Information processing method, information processing device, and information processing program for generating a learning model capable of identifying voids in composite materials

Abstract

This information processing method, information processing device, and information processing program acquire, for first distribution data indicating a physical amount distribution of a member, teaching data with a first label indicating a cell structural element attached thereto, for each cell constituting the first distribution data. From among the cells constituting the first distribution data, the proportion of cells having attached thereto a special label indicating a void is calculated on the basis of the teaching data. If the proportion is less than a prescribed threshold value, corrected teaching data is generated from the teaching data by changing the first label attached to a cell located outside the member to the special label. By performing machine learning on the basis of the corrected teaching data, a learning model for estimating the first label on the basis of the first distribution data is generated.

Description

[Technical Field]

[0001] The present disclosure relates to an information processing method, an information processing device, and an information processing program for generating a learning model capable of identifying voids in a composite material. [Background technology]

[0002] Patent Document 1 discloses a technique for extracting voids from a three-dimensional image obtained by capturing an image of a fiber-reinforced composite material using an X-ray CT scanner. According to this technique, the three-dimensional image is binarized to create a binary image, and the binary image is subjected to distance transformation to create a distance image. Then, a closing process is performed on the binary image using the distance image, and voids are extracted from the difference between the images before and after the closing process. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2015-068755 Summary of the Invention [Problem to be solved by the invention]

[0004] In Patent Document 1, in order to improve the accuracy of void extraction, it is conceivable to generate a learning model by machine learning using 3D images as training data. However, when there are few voids in the 3D image, there is a problem that the accuracy of void identification using the learning model decreases due to a lack of training data. This problem becomes more pronounced when generating a learning model for an inspection target of a high-quality composite material with fewer voids.

[0005] The present disclosure has been made in view of the above-mentioned circumstances, and aims to provide an information processing method, an information processing device, and an information processing program for generating a learning model for accurately identifying voids contained in a composite material including a reinforcing material and a matrix. [Means for solving the problem]

[0006] An information processing method according to the present disclosure controls a controller to receive training data, in which first distribution data indicating a physical quantity distribution related to a component including a plurality of structural elements is assigned a first label indicating the structural element of each cell constituting the first distribution data. The controller calculates, based on the training data, a proportion of cells constituting the first distribution data that are assigned a specific label indicating a void. The controller then determines whether the proportion is less than a predetermined threshold. If the proportion is less than the predetermined threshold, the controller generates corrected training data from the training data by changing the first labels assigned to cells located outside the component to the specific labels. Machine learning is performed based on the corrected training data to generate a learning model that estimates the first labels based on the first distribution data.

[0007] The learning model may perform semantic segmentation to estimate the first label based on the first distribution data.

[0008] The controller may change the first label assigned to a cell located outside the component and within a predetermined distance from the component to the specific label when the ratio is less than the predetermined threshold, and the predetermined distance may be set based on a kernel size in the semantic segmentation.

[0009] The physical quantity distribution may be a distribution representing the absorbance of X-rays at each point inside the member, obtained by an X-ray CT scanner.

[0010] The controller may identify cells that are located outside the component among the cells that make up the first distribution data, based on design data of the component or three-dimensional measurement data of the component.

[0011] The structural elements may be one or more fiber bundles, a matrix, and / or the voids.

[0012] The controller may acquire second distribution data indicating a physical quantity distribution of the component, and may calculate an output from the learning model corresponding to an input based on the second distribution data, thereby estimating a second label indicating whether or not each cell constituting the second distribution data is a void.

[0013] The controller may perform morphology processing on the second distribution data in which the second label has been estimated for each cell, and update the second label for each cell constituting the second distribution data.

[0014] An information processing device according to the present disclosure includes an acquisition unit and a controller. The acquisition unit acquires training data for first distribution data showing a physical quantity distribution of a component including a plurality of structural elements, in which a first label indicating the structural element of each cell constituting the first distribution data is assigned to the cell. The controller calculates, based on the training data, a proportion of cells constituting the first distribution data that are assigned a specific label indicating a void. The controller determines whether the proportion is less than a predetermined threshold. If the proportion is less than the predetermined threshold, the controller generates corrected training data from the training data by changing the first label assigned to cells located outside the component to the specific label. Machine learning is performed based on the corrected training data to generate a learning model that estimates the first label based on the first distribution data.

[0015] An information processing program according to the present disclosure processes training data in which first distribution data indicating a physical quantity distribution of a component including a plurality of structural elements is assigned a first label indicating the structural element of each cell constituting the first distribution data. The program causes a computer to execute a step of calculating, based on the training data, a proportion of cells constituting the first distribution data that are assigned a specific label indicating a void. The program also causes a computer to execute a step of determining whether the proportion is less than a predetermined threshold. If the proportion is less than the predetermined threshold, the program causes a computer to execute a step of generating corrected training data from the training data by changing the first labels assigned to cells located outside the component to the specific labels. The program also causes a computer to execute a step of generating a learning model that estimates the first labels based on the first distribution data by performing machine learning based on the corrected training data. [Effects of the Invention]

[0016] According to the present disclosure, it is possible to generate a learning model for accurately identifying voids contained in a composite material including a reinforcing material and a matrix. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a block diagram illustrating a configuration of an information processing device according to an embodiment of the present disclosure. [Figure 2] 10 is a flowchart illustrating a processing procedure when generating a learning model in an information processing device according to an embodiment of the present disclosure. [Figure 3] 10 is a flowchart illustrating a processing procedure for label estimation performed by an information processing device according to an embodiment of the present disclosure. [Figure 4] FIG. 2 is a diagram showing an example of first distribution data constituting training data. [Figure 5] FIG. 1 shows an arrangement of cells with specific labels indicating voids. [Figure 6] A diagram showing the arrangement of cells with specific labels indicating voids in the corrected training data. DETAILED DESCRIPTION OF THE INVENTION

[0018] Hereinafter, several exemplary embodiments will be described with reference to the drawings. Note that common parts in the drawings are given the same reference numerals, and duplicated explanations will be omitted.

[0019] [Subject to processing by information processing device] 1, the information processing device 20 is connected to an imaging unit 10. Here, the imaging unit 10 irradiates an electromagnetic wave (e.g., X-ray) onto a member to be measured, and detects the electromagnetic wave that has passed through, thereby obtaining the distribution of physical quantities inside the member.

[0020] For example, the imaging unit 10 is an X-ray CT device. In this case, the imaging unit 10 captures an image to acquire the physical quantity distribution inside the component. The "physical quantity distribution" inside the component represents the luminance at each point inside the component. The luminance at each point inside the component is associated with a "physical quantity" at each point, such as the X-ray absorbance or the degree of X-ray attenuation at each point. The luminance at each point inside the component itself may be treated as a "physical quantity."

[0021] The imaging unit 10 may acquire the luminance at each point on a predetermined cross section of a component and generate a two-dimensional image corresponding to the cross section. Here, the predetermined cross section may be arbitrarily set. For example, the coordinate axes of the component's stationary system may be defined as the X-axis, Y-axis, and Z-axis, and the predetermined cross section may be set as a plane parallel to any of the XY plane, YZ plane, and XZ plane. Alternatively, the imaging unit 10 may set multiple cross sections parallel to each other, group the two-dimensional images generated for each cross section, and generate a physical quantity distribution in three-dimensional space.

[0022] The component (composite material) from which the physical quantity distribution is acquired by the imaging unit 10 includes multiple structural elements. For example, the structural elements are at least one of a reinforcing material, a matrix (base material), and voids. The component may be obtained by impregnating a molded reinforcing material with a matrix to combine the reinforcing material and the matrix. The reinforcing material may be, for example, graphite, boron nitride, or ceramics such as silicon carbide. The reinforcing material may also be a resin such as Kevlar (registered trademark), or an appropriate metal or alloy. The reinforcing material may also be made of one or more fiber bundles.

[0023] The matrix can be, for example, a thermosetting resin, a thermoplastic resin, or a ceramic such as silicon carbide. Components with a ceramic matrix are specifically called ceramic matrix composites (CMCs). Voids are gap-like defects that occur in components.

[0024] The information processing device 20 acquires a physical quantity distribution inside a component including a plurality of structural elements using the imaging unit 10. The information processing device 20 may have a function of identifying a structural element to which a cell included in the component belongs, based on the physical quantity distribution acquired by the imaging unit 10.

[0025] A cell is the smallest unit of data used to represent a solid object on a computer. For example, a cell is a pixel or a voxel. A pixel may be a square or a rectangle, which are unit elements of an orthogonal lattice. A pixel may also be another polygon. A voxel may be a cube or a rectangular parallelepiped, which are unit elements of an orthogonal lattice. A voxel may also be another polyhedron.

[0026] On a computer, a component including multiple structural elements is represented as a set of cells, each of which has a scalar value and a vector value. By assigning the above-mentioned physical quantities to each cell, the entire set of cells can represent the above-mentioned "physical quantity distribution."

[0027] [Configuration of information processing device] 1 is a block diagram showing the configuration of an information processing device 20. The information processing device 20 includes a receiving unit 21 (acquiring unit), a database 23 (storage unit), a controller 25, an operation unit 27, and a presentation unit 29. The controller 25 is connected to the receiving unit 21, the database 23, the operation unit 27, and the presentation unit 29 so as to be able to communicate with them.

[0028] In addition, the operation unit 27 and the presentation unit 29 may be provided in the information processing device 20 itself, or may be installed outside the information processing device 20 and connected to the information processing device 20.

[0029] The receiving unit 21 is connected to the imaging unit 10 wirelessly or via a wire so as to be able to communicate with the imaging unit 10. The receiving unit 21 receives data (first distribution data, second distribution data) indicating the distribution of physical quantities inside the member from the imaging unit 10. The receiving unit 21 may receive two-dimensional images generated for each predetermined cross section of the member, or may receive data indicating the distribution of physical quantities in three-dimensional space obtained by grouping two-dimensional images.

[0030] The receiving unit 21 may also receive data indicating the physical quantity distribution (first distribution data, second distribution data) from a storage medium, etc. In this case, the data indicating the physical quantity distribution stored in the storage medium, etc. may be measured using an imaging unit 10 that is not connected to the information processing device 20.

[0031] The receiving unit 21 may receive design data of a component or three-dimensional measurement data of the component. The design data of the component is, for example, data generated by CAD (Computer-aided design) when the component is designed. The three-dimensional measurement data of the component is, for example, data indicating the outer shape of the component obtained by actually measuring the component using a three-dimensional measuring device.

[0032] Alternatively, the receiving unit 21 may receive "teacher data." Here, "teacher data" refers to data in which a first label indicating the structural element of each cell is assigned to first distribution data indicating a physical quantity distribution related to a component including a plurality of structural elements. A method for assigning a first label to each cell of the first distribution data will be described later.

[0033] The database 23 stores data indicating the distribution of physical quantities received by the receiving unit 21. That is, the database 23 may store two-dimensional images generated for each predetermined cross section of a component, or may store data indicating the distribution of physical quantities in three-dimensional space. In addition, the database 23 may store, together with the received data, a timestamp specifying the date and time of imaging by the imaging unit 10, an identification number specifying the component that was the subject of measurement by the imaging unit 10, and the like.

[0034] The operation unit 27 is an input device that can be operated by a user, such as a monitor or maintenance staff of the evaluation system. For example, the operation unit 27 is a keyboard, a mouse, a trackball, a touch panel, or the like. The operation unit 27 is not limited to the examples given here. The content of the user's operation input via the operation unit 27 is transmitted to the controller 25.

[0035] The operation unit 27 may receive an operation from the user to set (attach) a first label indicating a structural element of the cell for each cell constituting the first distribution data indicating the distribution of physical quantities inside the component. The user sets, via the operation unit 27, a first label indicating a structural element of the cell for each cell constituting the first distribution data, while referring to the first distribution data presented by the presentation unit 29 described below. The first distribution data to which the first label indicating the structural element of the cell has been attached for each cell is input to the controller 25 as "teaching data."

[0036] The presentation unit 29 displays information generated by the controller 25. The presentation unit 29 may also present to the user data indicating the distribution of physical quantities inside the component (first distribution data, second distribution data) and labels identifying the structural elements to which voxels included in the component belong. Alternatively, the presentation unit 29 may present to the user a label identifying whether a voxel is a void or not. For example, the presentation unit 29 may be a display that displays figures and characters by combining a plurality of display pixels. The presentation unit 29 is not limited to the examples given here.

[0037] The information processing device 20 may include an output unit (not shown) that outputs information generated by the controller 25 to the outside of the information processing device 20. The output unit may output a label that identifies a structural element to which a cell belongs. The output unit may also output a label that identifies whether a cell is a void. The output unit may also output a learning model.

[0038] The controller 25 (control unit) is a general-purpose computer equipped with a CPU (Central Processing Unit), a memory, and an input / output unit. A computer program (information processing program) for functioning as the information processing device 20 is installed in the controller 25. By executing the computer program, the controller 25 functions as multiple information processing circuits (251, 253, 255, 257, 259) included in the information processing device 20. The computer program may be stored in a storage medium that can be read and written by a computer. Alternatively, the computer program may be distributed via a telecommunications line.

[0039] In the present disclosure, an example is shown in which multiple information processing circuits (251, 253, 255, 257, 259) are realized by software. However, it is also possible to configure the information processing circuits (251, 253, 255, 257, 259) by preparing dedicated hardware for executing each of the information processes described below. Also, the multiple information processing circuits (251, 253, 255, 257, 259) may be configured by individual hardware. Furthermore, the information processing circuits (251, 253, 255, 257, 259) may also serve as a control unit used for monitoring or controlling the imaging unit 10.

[0040] As shown in FIG. 1, the controller 25 includes a ratio calculation unit 251, a determination unit 253, a label change unit 255, a learning model generation unit 257, and an evaluation unit 259 as a plurality of information processing circuits (251, 253, 255, 257, 259).

[0041] Note that, when label estimation using a learning model is not performed in the information processing device 20, the evaluation unit 259 may be omitted. A controller for estimating labels using a learning model may be prepared separately from the controller 25 including the ratio calculation unit 251, the determination unit 253, the label change unit 255, and the learning model generation unit 257, and the evaluation unit 259 may be provided in the controller.

[0042] The ratio calculation unit 251 calculates, based on the teacher data, the ratio of cells to which a specific label indicating that the cells are voids to the cells that make up the first distribution data.

[0043] For example, the percentage calculation unit 251 may determine the percentage by dividing the volume occupied by all cells labeled with a specific label indicating that they are voids by the volume of all cells constituting the first distribution data. Alternatively, the percentage calculation unit 251 may determine the percentage by dividing the total number of cells labeled with a specific label indicating that they are voids by the total number of cells constituting the first distribution data.

[0044] The ratio calculation unit 251 may determine the ratio by dividing the volume occupied by all cells labeled with a specific label indicating that they are voids by the internal volume of the component. The ratio calculation unit 251 may also determine the ratio by dividing the total number of cells labeled with a specific label indicating that they are voids by the total number of cells located inside the component.

[0045] The determination unit 253 determines whether or not the ratio calculated by the ratio calculation unit 251 is less than a predetermined threshold value. Note that the predetermined threshold value is set in advance by, for example, a user.

[0046] The label change unit 255 generates corrected training data from the training data. More specifically, when the determination unit 253 determines that the ratio is less than a predetermined threshold, the label change unit 255 changes the first label assigned to the cell located outside the component in the training data to a specific label. Therefore, the specific label is assigned to the cell located outside the component in the corrected training data.

[0047] When the determination unit 253 determines that the ratio is less than the predetermined threshold, the label change unit 255 may change, in the training data, the first labels assigned to cells that are outside the component and within a predetermined distance from the component to specific labels. The label change unit 255 may not change the first labels assigned to cells that are outside the predetermined distance range.

[0048] The predetermined distance may be set based on the kernel size in semantic segmentation. "Semantic segmentation" and "kernel size" will be described later.

[0049] Alternatively, the label changing unit 255 may specify cells located outside the component among the cells constituting the first distribution data based on the design data of the component or the three-dimensional measurement data of the component.

[0050] The manner in which the label changing unit 255 changes the first label attached to a cell located outside a component to a specific label will be described with reference to FIGS. 4 to 6. FIG. 4 is a diagram showing an example of first distribution data constituting the training data BD1. FIG. 5 is a diagram showing the arrangement of cells attached with a specific label indicating a void. FIG. 6 is a diagram showing the arrangement of cells attached with a specific label indicating a void in the corrected training data BD2. The pixels constituting FIGS. 4 to 6 correspond to the cells in this embodiment.

[0051] In Figure 4, the different colors of each cell represent the different first labels assigned to each cell. In Figure 4, cells made of reinforcing material exist in regions R1, R2, and R3. Region R1 corresponds to the fiber bundles extending in the vertical direction of the paper in Figure 4. Region R2 corresponds to the fiber bundles extending in the horizontal direction of the paper in Figure 4. Region R3 corresponds to the fiber bundles extending in the direction perpendicular to the paper in Figure 4. In addition, cells made of matrix exist in region RM. Cells made of voids exist in region RV. The region of cells located outside the component is shown as region RG.

[0052] In Figure 5, corresponding to Figure 4, cells consisting of voids are shown in black. Boundaries between regions are also shown. Cells with specific labels indicating voids are present in the black regions. According to Figures 4 and 5, the volume occupied by cells consisting of voids is smaller than the volume occupied by cells consisting of reinforcing material or matrix. Therefore, if a learning model is generated by performing machine learning based on training data such as that shown in Figure 5, the accuracy of void identification using the learning model may be reduced due to a lack of training data.

[0053] Therefore, the label changing unit 255 changes the first label attached to the cells present in the region RG to a specific label, as shown in Fig. 6. As a result, the region RG in Fig. 6 becomes a region where cells attached with a specific label indicating a void exist, similar to the region RV.

[0054] In this way, by changing the first label attached to the cell present in the region RG to a specific label, a decrease in the accuracy of void identification using the learning model is suppressed.

[0055] The learning model generation unit 257 performs machine learning based on the training data or the corrected training data to generate a learning model that estimates the first label based on the first distribution data. The learning model is represented by, for example, a neural network.

[0056] More specifically, the learning model generation unit 257 performs machine learning based on corrected teacher data generated based on teacher data whose proportion is determined by the determination unit 253 to be less than a predetermined threshold. Also, the learning model generation unit 257 performs machine learning based on teacher data whose proportion is determined by the determination unit 253 to be equal to or greater than a predetermined threshold.

[0057] The learning model generated by the learning model generation unit 257 may perform semantic segmentation to estimate a first label based on the first distribution data. Here, "semantic segmentation" refers to a deep learning algorithm that associates a first label with a cell that constitutes the first distribution data. Examples of learning models that perform semantic segmentation include FCN (Fully Convolutional Network) and U-Net, but the present embodiment is not limited to these.

[0058] In "semantic segmentation," "kernel" refers to a filter used when extracting features from the first distribution data. The filter extracts features from multiple cells that make up the first distribution data. The arrangement of cells to be processed by the filter and the number of cells can be set arbitrarily. "Kernel size" is a size that characterizes the size of the arrangement of cells to be processed, and represents, for example, the width of the arrangement of cells to be processed.

[0059] As an example of machine learning processing, the learning model generation unit 257 calculates the output obtained when the first distribution data is input to the learning model, and the error or likelihood of the first label for each cell corresponding to the input first distribution data. Then, the learning model generation unit 257 adjusts parameters that define the neural network so as to minimize the error or maximize the likelihood. As a result of such machine learning processing, the neural network learns features that represent the training data or the corrected training data.

[0060] The neural network includes an input layer to which first distribution data is input, an output layer to which output values ​​are output, and at least one or more hidden layers provided between the input and output layers, with signals propagating in the order of the input layer, hidden layer, and output layer. Each of the input, hidden, and output layers is composed of one or more units. The units between the layers are connected to each other, and each unit has an activation function (e.g., a sigmoid function, a normalized linear function, a softmax function, etc.). A weighted sum is calculated based on multiple inputs to the unit, and the value of the activation function, whose variable is the sum, becomes the output of the unit.

[0061] The learning model generation unit 257 minimizes the error between the output of the neural network and the first label or maximizes the likelihood of the output by adjusting the weights used to calculate the weighted sum for each unit among the parameters that define the neural network. To minimize the error or maximize the likelihood of the output of the neural network for multiple pieces of training data or modified training data, a maximum likelihood estimation method or the like can be applied.

[0062] To minimize errors related to the output of the neural network, for example, the learning model generation unit 257 may use gradient descent, stochastic gradient descent, etc. The learning model generation unit 257 may use backpropagation for gradient calculation in gradient descent or stochastic gradient descent.

[0063] In machine learning using neural networks, generalization performance (the ability to discriminate against unknown data) and overfitting (the phenomenon in which generalization performance does not improve while the model adapts to training data) can become problems.

[0064] Therefore, in generating a learning model in the learning model generation unit 257, a method such as regularization that restricts the degree of freedom of weights during learning may be used to mitigate overfitting. Other methods may also be used, such as dropout, which probabilistically selects units in a neural network and disables other units. Furthermore, to improve generalization performance, methods such as data regularization, data standardization, and data augmentation that eliminate bias in training data or modified training data may be used.

[0065] The learning model generated by the learning model generation unit 257 may be stored in the database 23. In addition, the generated learning model may be output to the outside of the information processing device 20 via an output unit (not shown).

[0066] The evaluation unit 259 calculates an output from the learning model corresponding to an input based on second distribution data indicating the physical quantity distribution of the component. That is, the evaluation unit 259 estimates a second label indicating whether the structural element of the cell is a void or not for each cell constituting the second distribution data based on the output from the learning model. The evaluation unit 259 sets the output from the learning model as an estimated value of the second label.

[0067] The evaluation unit 259 may acquire a generated learning model from the database 23 instead of acquiring the learning model from the learning model generation unit 257. The evaluation unit 259 may also acquire a learning model from outside the information processing device 20.

[0068] Alternatively, the evaluation unit 259 may update the second label by performing morphology processing on the second distribution data in which the second label is estimated for each cell (labeled second distribution data).

[0069] For example, the evaluation unit 259 performs "Erosion" processing or "Dilation" processing as the morphology processing. The evaluation unit 259 may perform the morphology processing on any one type of cell among the reinforcing material cells, the matrix cells, and the void cells included in the second distribution data.

[0070] The evaluation unit 259 may repeatedly perform the "Erosion" process and the "Dilation" process a predetermined number of times. The evaluation unit 259 may perform a combination of the "Erosion" process and the "Dilation" process. By performing the "Erosion" process, it is possible to remove isolated cells included in the second distribution data for which the second label has been estimated. Furthermore, by performing the "Dilation" process, it is possible to connect discontinuous cells and fill in gaps in the cells.

[0071] The evaluation unit 259 may perform an "Opening" process or a "Closing" process. The "Opening" process is a process in which "Erosion" process is performed a predetermined number of times in succession, followed by "Dilation" process a predetermined number of times in succession. The "Closing" process is a process in which "Dilation" process is performed a predetermined number of times in succession, followed by "Erosion" process a predetermined number of times in succession.

[0072] Morphological processing can remove noise contained in the second distribution data from which the second label has been estimated. Morphological processing can detect voids that are open to the outside of the component. Furthermore, even if a continuous structure in reality is expressed as a discontinuous structure in the second distribution data, morphological processing can correct the discontinuous parts to a continuous structure, making it possible to extract fine structures.

[0073] [Processing procedure of information processing device] Next, a processing procedure of the information processing device 20 according to the present disclosure will be described with reference to the flowcharts of FIGS.

[0074] [Processing procedure for generating learning model] 2 is a flowchart showing a processing procedure for generating a learning model by the information processing device 20 according to an embodiment of the present disclosure. The processing of the flowchart shown in FIG. 2 is started based on, for example, a user instruction.

[0075] In step S101, the controller 25 acquires a plurality of first distribution data via the receiving unit 21.

[0076] In step S103, the controller 25 sets a first label indicating a structural element of the cell for each cell constituting the first distribution data. Then, the controller 25 creates a plurality of pieces of training data. When creating the training data, the operation unit 27 accepts an operation from the user to set a first label for each cell.

[0077] It should be noted that instead of steps S101 and S103, the controller 25 may acquire the training data stored in the database 23.

[0078] In step S105, the ratio calculation unit 251 selects one piece of teacher data from among the plurality of pieces of teacher data.

[0079] In step S107, the ratio calculation unit 251 calculates the ratio of cells with specific labels to the selected teacher data.

[0080] In step S109, the determination unit 253 determines whether the calculated ratio is less than a predetermined threshold value.

[0081] If it is determined that the calculated ratio is less than the predetermined threshold (YES in step S109), the label change unit 255 generates corrected training data from the training data in step S111. Then, in step S113, the learning model generation unit 257 performs machine learning based on the corrected training data.

[0082] On the other hand, if it is not determined that the calculated ratio is less than the predetermined threshold (NO in step S109), in step S115, the learning model generation unit 257 performs machine learning based on the training data.

[0083] In step S117, it is determined whether or not there is any unselected teacher data.

[0084] If there is any unselected training data, the process returns to step S105.

[0085] If there is no unselected training data, it is understood that processing based on all training data has been completed. Therefore, in step S119, the learning model generation unit 257 outputs the generated learning model to the database 23, and the database 23 stores the generated learning model. Alternatively, the learning model generation unit 257 may output the generated learning model to the outside of the information processing device 20.

[0086] [Processing procedure for label estimation] Fig. 3 is a flowchart showing a processing procedure for label estimation by the information processing device 20 according to an embodiment of the present disclosure. The processing of the flowchart shown in Fig. 3 is started after a learning model is generated by the flowchart in Fig. 2. In addition, the processing of the flowchart shown in Fig. 3 is started based on, for example, a user instruction.

[0087] In step S201, the evaluation unit 259 acquires a learning model.

[0088] In step S203, the controller 25 acquires the second distribution data via the receiving unit 21.

[0089] In step S205, the evaluation unit 259 calculates an output from the learning model corresponding to the input based on the second distribution data indicating the physical quantity distribution of the component. The evaluation unit 259 sets the output from the learning model as an estimated value of the second label.

[0090] In step S207, the evaluation unit 259 outputs, for each cell constituting the second distribution data, a second label indicating whether the structural element of the cell is a void or not.

[0091] [Effects of the embodiment] As described above in detail, the information processing device, information processing method, and information processing program according to the present disclosure acquire training data. Here, the training data is data in which first distribution data showing a physical quantity distribution related to a component including multiple structural elements is assigned a first label indicating the structural element of the cell for each cell constituting the first distribution data. Then, based on the training data, the proportion of cells constituting the first distribution data that are assigned a specific label indicating a void is calculated, and it is determined whether the proportion is less than a predetermined threshold. If the proportion is less than the predetermined threshold, the first labels assigned to cells located outside the component are changed to specific labels, thereby generating corrected training data from the training data. Machine learning is performed based on the corrected training data to generate a learning model that estimates the first labels based on the first distribution data.

[0092] This allows for the generation of a learning model for accurately identifying voids contained in a composite material containing a reinforcing material and a matrix. In particular, even when generating a learning model for inspecting a high-quality composite material with fewer voids, the accuracy of void identification using the learning model is prevented from decreasing. Furthermore, the learning model can be used to identify voids inside a component and evaluate the void content with high accuracy.

[0093] In the information processing device, information processing method, and information processing program according to the present disclosure, the learning model may perform semantic segmentation to estimate the first label based on the first distribution data. This allows a first label indicating a structural element of each cell constituting the first distribution data to be output, thereby estimating the structural element of the cell. As a result, the learning model can be used to identify voids within a component and evaluate the void content with high accuracy.

[0094] The information processing device, information processing method, and information processing program according to the present disclosure may change the first label assigned to the cell to the specific label when the ratio is less than the predetermined threshold. The cell to be changed to the specific label may be located outside the component and within a predetermined distance from the component. The predetermined distance may be set based on a kernel size in the semantic segmentation. This prevents a decrease in the accuracy of void identification using the learning model, even when a learning model is generated for an inspection target of a high-quality composite material with fewer voids. Furthermore, it is possible to prevent an increase in computational costs when creating the learning model.

[0095] In the information processing device, information processing method, and information processing program according to the present disclosure, the physical quantity distribution may be a distribution representing the absorbance of X-rays at each point inside the component, acquired by an X-ray CT scanner. This allows the physical quantity distribution to be acquired by transmitting X-rays through the component. As a result, it is possible to identify voids contained in a composite material including a reinforcing material and a matrix, and to evaluate the void content with high accuracy.

[0096] The information processing device, information processing method, and information processing program according to the present disclosure may identify cells located outside the component among the cells constituting the first distribution data based on design data of the component or three-dimensional measurement data of the component. This makes it easy to change the first label attached to the cell located outside the component in the training data to a specific label. Therefore, the training data regarding voids included in the corrected training data can be amplified. As a result, a learning model for accurately identifying voids in a composite material including a reinforcing material and a matrix can be generated.

[0097] In the information processing device, information processing method, and information processing program according to the present disclosure, the structural element may be at least one of one or more fiber bundles, a matrix, and the voids, thereby enabling a list of structural elements that may be included in a component to generate a learning model.

[0098] The information processing device, information processing method, and information processing program according to the present disclosure may acquire second distribution data indicating a physical quantity distribution of the component. Then, the information processing device, information processing method, and information processing program may estimate, for each cell constituting the second distribution data, a second label indicating whether or not the cell is a void by calculating an output from the learning model corresponding to an input based on the second distribution data. This allows the learning model to identify voids inside the component and evaluate the void content with high accuracy.

[0099] The information processing device, information processing method, and information processing program according to the present disclosure may perform morphology processing on the second distribution data in which the second label has been estimated for each cell, to update the second label for each cell constituting the second distribution data. This makes it possible to remove noise contained in the second distribution data. It also makes it possible to detect voids that are open to the outside of the component. Furthermore, even if an actually continuous structure is represented as a discontinuous structure in the second distribution data, the morphology processing can correct the discontinuous portion to a continuous structure, making it possible to extract fine structures.

[0100] Each function described in this disclosure may be implemented by one or more processing circuits, including programmed processors, electrical circuits, and even devices such as application specific integrated circuits (ASICs), or circuit components arranged to perform the described functions.

[0101] According to the present disclosure, it is possible to identify voids in composite materials containing reinforcement and matrix and evaluate the void content with high accuracy. As a result, it is possible to promote the production of high-quality composite materials. Therefore, for example, it can contribute to Goal 9 of the Sustainable Development Goals (SDGs) led by the United Nations, "Build resilient infrastructure, promote inclusive and sustainable industrialization and foster innovation."

[0102] Although several embodiments have been described, the embodiments can be modified or varied based on the above disclosure. All components of the above embodiments and all features described in the claims may be individually extracted and combined, unless they contradict each other.

[0103] The entire contents of Patent Application No. 2022-179745 (filing date: November 9, 2022) are incorporated herein by reference. [Explanation of symbols]

[0104] 10. Imaging unit 20 Information processing equipment 21 Receiving unit 23 Database 25 Controller 27 Control section 29 Presentation section 251 Percentage Calculation Section 253 Judgment section 255 Label Change Section 257 Learning Model Generation Unit 259 Evaluation Department

Claims

1. 1. An information processing method executed by a controller in which teacher data is input, for each cell constituting first distribution data indicating a physical quantity distribution of a component including a plurality of structural elements, with a first label indicating a structural element of the cell, the method comprising: The controller Calculating a ratio of cells having a specific label indicating a void to the cells constituting the first distribution data based on the teacher data; determining whether the ratio is less than a predetermined threshold; If the ratio is less than the predetermined threshold, the first label assigned to the cell located outside the component is changed to the specific label, thereby generating corrected teacher data from the teacher data; An information processing method that generates a learning model that estimates the first label based on the first distribution data by performing machine learning based on the corrected training data.

2. The information processing method according to claim 1 , wherein the learning model performs semantic segmentation to estimate the first label based on the first distribution data.

3. when the ratio is less than the predetermined threshold, the controller changes the first label attached to a cell located outside the component and within a predetermined distance from the component to the specific label; The information processing method according to claim 2 , wherein the predetermined distance is set based on a kernel size in the semantic segmentation.

4. The information processing method according to claim 1 , wherein the physical quantity distribution is a distribution representing X-ray absorbance at each point inside the component, the distribution being acquired by an X-ray CT scanner.

5. The information processing method according to claim 1 , wherein the controller identifies cells that are located outside the component among the cells that constitute the first distribution data based on design data of the component or three-dimensional measurement data of the component.

6. The information processing method according to claim 1 , wherein the structural element is at least one of one or more fiber bundles, a matrix, and the voids.

7. The controller After generating the learning model, second distribution data indicating a physical quantity distribution of the component is acquired; The information processing method according to any one of claims 1 to 6, wherein a second label indicating whether or not each cell constituting the second distribution data is a void is estimated by using the generated learning model to calculate an output from the learning model corresponding to an input based on the second distribution data.

8. The information processing method according to claim 7 , wherein the controller performs morphological processing on the second distribution data in which the second label has been estimated for each cell, and updates the second label for each cell constituting the second distribution data.

9. An information processing device including an acquisition unit and a controller, The acquisition unit acquiring teacher data for each cell constituting first distribution data, the teacher data being assigned a first label indicating the structural element of the cell, for first distribution data indicating a physical quantity distribution related to a member including a plurality of structural elements; The controller Calculating a ratio of cells having a specific label indicating a void to the cells constituting the first distribution data based on the teacher data; determining whether the ratio is less than a predetermined threshold; If the ratio is less than the predetermined threshold, the first label assigned to the cell located outside the component is changed to the specific label, thereby generating corrected teacher data from the teacher data; An information processing device that generates a learning model that estimates the first label based on the first distribution data by performing machine learning based on the corrected training data.

10. For first distribution data indicating a physical quantity distribution related to a member including a plurality of structural elements, using training data in which a first label indicating a structural element of each cell constituting the first distribution data is assigned, On the computer, calculating, based on the teacher data, a ratio of cells that are labeled with a specific label indicating that they are voids to the cells that make up the first distribution data; determining whether the ratio is less than a predetermined threshold; generating corrected training data from the training data by changing the first label assigned to a cell located outside the component to the specific label when the ratio is less than the predetermined threshold value; generating a learning model that estimates the first label based on the first distribution data by performing machine learning based on the corrected training data; An information processing program for executing the above.

Images