By using data on a three-dimensional radiation source distribution in the first learning data and the second learning data, the radiation source distribution can be estimated three-dimensionally with high accuracy.

[0053] Fig. 4 is a diagram showing an example of machine learning of the second learning model 2. Like the first learning model 1 shown in Fig. 2, the second learning model 2 is also represented by a three-layer neural network composed of, for example, an input layer, an intermediate layer, and an output layer.

[0054] The number of nodes x in the input layer in the second learning model 2 is set to the number of elements when the intensities of each channel of the energy band in the calculation data of each of the 16 divided regions calculated under the same radiation source distribution are arranged in order for all regions. In the example of the present disclosure, the number of channels of the energy band in the calculation data of one region is 64 channels, and the number of regions for which calculation data is to be calculated under the same radiation source distribution is 16, so 64×16=1024 nodes x are used.

[0055] The number of nodes y in the output layer in the second learning model 2 is set to the number necessary to represent the radiation source distribution in the situation where the calculation data input to the input layer is calculated. In the case of the example of the present disclosure, since there are 16 regions in which the radiation source 8 can be placed, for example, if the value of the region in which the radiation source 8 is placed is represented by "1" and the value of the region in which the radiation source 8 is not placed is represented by "0", the distribution of the radiation source 8 can be represented by a 16-bit bit string. As already described, in the present disclosure, as an example, two types of radiation sources 8, cesium 137 and cobalt 60, are used, so that a 32-bit bit string is required to represent the distribution of the radiation source 8 for each type of radiation source 8.

[0056] The same can be said about the distribution of the shielding bodies 20. If the type of shielding body 20 used in the calculation of the calculation data is only lead, the distribution of the shielding bodies 20 is represented by a bit string of 16 bits. Therefore, a bit string of 48 bits is sufficient to represent the radiation source distribution using the shielding bodies 20 and two types of radiation sources 8. If one node y of the second learning model 2 is assigned to each bit representing the radiation source distribution, the output layer of the second learning model 2 represents the radiation source distribution in the situation where the calculation data is calculated, and therefore the number of nodes y is set to 48.

[0057] Although there is no restriction on the number of nodes z in the intermediate layer of the second learning model 2, in the present disclosure, as an example, the number is set to 1024, which corresponds to the number of nodes x. That is, the second learning model 2 of the present disclosure has nodes x1 to x 1024 , Node z1 to Node z in the middle layer 1024 , Node y1 to Node y in the output layer 48 The learning model has the following structure. The calculation data is used as input data, and the radiation source distribution is correct data a1 to correct data a 48 The learning data set as above is an example of second learning data.

[0058] In the present disclosure, as an example, the positions of the radiation source 8 and the shielding body 20 are changed in about 15,000 patterns, and 15,000 pieces of learning data are generated by acquiring calculation data for each of the 16 regions for each pattern. That is, calculation data (also called "integrated calculation data") that is a compilation of calculation data for each region calculated under a known radiation source distribution is used as input data, and 15,000 pieces of learning data are generated by associating the radiation source distribution in the situation in which the calculation data is calculated as the correct answer data a for the input data while changing the radiation source distribution. The second learning model 2 is a learning model obtained by performing machine learning using the 15,000 pieces of learning data thus generated.

[0059] Specifically, when each integrated calculation data included in 15,000 pieces of learning data is input to node x of the input layer, the second learning model 2 detects the radiation source distribution (i.e., the correct answer data a1 to a2) included in the same learning data as the input integrated calculation data. 48 ) is output from node y in the output layer.

[0060] The number of pieces of learning data used in the machine learning of the second learning model 2 is an example and is not limited to 15000. The user may prepare the number of pieces of learning data that is considered optimal for the machine learning of the second learning model 2.

[0061] The distribution of radiation sources 8 represented by the output layer of the second learning model 2 is not limited to two types of radiation sources 8, but may be a distribution of one type of radiation source 8, a distribution of radiation sources 8 when there are multiple radiation sources 8 of the same type, or a distribution of three or more types of radiation sources 8. That is, the learning data of the second learning model 2 includes a known distribution of at least one type of radiation source 8 as an output for the integrated calculation data. In addition, when a shielding object 20 is present, the learning data of the second learning model 2 includes a known distribution of the shielding object 20 as an output for the integrated calculation data.

[0062] Next, the operation of the estimation device 10 for estimating a radiation source distribution will be described in detail. Fig. 5 is a flow chart showing an example of a process of estimating a radiation source distribution executed by the estimation device 10 when an instruction to estimate a radiation source distribution for measurement data measured under an unknown radiation source distribution is received from a user in the measurement environment of the measurement data shown in Fig. 3. The CPU 11A of the control unit 11 in the estimation device 10 reads an estimation program 11D from the non-volatile memory 11B and executes the process of estimating a radiation source distribution.

[0063] It is assumed that measurement data measured under an unknown radiation source distribution is stored in advance in the storage unit 16. In addition, in the description of the radiation source distribution estimation process, "measurement data" refers to measurement data measured under an unknown radiation source distribution.

[0064] First, in step S10, the control unit 11 acquires the first learning model 1 from the storage unit 16 and expands it in the RAM 11C. The control unit 11 inputs measurement data to the input layer of the acquired first learning model 1. As a result, an estimated value of the calculation data obtained under the same measurement environment as the measurement environment in which the measurement data was measured is output from the output layer of the first learning model 1.

[0065] Therefore, in step S20, the control unit 11 obtains an estimate of the calculation data for the measurement data output by the first learning model 1. That is, the measurement data is converted into the calculation data by the first learning model 1. Hereinafter, the estimate of the calculation data obtained by inputting the measurement data into the first learning model 1 is referred to as the “converted data” to distinguish it from the calculation data actually calculated using PHITS.

[0066] In step S30, the control unit 11 acquires the second learning model 2 from the storage unit 16 and expands it in the RAM 11C. The control unit 11 inputs the conversion data acquired by the process of step S20 to the input layer of the second learning model 2. When the control unit 11 inputs the conversion data to the input layer of the second learning model 2, the conversion data may be normalized before inputting it to the second learning model 2 instead of inputting it as is. For example, the intensity of the channel of the conversion data with the maximum intensity may be set to "1", and the intensity of each channel at each position may be converted to a range of 0 to 1 before inputting it to the second learning model 2. Note that when the normalized conversion data is input to the second learning model 2, it is preferable that the machine learning of the second learning model 2 is also performed by learning data in which the normalized conversion data is input as input data.

[0067] As a result, an estimate of the unknown radiation source distribution in the measurement environment in which the measurement data was measured is output from the output layer of the second learning model 2.

[0068] Therefore, in step S40, the control unit 11 acquires an estimated value of the radiation source distribution output by the second learning model 2.

[0069] In step S50, the control unit 11 controls the display unit 13 to visualize the estimated value of the radiation source distribution acquired by the process of step S40, and displays it on the display device 5 as a radiation source distribution estimated from the measurement data. Specifically, the control unit 11 controls the display unit 13 to display the presence or absence of the radiation source 8 and the shielding body 20 for each of the 16 divided regions. This ends the process of estimating the radiation source distribution shown in Fig. 5.

[0070] In this way, the estimation device 10 first generates calculated data, i.e., converted data, from the measurement data, which is obtained under the same measurement environment as the measurement environment in which the measurement data was measured, and then estimates the radiation source distribution by inputting the generated converted data into the second learning model 2. The radiation source distribution to be estimated in the present invention refers not only to the position of the radiation source 8 but also to the strength of the radiation source 8 (i.e., the strength of the radioactivity of each radiation). Therefore, the present invention can also estimate the distribution of radioactive fuel debris at the Fukushima Daiichi Nuclear Power Plant, for example.

[0071] <Example of Estimation of Radiation Source Distribution> Next, the estimation result of the radiation source distribution estimated by the estimation process of the radiation source distribution shown in Fig. 5 will be shown. Hereinafter, the estimation result of the radiation source distribution estimated by the estimation process of the radiation source distribution shown in Fig. 5 will be referred to as the "estimation result by the estimation device 10".

[0072] Fig. 6 is a diagram showing an example of a correct radiation source distribution. As shown in Fig. 6, cesium 137 and cobalt 60 were placed in areas 9A and 9B of the measurement range, respectively. Note that no shielding body 20 was placed in the measurement range.

[0073] FIG. 7 is a diagram showing an example of measurement data measured at positions (1), (2), and (3) in the measurement range shown in FIG.

[0074] FIG. 8 is a diagram showing an example of converted data obtained by inputting the measurement data at positions (1), (2), and (3) in the measurement range shown in FIG.

[0075] FIG. 9 is a diagram showing examples of calculation data at positions (1), (2), and (3) in the measurement range, calculated under the radiation source distribution shown in FIG.

[0076] On the other hand, Figure 10 shows the correct radiation source distribution, the radiation source distribution estimated by inputting the measurement data shown in Figure 7 into the second learning model 2, and the radiation source distribution estimated by inputting the calculation data shown in Figure 9 into the second learning model 2, for each type of radiation source 8.

[0077] FIG. 11 is a diagram showing the correct radiation source distribution and the radiation source distribution estimated by inputting the converted data shown in FIG. 8 into the second learning model 2 for each type of radiation source 8. In FIG.

[0078] In Figure 10, the error between the radiation source distribution estimated from the measured data and the correct radiation source distribution was 1.06. Also, the error between the radiation source distribution estimated from the calculated data and the correct radiation source distribution was 6.54E-7 (E is a power of 10). In the case of the radiation source distribution estimated from the measured data, it was estimated that cesium-137 was present in a position where it did not actually exist.

[0079] The reason why the radiation source distribution estimated from the calculation data has a higher estimation accuracy than the radiation source distribution estimated from the measurement data is that the second learning model 2 is a learning model that has learned the correspondence between the calculation data and the radiation source distribution. As can be seen by comparing the measurement data and the calculation data shown in Figures 7 and 9, the measurement data and the calculation data show different trends even at the same position. Therefore, if measurement data that deviates from the calculation data is input to the second learning model 2, the estimated radiation source distribution also shows a tendency to deviate from the correct radiation source distribution that is associated with the calculation data.

[0080] On the other hand, in FIG. 11, the error between the radiation source distribution estimated from the converted data and the correct radiation source distribution was 7.83E-6. The reason why the radiation source distribution estimated from the converted data has a higher estimation accuracy than the radiation source distribution estimated from the measured data is due to the difference in waveform between the converted data and the measured data. As can be seen by comparing the measured data, converted data, and calculated data shown in FIG. 7 to FIG. 9, at any of positions (1) to (3), the waveform of the converted data shown in FIG. 8 tends to be closer to the waveform of the calculated data shown in FIG. 9 than the waveform of the measured data shown in FIG. 7. Therefore, the radiation source distribution estimated from the converted data has a higher estimation accuracy than the radiation source distribution estimated from the measured data.

[0081] Although one aspect of the estimation device 10 has been described above using the embodiment, the disclosed aspect of the estimation device 10 is merely an example, and the aspect of the estimation device 10 is not limited to the scope described in the embodiment. For example, the first learning model and the second learning model may be changed from those in the embodiment, or in addition to the first learning model and the second learning model, another learning model may be added, and the number of additional models is not particularly limited. For example, a learning model in which machine learning is performed using learning data in which measured radiation energy spectra obtained from a plurality of types of radiation sources 8 are input and radiation source spectra classified by type of radiation source 8 are output may be further used. Furthermore, for example, in obtaining the second calculated radiation energy spectrum, the 0th learning model described below may be used. Here, the 0th learning model is a learning model used to calculate the second calculated radiation energy spectrum.

[0082] In the above example, when the second calculated radiation energy spectrum, which is the input data of the second learning data, is generated, the calculation data is simulated for each distribution pattern of the radiation source 8. Therefore, when multiple radiation sources 8 are present in the 16-divided region of FIG. 3, the calculation data is simulated for each combination of the multiple radiation sources 8 while changing the distribution pattern of the radiation source 8, which may make the calculation data simulation complicated. However, when the 0th learning model is used to generate the second calculated radiation energy spectrum, the control unit 11 calculates the radiation energy spectrum for each distribution of the radiation sources 8, for example, and synthesizes the calculation data for each radiation source 8 for each region to obtain the second calculated radiation energy spectrum. Therefore, the second calculated radiation energy spectrum when multiple radiation sources 8 are distributed can be obtained without performing a simulation of the radiation energy spectrum when multiple radiation sources 8 are arranged in the 16-divided region.

[0083] For example, generation of the second calculated radiation energy spectrum in a case where three radiation sources 8 are arranged in an area divided into 16 parts, such as area 9A and area 9B in FIG. 6, will be described.

[0084] When the 0th learning model is not used, it is necessary to perform a simulation calculation assuming that three radiation sources 8 are arranged as they are. However, when the 0th learning model is used, the control unit 11 individually calculates three pieces of calculation data, namely, calculation data when only cobalt 60 is arranged in the region 9B (hereinafter also referred to as "basic calculation data A"), calculation data when only cesium 137 is arranged in one of the two regions 9A (hereinafter also referred to as "basic calculation data B"), and calculation data when only cesium 137 is arranged in the other region 9A (hereinafter also referred to as "basic calculation data C"). Then, the control unit 11 generates a synthesis of the basic calculation data A, the basic calculation data B, and the basic calculation data C for each region and for each channel (hereinafter also referred to as "synthetic basic calculation data") as an input of the 0th learning model. "Synthesis" is an act of generating another data (in this case, "synthetic basic data") from a plurality of basic calculation data, and means, for example, adding each of the basic calculation data A to C for each region and for each channel.

[0085] By inputting the composite basic calculation data to the 0th learning model, the control unit 11 can obtain the second calculated radiation energy spectrum in the radiation source distribution in which three radiation sources 8 are arranged at the same time as the output of the 0th learning model, as shown in FIG. 6. In other words, the control unit 11 decomposes the radiation sources 8 that are simultaneously present in the 16 divided regions for each radiation source 8 (for example, in the example shown in FIG. 6, the three radiation sources 8 are treated as being present individually), and synthesizes the basic calculation data A to C calculated assuming that any one radiation source 8 exists in the 16 divided regions, thereby obtaining the second calculated radiation energy spectrum in the case where multiple radiation sources 8 exist at each position at the same time. Therefore, the control unit 11 can obtain the second calculated radiation energy spectrum without simulating the calculation data of each region in the radiation source distribution including multiple radiation sources 8, and therefore the calculation load of the second calculated radiation energy spectrum can be reduced. The effect of reducing the calculation load becomes greater as the number of radiation sources 8 and shields 20 increases, and as the number of divided regions increases.

[0086] The input data for the learning data for machine learning of the 0th learning model is, for example, data obtained by synthesizing the radiation energy spectrum for each channel in each of 16 divided regions in the case where a plurality of known radiation sources 8 exist independently, for each radiation source 8. The correct answer data for the learning data for machine learning of the 0th learning model is calculated data obtained by performing a simulation assuming that a plurality of known radiation sources 8 exist simultaneously within the 16 divided regions. The 0th learning model can also handle three-dimensional radiation source distribution data, like the first and second learning models.

[0087] Furthermore, like the first learning model and the second learning model, the 0th learning model may be represented, for example, by a three-layer neural network consisting of an input layer, an intermediate layer, and an output layer, or may be represented by even more layers.

[0088] In the above, an example has been described in which calculation data for a case where multiple radiation sources 8 are distributed is calculated under a situation where only one radiation source 8 exists in each region, but the position of the radiation source 8 is not limited to this. For example, multiple radiation sources 8 may exist in the same region, and the multiple radiation sources 8 may be the same type or different types. Also, calculation data for a case where a radiation source 8 and a shielding body 20 exist in the same region can be used as basic calculation data.

[0089] For the sake of convenience, an example has been described in which the estimated range of the radiation source distribution is limited to the range of an area divided into 16, but the estimated range of the radiation source distribution may be expanded to a wider range. In this case, the wider the estimated range of the radiation source distribution, the more complicated the calculation of the calculation data when multiple radiation sources 8 exist simultaneously. In this case, for example, the entire area divided into 16 may be treated as one unit representing whether or not a radiation source 8 exists. In addition, if the second calculated radiation energy spectrum is generated using the 0th learning model, the calculation load of the second calculated radiation energy spectrum can be reduced, even if the estimated range of the radiation source distribution is wide, compared to the case where the second calculated radiation energy spectrum is calculated assuming that multiple radiation sources 8 exist simultaneously in the area.

[0090] In the above, addition of basic calculation data has been exemplified as a method of synthesizing basic calculation data, but it goes without saying that when, for example, a shielding object 20 is included, when the intensity of the basic calculation data is corrected, or when standardized basic calculation data is used, the basic calculation data can be synthesized using other calculation methods such as subtraction, multiplication and division depending on the situation.

[0091] In particular, when the number or types of radiation sources 8 are small, instead of a measured or calculated radiation energy spectrum, an integrated value of the radiation energy spectrum obtained over a certain period of time can be used as each learning data and as input data for each learning model when performing estimation processing.

[0092] The present disclosure can also be applied to estimating a radiation source using radiation intensity obtained by a survey meter or the like instead of an energy spectrum.

[0093] Various modifications or improvements can be made to the embodiments without departing from the gist of the present disclosure, and such modifications or improvements are also included in the technical scope of the disclosure.

[0094] For example, other processes may be added to the process of estimating the radiation source distribution shown in FIG. 5 without departing from the gist of the present disclosure.

[0095] In the above embodiment, as an example, the radiation source distribution estimation process shown in Fig. 5 is implemented by software processing. However, the same process as that shown in the flowchart of the radiation source distribution estimation process may be implemented by hardware. In this case, the processing speed can be increased compared to when the radiation source distribution estimation process is implemented by software processing.

[0096] In the above embodiment, the term "processor" refers to a processor in a broad sense, and includes a general-purpose processor (e.g., CPU 11A) and a dedicated processor (e.g., a GPU: Graphics Processing Unit, an ASIC: Application Specific Integrated Circuit, an FPGA: Field Programmable Gate Array, a programmable logic device, etc.).

[0097] In addition, the operations of the processor in the above embodiment may be performed not only by one processor but also by multiple processors located at physically separate locations working together. The order of the operations of the processor is not limited to the order described in the above embodiment and may be changed as appropriate.

[0098] In the above embodiment, an example in which the estimation program 11D is stored in the non-volatile memory 11B has been described. However, the storage destination of the estimation program 11D is not limited to the non-volatile memory 11B. The estimation program 11D of the present disclosure can also be provided in a form recorded in a computer-readable storage medium.

[0099] For example, the estimation program 11D may be provided in a form recorded on an optical disk such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a Blu-ray disk. The estimation program 11D may also be provided in a form recorded on a portable semiconductor memory such as a USB (Universal Serial Bus) memory or a memory card. The non-volatile memory 11B, a CD-ROM, a DVD-ROM, a Blu-ray disk, a USB, and a memory card are examples of non-transitory storage media.

[0100] Furthermore, the control unit 11 may download the estimation program 11D from the external device 4 connected to the communication line through the communication unit 12, and store the downloaded estimation program 11D in the nonvolatile memory 11B of the control unit 11. In this case, the CPU 11A of the control unit 11 reads the estimation program 11D downloaded from the external device 4 from the nonvolatile memory 11B, and executes the process of estimating the radiation source distribution.

[0101] The following are notes related to this embodiment.

[0102] (Appendix 1) A radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution is input to a first learning model that has been machine-learned using first learning data, the first learning model having as input a measured radiation energy spectrum that is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, and an output of a first calculated radiation energy spectrum that is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation where the measured radiation energy spectrum that is the input is measured; a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, is input, and an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution is input to a second learning model that has been machine-learned using second learning data that outputs the radiation source distribution in the situation in which the second calculated radiation energy spectrum was obtained; A process of outputting an output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution as an estimate of the unknown radiation source distribution is performed by a computer. Methods for estimating radiation source distribution.

[0103] (Appendix 2) the correct answer data of the second learning data includes a known radiation source distribution configured by a plurality of types of radiation sources, The second learning model outputs a distribution of the plurality of types of radiation sources estimated from the output of the first learning model. A method for estimating the distribution of radiation sources as described in Appendix 1.

[0104] (Appendix 3) The correct answer data of the second learning data further includes a known shielding object distribution of shielding objects that shield radiation emitted from a radiation source, The second learning model outputs a distribution of the occluding objects estimated from an output of the first learning model. A method for estimating radiation source distribution according to appendix 1 or appendix 2.

[0105] (Appendix 4) The radiation source includes one that emits gamma rays, and the shielding includes uranium. A method for estimating the distribution of radiation sources as described in Appendix 3.

[0106] (Appendix 5) A zeroth learning model in which machine learning is performed using learning data in which synthetic basic calculation data is synthesized, the respective basic calculation data being radiation energy spectra generated by simulation assuming that a plurality of radiation sources included in a known radiation source distribution exist independently, and a radiation energy spectrum generated by simulation assuming that the plurality of radiation sources exist simultaneously is output, is used to input the synthetic basic calculation data to the zeroth learning model, and the output of the zeroth learning model is used as the input of the second learning data. A method for estimating a radiation source distribution according to any one of Supplementary Note 1 to Supplementary Note 4.

[0107] (Appendix 6) The output of the first learning model is normalized and then input to the second learning model. A method for estimating a radiation source distribution according to any one of Supplementary Note 1 to Supplementary Note 5.

[0108] (Appendix 7) A processor is provided. the processor inputs a measured radiation energy spectrum, which is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, into a first learning model that has been machine-learned using first learning data and outputs a first calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation where the measured radiation energy spectrum, which is the input, is measured; and a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, is input, and an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution is input to a second learning model that has been machine-learned using second learning data that outputs the radiation source distribution in the situation in which the second calculated radiation energy spectrum was obtained; An output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution is output as an estimate of the unknown radiation source distribution. A device for estimating radiation source distribution.

[0109] (Appendix 8) A radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution is input to a first learning model that has been machine-learned using first learning data, the first learning model having as input a measured radiation energy spectrum that is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, and an output of a first calculated radiation energy spectrum that is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation where the measured radiation energy spectrum that is the input is measured; a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, is input, and an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution is input to a second learning model that has been machine-learned using second learning data that outputs the radiation source distribution in the situation in which the second calculated radiation energy spectrum was obtained; and causing a computer to execute a process of outputting an output of the second learning model for a radiation energy spectrum measured under the unknown radiation source distribution as an estimate of the unknown radiation source distribution. A program for estimating the distribution of radiation sources.

[0110] (Appendix 9) A non-transitory storage medium storing a program executable by a computer to execute a process for estimating a radiation source distribution, The process of estimating a distribution of radiation sources comprises: a first input step of inputting a radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution into a first learning model that has been machine-learned using first learning data in which a measured radiation energy spectrum, which is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, is input, and a first calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation in which the input measured radiation energy spectrum is measured, is output; a second input step of inputting the output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution into a second learning model that has been machine-learned using second learning data, the second learning model inputting a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, and outputting the radiation source distribution in the situation in which the second calculated radiation energy spectrum was obtained; an output step of outputting an output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution as an estimate of the unknown radiation source distribution; Non-transitory storage media including: [Explanation of symbols]

[0111] 1. First Learning Model 2. Second Learning Model 4 External device 5 Display device 6 Input Devices 7 Sodium iodide (NaI) spectrometer 8 Radiation source Area 9A (where cesium 137 is located) 9B (Cobalt 60) Area 9C (shielded) area 10 Estimation device 11 Control section 11A CPU 11B Non-volatile memory 11C RAM 11D Estimation Program 12 Communications Department 13 Display section 14 Reception Department 15 I / F section 16 Memory section 17 Bus 18 Loading surface 19 Detection surface 20 Shield

Claims

1. a first learning model that has been machine-learned using first learning data in which a measured radiation energy spectrum, which is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, is input, and a first calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation in which the input measured radiation energy spectrum is measured, is output; Input the radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution, A second learning model is obtained by machine learning using second learning data, which is an input of a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, and an output of a radiation source distribution in a situation where the second calculated radiation energy spectrum is obtained. An output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution is input; The output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution is and outputting the estimated value of the unknown radiation source distribution as the estimated value of the unknown radiation source distribution. Methods for estimating radiation source distribution.

2. the correct answer data of the second learning data includes a known radiation source distribution configured by a plurality of types of radiation sources, The second learning model outputs a distribution of the plurality of types of radiation sources estimated from the output of the first learning model. The method for estimating a radiation source distribution according to claim 1 .

3. The correct answer data of the second learning data further includes a known shielding object distribution of shielding objects that shield radiation emitted from a radiation source, The second learning model outputs a distribution of the occluding objects estimated from an output of the first learning model. The method for estimating a radiation source distribution according to claim 2 .

4. The radiation source includes one that emits gamma rays, and the shielding includes uranium. The method for estimating a radiation source distribution according to claim 3.

5. A 0th learning model in which machine learning is performed using learning data in which synthetic basic calculation data is synthesized, the respective basic calculation data being radiation energy spectra generated by simulation assuming that a plurality of radiation sources included in a known radiation source distribution exist independently, and a radiation energy spectrum generated by simulation assuming that the plurality of radiation sources exist simultaneously is output, the output of the 0th learning model when the synthetic basic calculation data is input is used as an input of the second learning data. The method for estimating a radiation source distribution according to claim 4.

6. The output of the first learning model is normalized and then input to the second learning model. The method for estimating a radiation source distribution according to any one of claims 1 to 5.

7. A processor is provided. the processor inputs a measured radiation energy spectrum, which is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, into a first learning model that has been machine-learned using first learning data and outputs a first calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation where the input measured radiation energy spectrum is measured; and a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, is input, and an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution is input to a second learning model that has been machine-learned using second learning data, the second calculated radiation energy spectrum being an input, and the radiation source distribution in the situation in which the second calculated radiation energy spectrum was obtained is output; An output of the second learning model for the radiation energy spectrum measured under the unknown radiation source distribution is output as an estimate of the unknown radiation source distribution. A device for estimating radiation source distribution.

8. a radiation energy spectrum measured using a radiation detector under an unknown radiation source distribution is input to a first learning model that has been machine-learned using first learning data, the first learning model having as input a measured radiation energy spectrum that is a radiation energy spectrum measured using a radiation detector for a known radiation source distribution, and an output of a first calculated radiation energy spectrum that is a radiation energy spectrum generated by simulation from a radiation source distribution in a situation where the input measured radiation energy spectrum is measured; a second calculated radiation energy spectrum, which is a radiation energy spectrum generated by simulation from a known radiation source distribution, is input, and an output of the first learning model for the radiation energy spectrum measured under the unknown radiation source distribution is input to a second learning model that has been machine-learned using second learning data, the second calculated radiation energy spectrum being an input, and the radiation source distribution in the situation in which the second calculated radiation energy spectrum was obtained is output; and causing a computer to execute a process of outputting an output of the second learning model for a radiation energy spectrum measured under the unknown radiation source distribution as an estimate of the unknown radiation source distribution. A program for estimating the distribution of radiation sources.

Images