JP6889096B2 - Learning model manufacturing method, pollution density calculation method and pollution density calculation device - Google Patents

Description

Embodiments of the present invention relate to a learning model manufacturing method, a pollution density calculation method, and a pollution density calculation device.

When radioactive contamination occurs due to leakage of radioactive material from a facility that handles radioactive material, the contamination density is evaluated to manage the leaked radioactive material, and access is restricted, shielded and removed as necessary. Take measures to reduce the exposure of workers such as dyeing.

However, when the pollution range is wide, performing work such as decontamination itself causes an increase in the exposure of workers. Therefore, it is necessary to make work such as decontamination more efficient and minimize the exposure of workers by formulating a work plan based on an appropriately evaluated surface contamination density distribution.

The surface contamination density distribution is generally calculated by measuring radiation while scanning the surface of a controlled area with a radiation inspection device such as a GM tube detector. However, since it takes time to scan the inspection device over the entire controlled area, a technique for calculating the surface contamination density distribution based on the air dose rate, which can be measured relatively easily, has been proposed. ..

For example, a technique for estimating the surface pollution density using changes in air dose rate data measured at different heights and a function that can be calculated in advance by assuming the shape of the pollution source (radioactive source). It has been known. According to this technique, the surface contamination density can be estimated very easily because the measured dose data is only input to the mathematical formula calculated in advance.

JP 2015-001500

However, in the above-mentioned technique, not only the radiation directly incident from the radiation source but also the air dose rate affected by the radiation scattered or transmitted through the structure such as a wall or the ceiling is measured as the measured data. Therefore, if the influence of radiation other than the radiation directly incident from the radiation source is not taken into consideration, there is a problem that a regression equation (response function) for obtaining a surface contamination density distribution satisfying the required accuracy cannot be obtained.

On the other hand, if the response function is created in consideration of the influence of radiation other than the radiation directly incident from the radiation source, a more accurate surface contamination density distribution can be obtained, but the structure information around the measurement location can be obtained. It is necessary to reflect and consider the influence of radiation scattered or transmitted through the structure. Therefore, there is a problem that more time is required to obtain a more accurate response function in consideration of the influence of radiation other than the radiation directly incident from the radiation source.

When considering the application of the contamination density calculation device in various places, the position, material, size, shape, density, etc. of the structure around the measurement place differ depending on the individual measurement place, so it has sufficient accuracy. In order to obtain the surface contamination density distribution, it is necessary to create a response function according to each measurement environment.

However, creating a response function according to each measurement environment imposes an excessive burden on the creation of the response function and consumes a lot of time, which significantly reduces the measurement efficiency of the surface contamination density distribution. There is a problem that it is not realistic.

In this way, there is an antinomy relationship between the measurement accuracy and measurement efficiency of the surface contamination density distribution, and if you want to obtain the surface contamination density distribution with higher accuracy, it takes more time to create the response function. Become.

On the other hand, a method of obtaining a response function that satisfies the required accuracy at various measurement locations by machine learning a learning model containing learning data showing the relationship between pollution density and air dose is also conceivable, but this method is adopted. Even so, the antinomy relationship is established, and the problem remains.

The method of machine learning the above learning model is to set pollution densities in various places in the space on a simulation, calculate the air dose distribution obtained at that time, and accumulate the relationship between the pollution density and the air dose. Is a method of obtaining a response function that satisfies the required accuracy by machine learning the learning model. However, in order to improve the estimation accuracy for a complicated pollution pattern, the learning accuracy should be increased. That is, it is necessary to simulate the pattern of contamination density and air dose distribution for various locations within the evaluation area.

However, for example, in a general simulation method in which a space is divided into voxel structures and the contamination density and air dose are calculated for each voxel, the space is divided into voxel structures divided into 100 in each direction of depth, width and height. When setting an exhaustive pollution pattern (divided into 1,000,000 voxels), even assuming a pollution pattern with one voxel (single voxel pollution), the simulation required to manufacture a learning model The number of patterns is a huge number of 1,000,000 patterns.

Further, in order to further improve the learning accuracy, it is preferable to learn not only when the contaminated part is a single voxel but also a contamination pattern over a plurality of voxels, but when learning a contamination pattern over a plurality of voxels, a learning model The number of simulation patterns required to manufacture is explosively increased from the number of patterns of single voxel contamination.

Considering the general calculation speed of the conventional simulation, the calculation time required for one pattern simulation is estimated to be at least several minutes, and at most several hours. Based on this estimation, if the average calculation time required for one pattern simulation is 1 hour (= 1/24 days = 1/8760 years), even if it is a contamination pattern of single voxel contamination, it will take 100 to complete the calculation. It will take more than a year, and it is not realistic to manufacture a learning model that sets an exhaustive pollution pattern.

On the other hand, when a method of randomly selecting an appropriate voxel position and setting a pollution source at the selected voxel position and simulating it so that a learning model can be manufactured in a realistically feasible time is adopted, the learning pattern Since the absolute number is expected to decrease and the accuracy of the response function obtained with the decrease is expected to decrease, there is a high possibility that the response function satisfying the required accuracy cannot be obtained.

The present invention has been made in view of the above circumstances, and is capable of producing learning data capable of accurately estimating (calculating) the pollution density within a realistic time, and thus manufacturing a learning model for estimating the pollution density. It is an object of the present invention to provide a learning model manufacturing method, a pollution density calculation method, and a pollution density calculation device that achieve both efficiency and estimation accuracy of pollution density.

The learning model manufacturing method according to the embodiment of the present invention contaminates at least one coordinate position selected from all the coordinate positions included in the analysis target range set in the three-dimensional space in order to solve the above-mentioned problems. The step of setting the probability and the evaluation area set inside the analysis target range and representing the position information of the range in which the pollution source is estimated are selected when setting the pollution probability for the evaluation area based on the position information. Whether the step of calculating the dose contribution indicating the degree of influence on the air dose from at least one coordinate position and at least the dose contribution among the contamination probability and the dose contribution satisfy the set conditions. When it is determined in the learning model manufacturing necessity determination step that determines the necessity of manufacturing the learning model and the learning model manufacturing necessity determination step that the learning model is manufactured according to the determination result of whether or not it is determined. Of the pollution probability and the dose contribution compared in making the decision, at least the pollution model that sets the position, size, and intensity of the pollution source, the nuclei that give the dose contribution, and the pollution model are used as input data. The training model acquisition step of executing the above simulation to acquire the air dose distribution in the evaluation area as output data, associating the acquired output data with the input data, and acquiring the associated input data and output data as the training model. It is characterized by having.

In the pollution density calculation method according to the embodiment of the present invention, in order to solve the above-mentioned problems, a step of machine learning the learning model manufactured by the learning model manufacturing method to obtain a learned model and the learned model are used. It is characterized by comprising a step of inputting the measured air dose to obtain the position and intensity of the pollution source.

The pollution density calculation device according to the embodiment of the present invention is set to at least one coordinate position selected from all the coordinate positions included in the analysis target range set in the three-dimensional space in order to solve the above-mentioned problems. Evaluation area that is set inside the analysis target range and represents the position information of the range in which the pollution source is estimated. At least one location selected when setting the pollution probability for the evaluation area based on the position information. It is necessary to manufacture a learning model according to the judgment result of determining whether or not at least the dose contribution degree, which indicates the degree of influence on the air dose from the coordinate position of, satisfies the set condition. When the learning model manufacturing necessity determining means and the learning model manufacturing necessity determining means decide to manufacture the learning model, of the contamination probability and the dose contribution degree compared in making the decision. A simulation using a pollution model in which the position, size, and intensity of the pollution source that gives at least the dose contribution is set as input data is executed to acquire the air dose distribution in the evaluation area as output data, and the acquired output data is obtained. A learning model manufacturing means that associates with input data and manufactures the associated input data and output data as the learning model, and the learning model derived by machine learning the learning model manufactured by the learning model manufacturing means. A pollutant source estimation means for estimating the position, size, and intensity of the pollutant source, which is the measured air dose distribution, using the relationship between the air dose distribution and the position, size, and intensity of the pollutant source and the measured air dose distribution. It is characterized by having.

According to the embodiment of the present invention, it is possible to achieve both the manufacturing efficiency of the learning model for estimating the pollution density and the estimation accuracy of the pollution density.

The schematic diagram which shows roughly the structure of the pollution density calculation apparatus which concerns on embodiment. The perspective view which showed the example of the analysis target to which the pollution density calculation method which concerns on embodiment apply. FIG. 2 is a projection drawing on a plane of the analysis target when the analysis target exemplified in FIG. 2 is divided into 20 × 10 × 10 boxels, and (A) is a projection drawing (plan view) on the xy plane. (B) is a projection drawing on the xz plane (front view). The processing flow diagram which shows the flow of the processing step of the 1st learning model manufacturing processing procedure which is an example of the learning model manufacturing method which concerns on embodiment. It is a projection drawing corresponding to FIG. 3, and (A) and (B) are a plan view and a front view of an analysis target when the influence of a pollution source located at a distance from the evaluation area beyond the broken line BL is ignored, respectively. It is a projection drawing corresponding to FIG. 3, and (A) and (B) are plan views of analysis targets when the influence of a pollution source located at a distance from the evaluation area beyond the broken line BL is replaced with a pseudo-contamination model, respectively. And front view. The processing flow diagram which shows the flow of the processing step of the 2nd learning model manufacturing processing procedure which is an example of the learning model manufacturing method which concerns on embodiment. The processing flow diagram which shows the flow of the processing step of the pollution density calculation processing procedure which is an example of the pollution density calculation method which concerns on embodiment.

Hereinafter, the learning model manufacturing method, the pollution density calculation method, and the pollution density calculation device according to the embodiment of the present invention will be described with reference to the accompanying drawings.

(Device configuration)
FIG. 1 is a schematic view schematically showing the configuration of the pollution density calculation device 20 which is an example of the pollution density calculation device according to the embodiment.

The pollution density calculation device 20 obtains a regression equation (response function) for associating the air dose with the nuclide, position, size (range) and intensity (contamination density) of the radiation source (contamination source), and obtains the regression equation and the obtained regression equation. It is a device that calculates the nuclide, position, size, and intensity that cause contamination using the input air dose. The regression equation used in the pollution density calculation device 20 is, for example, a response function of a machine-learned learning model obtained by machine-learning a learning model manufactured by using the learning model manufacturing method according to the embodiment, and is machine-learning. It is required by obtaining a completed learning model.

The contamination density calculation device 20 includes, for example, an input means 21, an output means 22, a pollution probability setting means 23, a dose contribution calculation means 24, a learning model manufacturing necessity determination means 25, and a learning model manufacturing means 26. A machine learning means 27, a pollution source estimation means 28, a pseudo pollution model setting means 29, an accuracy determination means 31, a storage means 32, and a control means 33 are provided.

Here, the structure information 35, the contamination probability information 36, the dose contribution information 37, the learning model 38, the learned model 39, and the evaluation model 41 shown inside the storage means 32 are read (read) or written (write). This is information held in the storage means 32 that provides a storage area that can be used.

The input means 21 is realized by, for example, an input device connected to the computer via an interface or an input means such as a keyboard or a mouse included in the computer itself. The input means 21 accepts the input of information and gives the received information to the control means 33. The information input from the input means 21 includes, for example, an execution request given to a processing means such as the contamination probability setting means 23.

The output means 22 is realized by, for example, a display device connected to the computer via an interface, a display means such as a display provided by the computer itself, a printing means such as a printer connected to the computer via an interface, and the like. When the output means 22 receives the display request, the output means 22 displays the content corresponding to the display request. Further, when the output means 22 receives the print request, the output means 22 prints out the contents corresponding to the print request.

The pollution probability setting means 23 has a function of acquiring information that affects radiation attenuation and shielding among information about the structure by referring to the structure information 35, and is set in the three-dimensional space and included in the analysis target 1. It has a function to obtain the contamination probability at the coordinate position.

The coordinates in the three-dimensional space are represented by a coordinate system that can identify one point in the three-dimensional space. In the simplest case, for example, the x-axis direction representing the horizontal (width) direction and the y representing the vertical (depth) direction. It is represented by a three-dimensional Cartesian coordinate system (FIG. 2) consisting of three orthogonal axes in the z-axis direction representing the axial direction and the height direction.

The contamination probability setting means 23 acquires information representing the characteristics of the structure such as position, material, size, shape and density, which affect the attenuation and shielding of radiation, from the structure information 35, while at least The contamination probability at the specified coordinate position is obtained by referring to the contamination probability information 36 in which the type and shape of the substance and the contamination probability are associated with each other.

The pollution probability setting means 23 can consider several methods for setting the pollution probability. For example, as the first setting method, a method utilizing the fact that the adhesion of pollutants cannot occur in the airy region where the structure does not exist or the closed space surrounded by the structure, and the pollution probability becomes 0%. Is. More specifically, the area where the pollution probability is not 0% (more than 0% and 100% or less) other than the aerial area where the structure does not exist and the closed space surrounded by the structure is specified and specified. This is a method of obtaining the contamination probability at a desired position in the analysis target by acquiring the contamination probability associated with the type of the substance in the region or the like.

The second setting method is a method that utilizes the fact that pollutants may adhere to the surface of the structure in contact with the outside air (atmosphere) in the region where the structure exists. More specifically, the surface of the structure is specified based on the structure information 35, and the contamination probability associated with the type of the substance constituting the structure for which the surface is specified is acquired, and the like is included in the analysis target. This is a method for obtaining the contamination probability at a desired position.

When the pollution probability setting means 23 obtains the pollution probability at the designated coordinate position, the pollution probability setting means 23 sends the pollution probability at the obtained designated coordinate position to the learning model manufacturing necessity determination means 25.

The dose contribution calculation means 24 refers to the dose contribution information 37, and refers to the analysis target 1 with respect to the range included in the analysis target 1 and the range for estimating the pollution source (hereinafter referred to as “evaluation area”). Using the function to acquire an evaluation index such as dose contribution, which indicates the degree of influence on the air dose from the coordinate position included in, and the acquired evaluation index, the evaluation area is set to the analysis target 1. It has a function of quantifying the magnitude of the influence on the air dose from the included coordinate positions, that is, a function of obtaining the evaluation value of the acquired evaluation index.

As an evaluation index for evaluating the degree of influence on the air dose from the coordinate position included in the analysis target 1, for example, the dose contribution degree and the like are available. The information of the evaluation standard using the dose contribution is included in the dose contribution information 37, and the dose contribution calculation means 24 calculates the dose contribution with reference to the dose contribution information 37. When the dose contribution calculation means 24 calculates the dose contribution, the calculation result is sent to the learning model manufacturing necessity determination means 25.

The learning model manufacturing necessity determination means 25 includes a nuclide as a pollution source, a model in which the position, size and intensity of the pollution source are set (hereinafter referred to as a “contamination model”), and an air dose distribution in an evaluation area for the pollution model. It has a function of determining whether or not to manufacture the learning model 38 associated with, that is, whether or not the learning model 38 needs to be manufactured.

The learning model manufacturing necessity determination means 25 determines the learning model manufacturing necessity determination condition in which at least the dose contribution of the contamination probability and the dose contribution is set (hereinafter, simply referred to as “manufacturing necessity determination condition”). By determining whether or not the condition is satisfied, it is determined whether or not the learning model 38 is manufactured. Here, of the contamination probability and the dose contribution, at least the dose contribution is defined so that the contamination probability can be regarded as uniform within the analysis target range, such as when the structure does not exist within the analysis target range. In this case, it is sufficient to compare the dose contributions.

As a manufacturing necessity determination condition, for example, the product of the contamination probability and the dose contribution is determined in consideration of the fact that the higher the contamination probability and the higher the dose contribution, the greater the need to manufacture the learning model 38. It is conceivable that the threshold value v1 (v1 is a positive number) will be exceeded. That is, the learning model manufacturing necessity determination means 25 determines that it is necessary to manufacture the learning model 38 when the necessity of manufacturing the learning model 38 has reached a predetermined level, and has reached the predetermined level. If not, it is determined that the manufacturing of the learning model 38 is unnecessary.

When the learning model manufacturing necessity determination means 25 determines the necessity of manufacturing the learning model 38, the learning model manufacturing necessity determination means 25 sends the determination result to the learning model manufacturing means 26.

The learning model manufacturing means 26 has, for example, a function of simulating and obtaining an air dose distribution in an evaluation area for a pollution model, and a function of associating the simulated air dose distribution with a pollution model to manufacture a learning model 38. Has.

The learning model manufacturing means 26 simulates and obtains the air dose distribution in the evaluation area for the pollution model for the pollution model determined to manufacture the learning model 38 based on the judgment result of the necessity of manufacturing the learning model 38. And the obtained air dose distribution and the pollution model are associated with each other to manufacture the learning model 38.

The learning model manufacturing means 26 sets another different pollution model from the simulation results already obtained, obtains the air dose distribution of the evaluation area for the set pollution model, and obtains the air dose distribution and the pollution model. May have a function of manufacturing a new learning model 38 by associating the above.

For the production of the new learning model 38, for example, using the simulation results already obtained, the intensity of the pollution source and the air dose distribution are multiplied by n (n is a positive number) to obtain the new pollution model and the air dose distribution in the evaluation area. This is done by translating the position of the pollution source and the air dose distribution based on the simulation results already obtained to obtain a new pollution model and the air dose distribution in the evaluation area.

However, the manufacturing method of the new learning model 38 illustrated above is based on the premise that the effects of radiation scattering and shielding are the same. Therefore, it is preferable that the method for producing the new learning model 38 illustrated above is applied when the average atomic numbers and densities of the substances around the contamination setting area are similar.

The machine learning means 27 has a function of making the learning model 38 machine-learn, and makes the learning model 38 machine-learn to obtain the learned model 39. Applicable machine learning includes, for example, simple multivariate analysis and deep learning.

The timing at which the machine learning means 27 ends machine learning is determined depending on whether or not the set determination criteria are satisfied. For example, in the pollution density calculation device 20, when the accuracy determination means 31 determines that the accuracy of the contamination density derived from the learned model 39 has reached the required level, it is determined that the set determination criteria are satisfied.

When it is determined that the learning model 38 satisfies the judgment criteria set by the trained model 39 obtained by machine learning, the trained model 39 can, in principle, estimate (calculate) the contamination density with an accuracy that satisfies the required accuracy. Regression equation (response function) is provided.

The pollution source estimation means 28 uses the measured air dose distribution and the regression equation that can estimate the pollution density derived from the trained model 39, and the pollution density that is the measured air dose distribution, that is, the nuclide that is the pollution source, the pollution source. Estimate the position, size and strength of. The intensity of the pollution source is obtained by multiplying by n (n is a positive number) from the proportional relationship with the air dose.

Instead of the pollution source included in the set range, the pseudo-contamination model setting means 29 considers that the influence of the pollution source on the air dose in the evaluation area is substantially equivalent to the pseudo-contamination model (hereinafter, “pseudo-contamination”). It has a function of determining whether or not to set a "model") and a function of setting a pseudo-contamination model. The pseudo-contamination model is a pollution source whose nuclides are known to be the source of pollution and the size and intensity of the pollution source are known.

Assuming that a pseudo-pollution model is set, it is not necessary to set the pollution model individually to manufacture the learning model 38, but without setting the pollution model at all (the influence of the pollution source). When there is a concern that the error will be slightly larger if the learning model 38 is not manufactured (completely ignored), that is, the relative level of the need to manufacture the learning model 38 is low, medium, or high. This is the case where the level of necessity for manufacturing the learning model 38 is at the level corresponding to "medium" when classified into three stages.

For example, as one of the indexes for determining whether or not to set a pseudo-contamination model, the product of the contamination probability and the dose contribution is applied, and two different thresholds v1 and v2 (v2 is positive satisfying v2 <v1). When the number) is set, it can be determined that a pseudo-contamination model needs to be set when the product of the contamination probability and the dose contribution is equal to or less than the threshold value v1 and exceeds the threshold value v2. An example of determining whether or not to set this pseudo-contamination model can be performed together with the determination of manufacturing necessity of the learning model 38.

The function of determining whether or not to set the pseudo-contamination model is combined with the function of the learning model manufacturing necessity determining means 25 determining the manufacturing necessity of the learning model 38 instead of the pseudo-contamination model setting means 29. You may have.

The accuracy determination means 31 has a function of determining whether or not the pollution density derived from the trained model 39 has reached the level required by the accuracy of the regression equation that can be estimated. Whether or not the accuracy is at the required level is based on an accuracy determination index that defines the relationship between the pollution density value of a known pollution source given to the evaluation model 41 and the pollution density estimate derived from the trained model 39. judge.

The accuracy determination means 31 first gives the air dose derived by the evaluation model 41 when a known pollution source (nuclide, position, size and intensity) is set to the trained model 39 for determining the accuracy, and becomes a pollution source. Determine the nuclide, the location, size and intensity of the pollution source. Subsequently, the accuracy determination means 31 determines the obtained nuclide as the pollution source, the position, size and intensity (contamination density) of the pollution source as the known nuclide as the pollution source, and the position, size and intensity (contamination density) of the pollution source. Is applied to the accuracy determination index to determine whether or not the required accuracy is satisfied.

As the accuracy determination index, for example, an index for evaluating an error such as the sum of squares of the difference (error) between the estimated pollution density derived from the trained model 39 and the known pollution density value given to the evaluation model 41 is used. .. For example, when the sum of squares of the above errors is applied as an accuracy determination index, the accuracy determination means 31 requests if the value of the sum of squares of the errors is smaller than the predetermined value set as the threshold value for the accuracy determination. Judge that the accuracy is satisfied.

In addition, the accuracy judgment index indicates the amount of change and the change speed of the difference from the past accuracy judgment result obtained by performing the accuracy judgment using the same evaluation model 41 as the current accuracy judgment result obtained by using the same accuracy judgment index. Can also be used as. Since the amount of change and the rate of change are considered to decrease as the accuracy increases, it is determined that the required accuracy is satisfied when the amount of change or the rate of change falls below a predetermined amount of change or rate of change set as a threshold value for determining accuracy.

The accuracy determination means 31 is, for example, the sum of squares in the case where the accuracy determination index is the sum of squares of the difference between the estimated value of the contamination density obtained from the trained model 39 and the known contamination density value given to the evaluation model 41. If the value of is less than or equal to the allowable value to be set, it is determined that the required accuracy is satisfied, while if it exceeds the allowable value, it is determined that the required accuracy is not satisfied.

The storage means 32 has a storage circuit that provides a storage area capable of reading (reading) and writing (writing) information, and has a function of holding the information in the storage area.

The storage means 32 includes a pollution probability setting means 23, a dose contribution calculation means 24, a learning model manufacturing necessity determination means 25, a learning model manufacturing means 26, a machine learning means 27, a pollution source estimation means 28, and a pseudo pollution model setting means 29. The accuracy determination means 31 and the control means 33 are provided with a storage area in which information necessary for executing each process can be read (read) and written (written).

That is, the storage means 32 realizes a function incorporated in the pollution density calculation device 20, a ROM (Read Only Memory) that stores a program that controls the function, a RAM (Random Access Memory) that is an execution area of the program, and the like. Is conceptually included.

The control means 33 is a means for controlling the processing of the entire device of the pollution density calculation device 20, and is an input means 21, an output means 22, a pollution probability setting means 23, a dose contribution calculation means 24, and a learning model manufacturing necessity determination means. 25, the learning model manufacturing means 26, the machine learning means 27, the pollution source estimation means 28, the pseudo pollution model setting means 29, the accuracy determination means 31, and the storage means 32 have a function of exchanging information with each other and controlling them.

The control means 33 uses the above functions to input means 21, output means 22, pollution probability setting means 23, dose contribution calculation means 24, learning model manufacturing necessity determination means 25, learning model manufacturing means 26, and machine learning means. 27, a learning model manufacturing process procedure (FIGS. 4 and 7) and a contamination density calculation process procedure (FIG. 8) described later by controlling the pollution source estimation means 28, the pseudo-pollution model setting means 29, the accuracy determination means 31 and the storage means 32. Execute the processing procedure such as.

When the control means 33 receives the information from the input means 21, the control means 33 receives the output means 22, the contamination probability setting means 23, the dose contribution calculation means 24, and the learning model manufacturing necessity determination according to the type of information received by the input means 21. A request is given based on the information received by any of the means 25, the learning model manufacturing means 26, the machine learning means 27, the pollution source estimation means 28, the pseudo pollution model setting means 29, and the storage unit 31.

When there is a display request for information, the control means 33 receives the information for which the display request has been made from the processing means 21, 23 to 29 related to the information for which the display request has been made, and the received information is an example of the output means 22. Give a display command to display to a certain display means. The information for which the display request has been made is displayed on the display means that has received the display command.

The structure information 35 includes information on the structure that affects the attenuation and shielding of radiation. Information that affects radiation attenuation and shielding includes, for example, information that represents the characteristics of the structure such as the position, material, size, shape, and density of the structure.

In the present embodiment, at least information on the position, size, and shape of the structure is included as information representing the characteristic items of the structure from the viewpoint of making it possible to specify the range occupied by the structure. Further, information on the material of the structure may be further included from the viewpoint of making it possible to automatically set the contamination probability.

Contamination probability information 36 includes at least information on contamination probabilities associated with the type and shape of the substance. The contamination probability information 36 is referred to, for example, when the contamination probability setting means 23 sets the contamination probability at the position where the structure exists.

The pollution probability information 36 does not need to individually include the pollution probability information corresponding to the type and shape of each substance. For example, the pollution probability information corresponding to a typical feature and the pollution. It may be an aspect including the information of the correction formula that corrects the probability according to the feature matter. That is, it is only necessary to uniquely identify the contamination probability corresponding to one characteristic item.

The dose contribution information 37 is information on an evaluation standard using the dose contribution, that is, information for obtaining the dose contribution I. The dose contribution I contains at least one of a first parameter corresponding to the distance from the radiation source and a second parameter corresponding to the effect of radiation attenuation or shielding, eg, from the distance from the source and from the source. It can be expressed by an arithmetic expression that takes into account the effects of attenuation and shielding of the emitted radiation.

The formula for obtaining the dose contribution I is, for example, using the set distance L between the pollution source and the evaluation area, the average line attenuation coefficient μ of the structure between the pollution source and the evaluation area, and the thickness d. It can be expressed as the following equation (1).

[Number 1]
I = exp (−μd) / L ² … (1)
In the above formula (1), ^{the term 1 / L 2 is} a term that takes into account the influence of the distance L from the pollution source, and the term exp (−μd) is the term for attenuation and shielding of radiation emitted from the pollution source. It is a term that takes into account the influence.

The evaluation index for evaluating the degree of influence on the air dose is not necessarily limited to the above formula (1), but is an index that reflects the distance from the radiation source (contamination source) and the influence of radiation attenuation or shielding. It should be.

The learning model 38 is set as the air dose distribution obtained by executing a simulation for the pollution source pattern in which the pollution source is set inside the analysis target range and which has a large influence on the accuracy of estimating the pollution density. Data associated with pollution sources. The trained model 39 is a machine-learned version of the learning model 38.

The evaluation model 41 is an analysis model for obtaining a regression equation capable of estimating the contamination density, and the analysis accuracy is known. The evaluation model 41 provides a determination criterion when determining whether or not the accuracy of the regression equation that can estimate the contamination density derived from the trained model 39 is at the required level.

The pollution density calculation device 20 is not limited to the configuration illustrated in FIG. 1, and includes at least a learning model manufacturing necessity determination means 25, a learning model manufacturing means 26, and a pollution source estimation means 28. For example, other components such as the machine learning unit 27 may be provided in an external device connected via a communication network.

For example, when the machine learning unit 27 is mounted on an external device, the trained model 39 is received by the machine learning unit 27 via a communication network and stored in the storage means 32.

Further, a part of the component includes not only a part of the processing means but also a part of the processing means. For example, the pollution density calculation device 20 may be constructed by mounting an element having a partial function of the processing means, such as a function of executing the simulation of the learning model manufacturing means 26, on an external computer (server).

When the pollution density calculation device 20 is premised on manufacturing up to the learning model 38, the pollution density calculation device 20 is used as a learning model manufacturing device by appropriately omitting information and functions applied only to the subsequent calculation of the pollution density. May be constructed as.

Further, in the pollution density calculation device 20, the pollution density calculation device 20 can be constructed by appropriately omitting functions that are not essential for manufacturing the learning model 38.

For example, when calculating the contamination density for an analysis target whose contamination probability can be regarded as uniform, the contamination density calculation device 20 does not necessarily have to be provided with the contamination probability setting means 23, and the contamination probability setting means 23. A pollution density calculation device 20 that does not include the 23 may be constructed.

For example, when the function of setting the pseudo-contamination model is not required, the contamination density calculation device 20 that does not include the pseudo-contamination model setting means 29 may be constructed.

For example, when the function of determining the accuracy of the trained model 39 as an analysis model for estimating the position, size, and intensity of the contamination source is unnecessary, the contamination density calculation device 20 not provided with the accuracy determination means 31 may be constructed. ..

The pollution density calculation device 20 shown in FIG. 1 is an example in which the storage means 32 holds the pollution probability information 36, but it is not always necessary for the storage means 32 to hold the pollution probability information 36. .. For example, when the user individually sets the contamination probability in consideration of the position, size, shape, and material of the structure grasped based on the structure information 35, the contamination probability without using the contamination probability information 36 is used. When setting, the contamination probability information 36 may be omitted.

In the pollution density calculation device 20 described above, for example, the processing means 21 to 29, 31 to 33 to be incorporated may be constructed by using hardware resources, or a processor such as a CPU (Central Processing Unit) and a ROM (Read-) may be constructed. As a hardware resource equipped with a storage device having a storage circuit such as Only Memory) and RAM (Random-Access Memory), and as a software resource that realizes the functions of the processing means 21 to 29, 31 to 33 incorporated in the computer. You may build it in cooperation with the program of.

(Analysis object)
Subsequently, the analysis target to which the present embodiment is applied will be described.

FIG. 2 is a perspective view schematically showing an example of the analysis target 1 to which the present embodiment is applied.
The x-axis, y-axis, and z-axis shown in FIG. 2 are axes of the three-dimensional Cartesian coordinate system, respectively, and are coordinate axes extending in the horizontal (width) direction, the vertical (depth) direction, and the height direction, respectively. Is.

The analysis target 1 is set in the three-dimensional space in which the coordinate position is defined. That is, the range of the analysis target 1 is in the three-dimensional space in which the coordinate position can be recognized. Inside the analysis target 1, for example, the wall 2, the floor 3, and the structures st1 to st4 installed on the wall 2 and the floor 3 are present, and the pollution source 5 is attached or accumulated.

The evaluation area 6 is a range in which the contamination density calculation device 20 calculates the contamination density, and is set inside the analysis target 1.

The voxel 7 is a three-dimensional division of the analysis target 1, and is divided into a rectangular parallelepiped shape in the three-dimensional coordinate system illustrated in FIG. The number of divisions of the voxels 7, that is, the size (volume) of each voxel 7 is set in consideration of the required accuracy and the amount of calculation. The minimum size of the voxel 7 is the minimum length recognized in each axis.
The voxel 7 serves as a reference when setting the position and size of the pollution source with respect to the analysis target 1. That is, in the three-dimensional coordinate system pollution density calculation device 20, the pollution source is set in units of 7 voxels.

The distance between the pollution source and the evaluation area 6 is the distance between the coordinate points in the voxel 7 in which the pollution source is set and the coordinate points in the voxel 7 included in the evaluation area 6, for example, the center point of the voxel 7 in which the pollution source is set. And the distance from the center point of the evaluation area 6 (the voxel 7 located in the center of), and the distance between the coordinate point (recent contact point) closest to the evaluation area 6 in the voxel 7 where the pollution source is set and the center point of the evaluation area 6. , There is a recent contact point with the evaluation area 6 in the voxel 7 in which the pollution source is set, and a coordinate point (recent contact point) closest to the pollution source in the evaluation area 6.

The analysis target 1 illustrated in FIG. 2 is recognized by the contamination density calculation device 20 as, for example, an aggregate of voxels 7 divided into a plurality of voxels 7 or one voxel 7.

FIG. 3 shows the analysis target 1 when the analysis target 1 illustrated in FIG. 2 is divided into a total of 2000 boxels 7 of 20 (x-axis direction) × 10 (y-axis direction) × 10 (z-axis direction). It is a projection drawing on a plane, FIG. 3A is a projection drawing on an xy plane (plan view), and FIG. 3B is a projection drawing on an xy plane (front view).

From the viewpoint of simplifying the explanation, the information representing the characteristic items of the structures of the structures st1 to st4 exemplified in FIG. 3 are all the same, and the contamination probability on the structure is, for example, a predetermined value p (0 <. A case where p ≦ 1) can be regarded as uniform and the dose contribution degree I is expressed by the above equation (1) will be described as an example.

In the analysis target 1 exemplified in FIG. 3 (FIGS. 3 (A) and 3 (B)), the evaluation area 6 is 8 (x-axis direction) × 6 (y-axis direction) × 10 (in the right corner of the figure). It is represented by a total of 480 voxels 7 (in the z-axis direction).

The structure st1 is represented by a total of 12 voxels 7 of 2 (x-axis direction) × 2 (y-axis direction) × 3 (z-axis direction), all of which exist inside the evaluation area 6.
The structure st2 is represented by a total of eight voxels 7 of 2 (x-axis direction) × 2 (y-axis direction) × 2 (z-axis direction), all of which exist inside the evaluation area 6.
The structure st3 is represented by a total of 16 voxels 7 of 2 (x-axis direction) × 2 (y-axis direction) × 4 (z-axis direction), all of which are outside the evaluation area 6 and with respect to the broken line BL. It exists on the side far from the evaluation area 6.
The structure st4 is represented by a total of 20 voxels 7 of 20 (x-axis direction) × 1 (y-axis direction) × 1 (z-axis direction), and 8 of the 20 voxels 7 are evaluated. Inside the area 6, seven are outside the evaluation area 6 and closer to the evaluation area 6 with respect to the broken line BL, and five are outside the evaluation area 6 and farther from the evaluation area 6 with respect to the broken line BL. To do.

The structure st4 does not occupy all of the voxels 7 in the y-axis direction and the z-axis direction, but the case of partially occupying the voxels 7 is the same as the case of occupying all of the voxels 7. Count.

(Learning model manufacturing method and pollution density calculation method)
Next, the learning model manufacturing method and the pollution density calculation method according to the embodiment will be described.

The learning model manufacturing method according to the embodiment determines whether or not the learning model 38 (FIG. 1) to be machine-learned is a pollution model that needs to be manufactured, and the learning model 38 needs to be manufactured. This is a method of manufacturing the learning model 38 by limiting the above to the target. The learning model manufacturing method according to the embodiment can be carried out using, for example, the contamination density calculation device 20.

In the pollution density calculation method according to the embodiment, the air dose is related to the position and intensity (contamination density) of the radiation source (contamination source) by using the learned model 39 (FIG. 1) obtained by machine learning the learning model 38. This is a method of calculating the position and intensity of the pollution source by obtaining the regression equation (response function) to be attached and using the obtained regression equation and the input air dose. The pollution density calculation method according to the embodiment can be carried out using, for example, the pollution density calculation device 20.

<First learning model manufacturing process procedure>
FIG. 4 is an example of the learning model manufacturing method according to the embodiment, and is a processing flow diagram showing the flow of the processing steps of the first learning model manufacturing processing procedure.

The first learning model manufacturing processing procedure described later is a case where the contamination density calculation device 20 (FIG. 1) manufactures the learning model 38 (FIG. 1) for the analysis target 1 exemplified in FIG. This is a description of the case where the product of the contamination probability and the dose contribution is used when determining the level of necessity for manufacturing the learning model 38.

The first learning model manufacturing processing procedure includes, for example, a contamination probability setting step (step S1), a dose contribution calculation step (step S2), a learning model manufacturing necessity determination step (step S3), and a learning model acquisition step. (Step S5), a pseudo-contamination model setting necessity determination step (step S6), and a pseudo-contamination model setting step (step S7) are provided.

In the first learning model manufacturing process procedure (FIG. 4), first, the contamination probability setting step (step S1) is performed. In the contamination probability setting step, the contamination probability setting means 23 (FIG. 1) sets the contamination probability at at least one coordinate position within the analysis target range.

In the contamination probability setting step (step S1: FIG. 4), the contamination probability setting means 23 (FIG. 1) refers to the structure information 35 (FIG. 1), and among the information related to the structures st1 to st4 (FIG. 3), While acquiring information that affects the attenuation and shielding of radiation by structures st1 to st4, such as information that represents the characteristics of the structure such as the position, material, size, shape, and density of structures st1 to st4, the probability of contamination With reference to the information 36 (FIG. 1), the contamination probability of the box cell 7 (FIG. 2) located at the specified coordinate position is obtained.

In this case, referring to the structure information 35 and the contamination probability information 36, the contamination probability of the voxels 7 representing the structures st1 to st4 is p, and the contamination probability of the other voxels 7 is 0 (zero). Here, when the contamination probability is set by using the first setting method described above, the contamination probability setting means 23 has a total of 56 voxels 7 representing the structures st1 to st4 out of the 2000 voxels 7 in the analysis target 1. The contamination probability p is set in the voxels 7 of the above, and the contamination probability 0 is set in the remaining voxels 7.

When the contamination probability setting step is completed, the dose contribution calculation step (step S2: FIG. 4) is subsequently performed. In the dose contribution calculation step, the dose contribution calculation means 24 (FIG. 1) refers to the dose contribution information 37 (FIG. 1), and the dose contribution I from the coordinate position selected when setting the contamination probability. Is calculated.

^{In this case, the magnitude of the dose contribution I is influenced by the term 1 / L 2} included in the right-hand side term of the above equation (1), and when L is sufficiently large, the dose contribution I is approximated to 0. Small enough to be possible. Therefore, in FIGS. 3, 5, and 6 in which the analysis target 1 is exemplified, the distance L from the evaluation area 6 increases and the dose contribution I decreases toward the left side.

When the dose contribution calculation step is completed, the learning model manufacturing necessity determination step (step S3: FIG. 4) is subsequently performed. In the learning model manufacturing necessity determination step, the learning model manufacturing necessity determination means 25 determines the level of necessity for manufacturing the learning model 38, and sets the relative level of the necessity to, for example, “high” or “medium”. If the relative level of the need corresponds to "high" in the case of classifying into three stages of "low" and "low", it is decided to manufacture the learning model 38.

Here, when it is decided to manufacture the learning model 38 (YES in step S3), subsequently, the learning model manufacturing means 26 having a simulation function determines to manufacture the learning model 38. A simulation is performed using the pollution model that gives at least the dose contribution among the pollution probability and the dose contribution as input data. As a result of this simulation, the learning model manufacturing means 26 acquires the air dose distribution in the evaluation area as output data (step S4: FIG. 4).

Subsequently, a learning model acquisition step (step S5: FIG. 4) is performed. In the learning model acquisition step, the learning model manufacturing means 26 acquires the air dose distribution in the evaluation area as the acquired output data and the contamination model as the input data as the learning model 38. That is, the learning model 38 is manufactured.

When the learning model 38 is manufactured, the first learning model manufacturing processing procedure ends all processing steps (END).

On the other hand, when it is determined not to manufacture the learning model 38 in the learning model manufacturing necessity determination step (NO in step S3), the pseudo-contamination model setting necessity determination step is subsequently performed (step S6: FIG. 4).

In the pseudo-contamination model setting necessity determination step, for example, the learning model manufacturing necessity determination means 25 having a function of determining whether or not to set the pseudo pollution model is the learning model 38 as in the learning model manufacturing necessity determination step. Determine the level of need to manufacture and, if the level of need falls under "medium" (though not "high"), decide to set up the required pseudo-contamination model.

Here, when it is determined to set the necessary pseudo-contamination model (YES in step S6), the pseudo-contamination model setting step is subsequently performed (step S7: FIG. 4).

In the pseudo-contamination model setting step, for example, the pseudo-contamination model setting means 29 sets the pseudo-contamination model according to the input for setting the pseudo-contamination model. When the setting of the pseudo-contamination model is completed, the processing flow proceeds from step S7 to step S4, and the processing steps after step S4 are performed.

On the other hand, when it is decided not to manufacture the learning model 38 in the learning model manufacturing necessity determination step, and it is decided not to set the necessary pseudo-contamination model in the subsequent pseudo-contamination model setting necessity determination step (NO in step S3). And NO in step S6), the processing flow proceeds from step S6 to END. That is, the first learning model manufacturing processing procedure ends without manufacturing the learning model 38 and without setting the pseudo-contamination model (END).

In the first learning model manufacturing processing procedure described above, the case where the dose contribution calculation step (step S2) is performed following the contamination probability setting step (step S1) has been described, but the contamination probability setting step and the dose contribution have been described. The degree calculation step may be performed in any order as long as it is before the learning model manufacturing necessity determination step is performed. That is, the dose contribution calculation step may be performed before the contamination probability setting step, or may be performed in parallel.

Further, the first learning model manufacturing process procedure described above is an example including the pseudo-contamination model setting necessity determination step (step S4) and the pseudo-contamination model setting step (step S5), but the pseudo-contamination model setting is required. The rejection step (step S4) and the pseudo-contamination model setting step (step S5) are arbitrary processing steps and may be omitted.

If the pseudo-contamination model setting necessity determination step (step S4) and the pseudo-contamination model setting step (step S5) are omitted, the relative levels of the need to manufacture the learning model 38 are at least “high” and “low”. It suffices if it is classified into two stages.

Here, the relative levels of the need to manufacture the learning model 38 in the learning model manufacturing necessity determination step (step S3) and the pseudo-contamination model setting necessity determination step (step S6) are “high”, “medium”, and Explain the "low" relationship.

FIG. 5 is a projection drawing corresponding to FIG. 3, and the distance L from the evaluation area 6 to the pollution source (voxel 7) is farther than the broken line BL (the broken line BL in the figure) with respect to the analysis target 1 exemplified in FIG. It is a projection drawing which shows the relationship between the analysis object 1 and voxel 7 (7a, 7c) in the case where the influence of the pollution source located in the area on the left side including is completely ignored.

More specifically, FIGS. 5 (A) and 5 (B) are a plan view and a front view of the analysis target 1 in the case where the influence of the pollution source located at the distance from the evaluation area 6 beyond the broken line BL is ignored, respectively. It is a figure.

For example, the distance L from the evaluation area 6 shown in FIG. 5 (FIGS. 5A and 5B) is equal to or greater than the distance L1 from the evaluation area 6 to the broken line BL (including the broken line BL in the figure). In voxels 7c (located to the left), the dose contribution I is small enough and the level of need to manufacture the learning model 38, which is the product of the contamination probability and the dose contribution, is relatively "medium" or less. ("Medium" or "Low"). Conversely, for voxels 7a with a distance of less than L1, the level of need to manufacture the learning model 38 is relatively high.

In this case, the pollution source set in the voxel 7c beyond the broken line BL is excluded from the target for manufacturing the learning model 38, and the more necessary distance L from the evaluation area 6 is the voxel 7a having a distance L1 or less. The learning model 38 is manufactured by narrowing down to the set pollution source.

More specifically, in the analysis target 1 illustrated in FIG. 5, the contamination probability p is set, and the number of voxels 7a in which the influence of the contamination source is not ignored is narrowed down to a total of 35 voxels.

Temporarily, a single voxel analysis is performed on the voxel 7a in which the contamination probability p is set and the influence of the contamination source is not ignored, that is, the analysis target 1 (aggregate of 2000 voxels 7) exemplified in FIG. 5 without narrowing down the contamination model. When implemented, analysis of 2000 patterns is required. On the other hand, the number of voxels 7a that satisfy the manufacturing necessity determination condition of the learning model 38, that is, the pollution model that manufactures the learning model 38 can be reduced to 35 patterns by narrowing down the pollution model.

FIG. 6 is a projection drawing corresponding to FIG. 3, and the distance L from the evaluation area 6 to the pollution source (voxel 7) is farther than the broken line BL (broken line BL in the figure) with respect to the analysis target 1 exemplified in FIG. It is a projection drawing showing the relationship between the analysis target 1 and the voxels 7 (7a, 7e) when the influence of the pollution source located in the area on the left side including) on the evaluation area 6 is replaced by the pseudo-pollution model 8.

More specifically, FIGS. 6 (A) and 6 (B) are analysis targets in the case where the influence of the pollution source whose distance L from the evaluation area 6 is located beyond the broken line BL is replaced with the pseudo-contamination model 8, respectively. It is a plan view and a front view of 1.

The pseudo-contamination model 8 is relatively “medium” in terms of the level of necessity for manufacturing the learning model 38, and an error occurs when the influence of a contamination source having a large distance L from the evaluation area 6 is completely ignored. It is set as necessary, such as when there is a concern that it will grow.

In the example shown in FIG. 6 (FIGS. 6A and 6B), the voxel 7 including the dashed line BL (located immediately to the left of the dashed line BL in the figure) has a distance L from the evaluation area 6. A pseudo-pollution model 8 is set to replace the pollution source located on the side farther from the broken line BL.

The pseudo-contamination model 8 is, for example, a contamination model that simulates uniform surface contamination. In the example shown in FIG. 6, the voxels 7e in which the pollution source assuming the surface pollution is set are arranged in the arrangement of 1 (x-axis direction) × 4 (y-axis direction) × 7 (z-axis direction) and the positions of the broken lines BL. The aggregate of voxels 7e arranged in is the pseudo-contamination model 8. That is, the total number of voxels 7e constituting the pseudo-contamination model 8 is 28.

The pseudo-contamination model 8 to be set is appropriately determined in consideration of the nuclide as the contamination source, the position, size and intensity of the contamination source, the distance L from the evaluation area 6, and the like, but is not necessarily uniquely determined.

For example, the range of the distance L from the evaluation area 6 where the level of necessity for manufacturing the learning model 38 is relatively “large”, “medium”, and “small” is L <L1, L1 ≦ L, respectively. When <L2 and L ≧ L2, the pseudo-contamination model 8 has a distance L from the evaluation area 6 of L1 ≦ L <L2, where the level of necessity for manufacturing the learning model 38 is relatively “medium”. It is usually set within the range, but this is not always the case.

When performing a single voxel analysis on the analysis target 1 (aggregate of 2000 voxels 7) exemplified in FIG. 6, a pseudo-contamination model 8 composed of 28 voxels 7e in addition to 35 voxels 7a is used. In order to set, the required number of analysis patterns is 63 patterns, which is larger than the case where the influence of the pollution source located on the side where the distance from the evaluation area 6 is farther than the broken line BL is completely ignored, but the voxels 7a are not narrowed down. The number of analysis patterns required can be significantly reduced as compared with the case where single voxel analysis is performed (2000 patterns).

The dimensions of the pseudo-contamination model 8 to be set are not necessarily limited to the dimensions exemplified in FIG. The pseudo-contamination model 8 is located at a nuclide that is a pollution source having a large distance L from the evaluation area 6, the position, size and intensity of the pollution source, a distance L from any one point included in the evaluation area 6, or a broken line BL. It can be arbitrarily set in consideration of conditions such as a solid angle for expecting the yz plane.

<Second learning model manufacturing process procedure>
FIG. 7 is an example of the learning model manufacturing method according to the embodiment, and is a processing flow diagram showing the flow of the processing steps of the second learning model manufacturing processing procedure.

In the second learning model manufacturing processing procedure exemplified in FIG. 7, the first learning model manufacturing processing procedure exemplified in FIG. 4 is combined into one processing step (step S10) from the viewpoint of simplifying the drawings and description. ).

The second learning model manufacturing processing procedure includes, for example, a step (step S11) of temporarily completing machine learning for the learning model 38 with respect to the first learning model manufacturing processing procedure, and the obtained learned model 39. A step (step S12) of reading the evaluation model 41 as a reference for accuracy determination, a step of determining whether the determination accuracy of the trained model 39 satisfies the required accuracy (step S13), and a trained model. The difference is that if the determination accuracy of 39 does not satisfy the required accuracy (NO in step S13), a step (step S14) for determining to manufacture the additional learning model 38 is further provided, but the other points are different. There is virtually no difference.

Therefore, in the description of the second learning model manufacturing process procedure, the description overlapping with the description of the first learning model manufacturing process procedure will be omitted, and the different steps S11 to S14 will be mainly described.

In the second learning model manufacturing processing procedure, first, the first learning model manufacturing processing procedure (step S10) is performed, and the learning model manufacturing means 26 (FIG. 1) manufactures the learning model 38. Subsequently, the machine learning means 27 (FIG. 1) machine-learns the learning model 38 to obtain the trained model 39 (step S11).

Subsequently, the accuracy determination means 31 reads the trained model 39 and the evaluation model 41 that gives the criteria for accuracy determination (step S12), and the accuracy of the regression equation that can estimate the pollution density derived by the trained model 39 is improved. , It is determined whether or not the level that is the reference for the accuracy determination given by the evaluation model 41, that is, the required level has been reached (step S13).

When the accuracy determination means 31 determines that the accuracy of the regression equation that can estimate the contamination density derived from the trained model 39 has reached the required level (YES in step S13), the additional learning model It is determined that the production of 38 is unnecessary. That is, the processing flow of the second learning model manufacturing processing procedure proceeds from step S13 to END and ends all processing steps (END).

On the other hand, when the accuracy determining means 31 determines that the accuracy of the regression equation that can estimate the contamination density derived by the trained model 39 has not reached the required level (NO in step S13), an additional It is determined to manufacture the learning model 38 (step S14). After the manufacturing of the additional learning model 38 is decided, the processing flow of the second learning model manufacturing processing procedure proceeds from step S14 to step S10, and in the contamination density calculation device 20, the first learning model manufacturing processing procedure (step). S10) is performed and an additional learning model 38 is manufactured.

If it is determined that the accuracy of the regression equation that can estimate the contamination density derived from the trained model 39 has not reached the required level, the evaluation data 41 used for the accuracy determination is added as the training model 38. It is preferable to (manufacture). Further, it is preferable to intensively add (manufacture) the learning model 38 around the voxels 7 whose accuracy determined for each voxel 7 does not reach the required level.

Learning that satisfies the required accuracy by adding the evaluation data 41 used for the accuracy determination as the learning model 38, or by adding the learning model 38 intensively around the voxels 7 that have not reached the required level. The model 38 can be manufactured efficiently.

The second learning model manufacturing processing procedure is carried out, for example, when setting a contamination distribution over not only a single voxel but also a plurality of voxels, or when it is desired to increase the learning model 38 until the estimation accuracy reaches the required accuracy. Then it is more useful.

<Pollution density calculation processing procedure>
FIG. 8 is a processing flow diagram showing the flow of processing steps of the contamination density calculation processing procedure, which is an example of the contamination density calculation method according to the embodiment.

The pollution density calculation processing procedure shown in FIG. 8 includes a learning model manufacturing processing procedure such as a first learning model manufacturing processing procedure and a second learning model manufacturing processing procedure, and performs the learning model manufacturing processing procedure. This is an example of processing steps when the learning model 38 is manufactured by.

The pollution density calculation processing procedure corresponds to, for example, the first or second learning model manufacturing processing procedure, and the processing step (step S100) for manufacturing the learning model 38 and the machine learning means 27 (FIG. 1) are manufactured. The step of machine learning the learned model 38 to obtain the trained model 39 (step S101) and the pollution source estimation means 28 (FIG. 1) can estimate the measured air dose distribution and the pollution density derived from the trained model 39. It is provided with a step (step S102) of estimating the pollution density which is the measured air dose distribution, that is, the nuclear species which is the pollution source, and the position, size and intensity of the pollution source by using the same regression equation.

In the contamination density calculation processing procedure, first, the learning model manufacturing processing procedure (step S100) such as the first or second learning model manufacturing processing procedure is performed to manufacture the learning model 38.

Subsequently, the machine learning means 27 machine-learns the manufactured learning model 38 to obtain the trained model 39 (step S101). Once the trained model 39 is obtained, the air dose distribution measured by the pollution source estimation means 28 is subsequently measured using the measured air dose distribution and a regression equation capable of estimating the pollution density derived from the trained model 39. The pollution density, that is, the nuclide that is the source of pollution, and the position, size, and intensity of the pollution source are estimated (step S102).

When the contamination source estimation means 28 completes the estimation of the contamination density, the processing flow of the contamination density calculation processing procedure proceeds from step S102 to END, and all processing steps are completed (END).

As described above, according to the present embodiment, a pollution source pattern in which a pollution source is set inside the analysis target range is prepared, and the magnitude of the influence on the accuracy of estimating the pollution density is evaluated with respect to the prepared pollution source pattern. It is possible to significantly reduce the pollution source pattern required for manufacturing the learning model 38 (Fig. 1) required to obtain a regression equation (response function) that can estimate (calculate) the pollution density with an accuracy that satisfies the required accuracy. Therefore, it is possible to achieve both the manufacturing efficiency of the learning model 38 and the estimation accuracy of the contamination density.

Further, since the learning model 38 can be manufactured by setting the pseudo-contamination model 8 (FIG. 6) that substitutes some pollution models in consideration of the magnitude of the influence on the accuracy of estimating the pollution density. It is possible to reduce the pollution source patterns required for manufacturing the learning model 38 while ensuring the required accuracy. Therefore, it is possible to manufacture the learning model 38 capable of acquiring a regression equation capable of estimating the contamination density with an accuracy satisfying the required accuracy in a time that can be realistically created.

According to the learning model manufacturing method, the pollution density calculation method, and the pollution density calculation device 20 (FIG. 1), the pollution source model required by the current trained model 39 is used with respect to the pollution source model required by the evaluation model 41 using the evaluation model 41. It is possible to determine whether to end or continue manufacturing the learning model 38 depending on whether or not the accuracy of is satisfied with the required accuracy.

Further, since it is determined from the result of the determination whether or not the learning model 38 needs to be added (manufactured), a necessary and sufficient amount of the learning model 38 that satisfies the required accuracy can be manufactured, and the learning model 38 can be manufactured with an accuracy that satisfies the required accuracy. It is possible to manufacture a learning model 38 in which a regression equation capable of estimating the pollution density can be obtained in a time that can be realistically created.

The "processor" described in the above-described embodiment is, for example, a dedicated or general-purpose CPU, a GPU (Graphics Processing Unit), an application specific integrated circuit (ASIC), or a programmable logic device (for example, a programmable logic device). Arithmetic processing circuit that can execute programs such as Simple Programmable Logic Device (SPLD), Complex Programmable Logic Device (CPLD), and Field Programmable Gate Array (FPGA). Means. The processor realizes various functions by reading and executing a program stored in a storage circuit.

Further, instead of storing the program in the program in the storage circuit, the program may be directly incorporated in the circuit constituting the processor. In this case, the processor realizes various functions by reading and executing a program embedded in the circuit.

Further, a plurality of independent processors may be combined to form an arithmetic processing circuit capable of executing a program, and each processor may execute the program to realize each function. When a plurality of processors are provided, a storage circuit for storing programs may be provided individually for each processor, or several programs corresponding to the functions of the plurality of processors may be collectively provided.

The present invention is not limited to the above-described embodiment as it is, and can be implemented in various forms other than the above-described embodiment at the implementation stage. The present invention can be omitted, added, replaced, or modified in various ways without departing from the gist of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are also included in the scope of the invention described in the claims and the equivalent scope thereof.

1 ... Analysis target, 2 ... Wall, 3 ... Floor, 5 ... Pollution source, 6 ... Evaluation area, 7 (7a, 7c, 7e) ... Boxel, 8 ... Pseudo-contamination model, 20 ... Pollution density calculation device, 21 ... Input means , 22 ... Output means, 23 ... Pollution probability setting means, 24 ... Dose contribution calculation means, 25 ... Learning model manufacturing necessity determination means, 26 ... Learning model manufacturing means, 27 ... Machine learning means, 28 ... Pollution source estimation means, 29 ... Pseudo-contamination model setting means, 31 ... Accuracy determination means, 32 ... Storage means, 33 ... Control means, 35 ... Structure information, 36 ... Pollution probability information, 37 ... Dose contribution information, 38 ... Learning model, 39 ... Learning Finished model, 41 ... Evaluation model, st1 to st4 ... Structure.

Claims (9)

A step of setting the contamination probability of at least one coordinate position selected from all the coordinate positions included in the analysis target range set in the three-dimensional space, and a step of setting the contamination probability.
Evaluation area set inside the analysis target range and representing the position information of the range for estimating the pollution source From at least one coordinate position selected when setting the pollution probability for the evaluation area based on the position information. Steps to calculate the dose contribution, which indicates the degree of influence on the air dose of
Whether or not the learning model needs to be manufactured is determined according to the determination result of determining whether or not the dose contribution meets the set conditions among the contamination probability and the dose contribution. The decision step and
When it is decided to manufacture the learning model in the learning model manufacturing necessity determination step, it becomes the pollution source that gives at least the dose contribution among the pollution probability and the dose contribution compared in making the decision. A simulation using a contamination model in which the position, size and intensity of the nuclei and the pollution source are set as input data is executed to acquire the air dose distribution in the evaluation area as output data, and the acquired output data is associated with the input data. A learning model manufacturing method comprising a learning model acquisition step of acquiring the associated input data and output data as the learning model. The step of setting the contamination probability includes a step of determining the presence or absence of the structure based on the structure information including information on the position, shape and size of the structure in the analysis target range.
The learning model manufacturing method according to claim 1, further comprising a step of setting a contamination probability of more than 0% and 100% or less at a coordinate position determined to have the structure in the analysis target range. The step of setting the contamination probability includes a step of extracting the surface of the structure based on the structure information including information on the position, shape and size of the structure in the analysis target range.
The learning model manufacturing method according to claim 1, further comprising a step of setting a contamination probability of more than 0% and 100% or less at a coordinate position where the surface of the structure is extracted based on the structure information. The structure information includes information on the material of the structure.
The step of setting the contamination probability refers to the contamination probability information associated with the substance and shape and the contamination probability, and obtains the contamination probability associated with the same substance and shape as the material and shape of the structure. The learning model manufacturing method according to claim 2 or 3 , further comprising a step and a step of setting the acquired contamination probability as the contamination probability. The learning model acquisition step translates the location of the pollution source when at least one association between the nuclide serving as the pollution source, the position, size and intensity of the pollution source and the air dose distribution in the evaluation area is obtained. The step of obtaining the air dose distribution after this is performed by translating the obtained air dose distribution in the same translation as the translation of the pollution source.
The nuclide serving as the pollution source, the position, size and intensity of the pollution source with respect to the coordinate position after translation obtained in the finding step, and the air dose distribution of the evaluation area are obtained with the associated input data and the above-mentioned input data. The learning model manufacturing method according to any one of claims 1 to 4, which includes a step of adding to the learning model as output data. The learning model acquisition step obtains the air dose when at least one association between the nuclide serving as the pollution source, the position, size and intensity of the pollution source and the air dose distribution in the evaluation area is obtained. And the step of multiplying the intensity of the contamination source by a positive number n and multiplying it by n to associate the nuclide, the position, size and intensity of the contamination source with the air dose distribution of the evaluation area.
The location, size and intensity of the pollution source after n times associated in the associating step and the air dose distribution of the evaluation area are added to the training model as the associated input and output data. The learning model manufacturing method according to any one of claims 1 to 5, which includes steps. The learning model acquisition step includes a pseudo-contamination model that is considered to be substantially equivalent to the effect of the pollution source included in the range set inside the analysis target range on the air dose in the evaluation area. Includes steps to set up instead of pollution source
It is a step of executing the simulation by substituting the pollution source included in the set range with the pseudo pollution model.
The pseudo-contamination model is any of claims 1 to 6, which is a model that substitutes the pollution source included in the set range and includes a pollution source whose size and intensity are known in the same range as the set range. The learning model manufacturing method described in item 1. A step of machine learning a learning model manufactured by the learning model manufacturing method according to any one of claims 1 to 7 to obtain a trained model.
A pollution density calculation method comprising a step of inputting the air dose into the trained model to obtain the position and intensity of the pollution source. The pollution probability set at at least one coordinate position selected from all the coordinate positions included in the analysis target range set in the three-dimensional space and the position of the range set inside the analysis target range and estimating the pollution source. Evaluation area representing information At least the above-mentioned dose contribution degree indicating the degree of influence on the air dose from at least one coordinate position selected when setting the contamination probability with respect to the evaluation area based on the position information. A learning model manufacturing necessity determination means for determining the necessity of manufacturing a learning model according to a judgment result of determining whether or not the dose contribution satisfies the set condition.
When the learning model manufacturing necessity determination means decides to manufacture the learning model, the position of the pollution source that gives at least the dose contribution among the pollution probability and the dose contribution compared in making the decision. A simulation using a pollution model with a set size and intensity as input data is executed to acquire the air dose distribution in the evaluation area as output data, the acquired output data is associated with the input data, and the associated input data and output. A learning model manufacturing means that manufactures data as the learning model, and
Using the relationship between the air dose distribution derived by the learning model after machine learning the learning model manufactured by the learning model manufacturing means and the position, size and intensity of the pollution source, and the measured air dose distribution. A pollution density calculation device comprising the means for estimating a pollution source for estimating the position, size, and intensity of a pollution source having the measured air dose distribution.

Images