Defect estimation device, defect estimation system, defect estimation method, and program

Abstract

To precisely estimate an internal defect that a laminate of fiber-reinforced plastic has.SOLUTION: A defect estimation device comprises: a prior learning section for learning a model for outputting position information when feature data are inputted on the basis of the feature data generated by analyzing a temperature change occurring on a surface of a laminate on the basis of a finite element model showing the laminate of fiber-reinforced plastic, and learning data including position information of an internal defect that the laminate has; a transfer learning section for updating a model learned by the prior learning section on the basis of feature data generated by measuring a temperature change occurring on the surface of the laminate and additional data including position information of the internal defect that the laminate has; a feature amount reception section for receiving input of feature data based on a temperature change occurring on the surface of the laminate; and a defect position estimation section for estimating the position information of the internal defect that the laminate has by inputting the feature data received by the feature amount reception section to a model.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present disclosure relates to a defect estimation device, a defect estimation system, a defect estimation method, and a program. [Background technology]

[0002] Carbon fiber reinforced plastic (CFRP) is a composite material that uses resin as the matrix and carbon fiber as the reinforcement material. Generally, CFRP is used by laminating prepregs that have strength and rigidity reinforced in one direction. Hereinafter, CFRP with a laminated structure will be referred to as a "CFRP laminate."

[0003] There is known a technique for estimating internal defects in a CFRP laminate using a machine learning model. For example, Patent Document 1 discloses a defect estimation device that estimates the three-dimensional position of an internal defect in a laminate by inputting feature data based on temperature changes that occur on the surface of the CFRP laminate into a model that has learned the relationship between the feature data and position information that represents the three-dimensional position of the internal defect. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7471031 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the estimation accuracy of the conventional techniques leaves room for improvement. For example, the conventional techniques create training data by measuring real samples, and therefore, unless a sufficient number of real samples are collected, internal defects cannot be accurately estimated. Furthermore, for example, the conventional techniques disclose a method for creating training data by numerical analysis, but do not disclose a method for creating training data similar to training data obtained from real samples.

[0006] One aspect of the present disclosure aims to provide a technology that can accurately estimate internal defects in a fiber-reinforced plastic laminate. [Means for solving the problem]

[0007] A defect estimation device according to one aspect of the present disclosure includes a pre-learning unit that learns a model that outputs position information when feature data is input, based on learning data including feature data generated by analyzing temperature changes that occur on the surface of the laminate based on a finite element model representing a fiber-reinforced plastic laminate and position information of internal defects in the laminate; a transfer learning unit that updates the model learned by the pre-learning unit based on additional data including feature data generated by measuring temperature changes that occur on the surface of the laminate and position information of internal defects in the laminate; a feature receiving unit that receives input of feature data based on the temperature changes that occur on the surface of the laminate; and a defect position estimation unit that estimates position information of internal defects in the laminate by inputting the feature data received by the feature receiving unit into the model. [Effects of the Invention]

[0008] According to one aspect of the present disclosure, internal defects in a fiber-reinforced plastic laminate can be efficiently estimated. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a block diagram showing an example of the overall configuration of a defect estimation system. [Figure 2] FIG. 1 is a diagram illustrating an example of a temperature measuring device. [Figure 3] FIG. 1 is a block diagram illustrating an example of a computer. [Figure 4] FIG. 1 is a diagram showing an example of a CFRP laminate. [Figure 5] FIG. 2 is a block diagram illustrating an example of a functional configuration of the defect estimation system. [Figure 6] FIG. 10 is a diagram illustrating an example of location information. [Figure 7]1 is a flowchart illustrating an example of a defect estimation method. [Figure 8] 10 is a flowchart illustrating an example of a pre-learning process. [Figure 9] 10 is a flowchart illustrating an example of a transfer learning process. [Figure 10] 10 is a flowchart illustrating an example of an estimation process. [Figure 11] FIG. 10 is a diagram showing a first example of an estimation result. [Figure 12] FIG. 10 is a diagram showing a second example of an estimation result. [Figure 13] FIG. 10 is a diagram illustrating an example of estimation accuracy. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functional configurations are designated by the same reference numerals, and redundant description will be omitted.

[0011] [Embodiment] One embodiment of the present disclosure is a defect estimation system that estimates an internal defect in an object to be inspected. In this embodiment, the object to be inspected is a fiber-reinforced plastic laminate (CFRP laminate). The defect estimation system has a function of estimating position information indicating the three-dimensional position of an internal defect in the CFRP laminate based on a machine learning model.

[0012] Carbon fiber reinforced plastics (CFRP) have a high specific modulus of elasticity and are widely used in the aerospace and aviation fields. In addition, as the electrification of automobiles progresses, their use is expected to expand in the automotive field as well.

[0013] CFRP laminates may develop internal defects during use or manufacturing. Internal defects that occur during use include, for example, delamination, fiber breakage, and base material cracking. Internal defects that occur during manufacturing include, for example, the inclusion of foreign matter.

[0014] Non-destructive testing such as ultrasonic measurement (see Reference 1) or radiography (see Reference 2) has been used to detect internal defects in CFRP laminates. However, ultrasonic measurement requires contact with materials such as water or oil, and the skill of the testing technician has a significant impact on the accuracy of the test results. In addition, radiography requires safety management, such as the establishment of a controlled area, and requires time for development, resulting in high economic and time costs.

[0015] [Reference 1] Shuichi Mikami, Toshiyuki Oshima, Noboru Sugawara, and Tomoyuki Yamazaki, "Quantitative Evaluation of Defect Detection by Ultrasonic Flaw Detection Methods Using Detailed Analysis of Echo Waveforms," ​​Transactions of the Japan Society of Civil Engineers, vol. 501, pp. 103-112, 1994. [Reference 2] Sultan, M., Worden, K., Pierce, S., Hickey, D., Staszewski,W., Dulieu-Barton, J., and Hodzic, A., "On impact damage detection and quantification for CFRP laminates", Mechanical Systems and Signal Processing, vol. 25, pp. 3135-3152, 2011.

[0016] Therefore, stress analysis using infrared thermography (see Reference 3) has attracted attention. Infrared thermography is a device that uses an infrared sensor to measure the distribution of infrared energy emitted from the surface of an object and converts it into a temperature distribution. Reference 3 discloses that defect detection using infrared thermography is useful for structures such as bridges. However, the technology disclosed in Reference 3 cannot pinpoint the detailed location of internal defects.

[0017] [Reference 3] Sakagami, T., Mizokami, Y., Shiozawa, D., Fujimoto, T., Izumi, Y., Hanai, T., and Moriyama, A., "Verification of the repair effect for fatigue cracks in members of steel bridges based on thermoelastic stress measurement", Engineering Fracture Mechanics, vol. 183, pp. 1-12, 2017.

[0018] Additionally, research is being conducted into a method using a machine learning model (see Reference 4). In Reference 4, a neural network is created that uses the first to third natural frequencies as input data to determine the location and amount of damage. However, in Reference 4, the CFRP laminate is divided into 10 sections in the longitudinal direction, and the location and amount of damage are determined for eight sections, excluding the two sections at the tip. Therefore, the technology disclosed in Reference 4 has low resolution for the output position information, making it impossible to identify the detailed location of internal defects.

[0019] [Reference 4] Goro, K. Nishi, Zheng Hwang, and Yumi Fujikawa, "Damage Identification of CFRP Laminates Using Neural Networks and Experimental Data," Transactions of the Japan Society of Mechanical Engineers, Series A, vol. 62, pp. 2338-2343, 1996.

[0020] Non-Patent Document 1 discloses a defect estimation device that estimates the three-dimensional location of an internal defect in a CFRP laminate from an image obtained by measuring the surface temperature of the laminate using infrared thermography. The estimation uses a convolutional neural network (CNN) that has learned the relationship between the surface temperature change distribution of the CFRP laminate and the three-dimensional location of an internal defect in the laminate. Non-Patent Document 1 makes it possible to precisely estimate the three-dimensional location of an internal defect in a CFRP laminate.

[0021] In Non-Patent Document 1, real samples with various internal defects are collected and analyzed using infrared thermography to create training data. Therefore, if a sufficient number of real samples are not collected, a highly accurate model cannot be trained, and internal defects cannot be accurately estimated. Non-Patent Document 1 also discloses a method for creating training data using numerical analysis based on the finite element method, but does not disclose a method for creating training data similar to training data obtained from real samples. Even if a model is trained based on training data that is different from training data obtained from real samples, internal defects cannot be accurately estimated.

[0022] The present embodiment aims to accurately estimate internal defects in a fiber-reinforced plastic laminate. To this end, in this embodiment, a model is pre-trained based on training data created by numerical analysis, and the model is then transferred based on additional data created by measuring actual samples. In one aspect, this embodiment makes it possible to accurately estimate internal defects in a fiber-reinforced plastic laminate.

[0023] <Overall structure> The overall configuration of the defect estimation system in this embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the overall configuration of the defect estimation system.

[0024] 1, the defect estimation system 100 includes a temperature measuring device 1 and a defect estimation device 2. The temperature measuring device 1 and the defect estimation device 2 are connected to each other so as to be able to communicate data with each other via a communication network N such as a LAN (Local Area Network) or the Internet.

[0025] The temperature measuring device 1 is an electronic device that measures the surface temperature of a CFRP laminate that is a learning or estimation target. The temperature measuring device 1 changes the surface temperature of the CFRP laminate and measures the surface temperature of the CFRP laminate using an infrared camera before and after the surface temperature change. The temperature measuring device 1 also calculates the surface principal stress sum distribution of the CFRP laminate based on two images (hereinafter also referred to as "surface temperature images") that represent the surface temperature distribution measured before and after the surface temperature change. The surface principal stress sum distribution is an example of feature data.

[0026] The method for changing the surface temperature of a CFRP laminate varies depending on the shape and application of the CFRP laminate. In this embodiment, the surface temperature is changed by applying a fixed external force to the CFRP laminate. Specifically, both ends of the plate-shaped CFRP laminate are supported and pulled in opposite directions.

[0027] The method for changing the surface temperature of a CFRP laminate is not limited to applying a fixed external force. For example, the surface temperature may be changed by applying heat to the CFRP laminate. In this case, infrared rays are irradiated onto the surface of the CFRP laminate to raise the surface temperature, and then the temperature change over time is measured.

[0028] The defect estimation device 2 is an example of an information processing device such as a personal computer, workstation, or server that estimates position information of an internal defect in a CFRP laminate. Based on the surface principal stress sum distribution of the CFRP laminate and position information of an internal defect in the laminate, the defect estimation device 2 learns an estimation model that inputs the surface principal stress sum distribution and outputs an estimated value of the position information of the internal defect. Furthermore, the defect estimation device 2 uses the learned estimation model to estimate the position information of the internal defect from the surface principal stress sum distribution of the CFRP laminate that is the estimation target.

[0029] The overall configuration of the defect estimation system 100 shown in FIG. 1 is one example, and various system configuration examples are possible depending on the application and purpose. For example, the defect estimation system 100 may include multiple units of one or more of the temperature measurement device 1 and the defect estimation device 2. For example, the defect estimation device 2 may be realized by multiple computers, or may be realized as a cloud computing service. For example, the temperature measurement device 1 and the defect estimation device 2 may be realized by a standalone computer. The classification of devices such as the temperature measurement device 1 and the defect estimation device 2 shown in FIG. 1 is one example.

[0030] <Hardware configuration> The hardware configuration of the defect estimation system 100 will be described with reference to FIGS.

[0031] ≪Temperature measuring device≫ 2 is a block diagram showing an example of the temperature measuring device 1. As shown in FIG. 2, the temperature measuring device 1 includes a control device 101, a pair of support parts 102, a drive control part 103, and an infrared camera 104.

[0032] The support part 102 is made up of a first support part 102A that supports one end of the CFRP laminate 1000 and a second support part 102B that supports the other end. The drive control part 103 controls the first support part 102A and the second support part 102B so that they each move linearly along a line connecting the first support part 102A and the second support part 102B.

[0033] With the first support part 102A and the second support part 102B supporting both ends of the CFRP laminate 1000, the drive control part 103 moves the first support part 102A and the second support part 102B in opposite directions to each other, thereby applying tensile stress to the CFRP laminate 1000. This causes a temperature drop in the CFRP laminate 1000 proportional to the stress variation, and the surface temperature of the CFRP laminate 1000 changes.

[0034] The infrared camera 104 receives infrared rays emitted from the surface of the CFRP laminate 1000 and generates a surface temperature image of the CFRP laminate 1000 by analyzing the infrared rays.

[0035] The control device 101 is realized by, for example, a computer. The control device 101 has a CPU 501, a ROM 502, a RAM 503, an input device 505, a display device 506, and an external I / F 508. The CPU 501, the ROM 502, and the RAM 503 form a so-called computer. The hardware components of the control device 101 are connected to each other via a bus line 509. The input device 505 and the display device 506 may be connected to the external I / F 508 for use.

[0036] The drive control unit 103 and the infrared camera 104 are connected to an external I / F 508 and are controlled by the control device 101. In addition, a surface temperature image acquired by the infrared camera 104 is input to the control device 101 via the external I / F 508.

[0037] Computer The defect estimation device 2 may be realized by, for example, a computer. Fig. 3 is a block diagram showing an example of a computer.

[0038] 3, the computer 500 includes a CPU (Central Processing Unit) 501, a ROM (Read Only Memory) 502, a RAM (Random Access Memory) 503, a HDD (Hard Disk Drive) 504, an input device 505, a display device 506, a communication I / F (Interface) 507, and an external I / F 508. The CPU 501, the ROM 502, and the RAM 503 form a so-called computer. The hardware components of the computer 500 are connected to each other via a bus line 509. The input device 505 and the display device 506 may be connected to the external I / F 508 for use.

[0039] The CPU 501 is a computing device that reads programs and data from a storage device such as the ROM 502 or the HDD 504 onto the RAM 503 and executes the processes, thereby realizing the overall control and functions of the computer 500. The computer 500 may have a GPU (Graphics Processing Unit) in addition to or instead of the CPU 501.

[0040] The ROM 502 is an example of a non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. The ROM 502 functions as a main storage device that stores various programs, data, etc. required for the CPU 501 to execute various programs installed in the HDD 504. Specifically, the ROM 502 stores boot programs such as a Basic Input / Output System (BIOS) and an Extensible Firmware Interface (EFI) that are executed when the computer 500 starts up, as well as data such as OS (Operating System) settings and network settings.

[0041] The RAM 503 is an example of a volatile semiconductor memory (storage device) in which programs and data are erased when the power is turned off. The RAM 503 is, for example, a dynamic random access memory (DRAM) or a static random access memory (SRAM). The RAM 503 provides a working area in which various programs installed in the HDD 504 are expanded when executed by the CPU 501.

[0042] The HDD 504 is an example of a non-volatile storage device that stores programs and data. The programs and data stored in the HDD 504 include an OS, which is basic software that controls the entire computer 500, and applications that provide various functions on the OS. Note that the computer 500 may use a storage device that uses flash memory as a storage medium (e.g., an SSD (Solid State Drive)) instead of the HDD 504.

[0043] The input device 505 is a touch panel, operation keys or buttons, keyboard or mouse, microphone for inputting audio data such as voice, etc. that are used by the user to input various signals.

[0044] The display device 506 is composed of a display such as a liquid crystal or organic EL (Electro-Luminescence) for displaying a screen, a speaker for outputting audio data such as voice, etc.

[0045] The communication I / F 507 is an interface for connecting to a communication network and enabling the computer 500 to perform data communication.

[0046] The external I / F 508 is an interface with an external device. Examples of external devices include a drive device 510, etc.

[0047] The drive device 510 is a device for setting a recording medium 511. The recording medium 511 here includes media that record information optically, electrically, or magnetically, such as CD-ROMs, flexible disks, magneto-optical disks, etc. Also, the recording medium 511 may include semiconductor memories that record information electrically, such as ROMs, flash memories, etc. Thereby, the computer 500 can read and / or write to the recording medium 511 via the external I / F 508.

[0048] In addition, various programs installed in the HDD 504 are installed, for example, by setting a distributed recording medium 511 in a drive device 510 connected to the external I / F 508 and reading out the various programs recorded on the recording medium 511 by the drive device 51*. Alternatively, various programs installed in the HDD 504 may be installed by being downloaded from another network different from the communication network via the communication I / F 507.

[0049] <Structure of the CFRP laminate> The structure of the CFRP laminate will be described with reference to Fig. 4. Fig. 4 is a diagram showing an example of a CFRP laminate.

[0050] As shown in Fig. 4, a CFRP laminate 1000 is formed by laminating a plurality of prepregs 1001. In each prepreg 1001, the strength and rigidity of a resin matrix 1002 is reinforced in one direction by carbon fiber reinforcing fibers 1003. When laminating the prepregs 1001, the fiber orientation in each layer is rotated by an equal angle, so that the CFRP laminate 1000 exhibits isotropy as a whole.

[0051] While Fig. 4 shows an example in which each prepreg is rotated by 90°, it may also be rotated by, for example, 30°, 45°, 60°, etc. Also, Fig. 4 shows an example in which 10 layers are laminated, but the number of layers is not limited as long as it is two or more. In this embodiment, a CFRP laminate is assumed, but the types of reinforcing fibers and resins may be changed as appropriate.

[0052] <Functional configuration> The functional configuration of the defect estimation system 100 will be described with reference to Fig. 5. Fig. 5 is a block diagram showing an example of the functional configuration of the defect estimation system.

[0053] <Functional configuration of the temperature measurement device> As shown in FIG. 5, the temperature measuring device 1 includes a temperature changing unit 11, a temperature measuring unit 12, and a feature amount calculating unit 13.

[0054] The temperature changing unit 11 is realized by the support unit 102 and the drive control unit 103 shown in Fig. 2. The temperature measuring unit 12 is realized by the infrared camera 104 shown in Fig. 2. The feature amount calculating unit 13 is realized by processing that is executed by the CPU 501 in accordance with a program loaded from the ROM 502 onto the RAM 503 shown in Fig. 2.

[0055] The temperature changing unit 11 changes the surface temperature of the CFRP laminate by the drive control unit 103 driving the support unit 102.

[0056] The temperature measurement unit 12 measures the surface temperature of the CFRP laminate using the infrared camera 104. The temperature measurement unit 12 generates a surface temperature image of the CFRP laminate. The temperature measurement unit 12 sends the generated surface temperature image to the feature calculation unit 13.

[0057] The feature amount calculation unit 13 calculates feature data based on the temperature change of the CFRP laminate based on two surface temperature images generated before and after the temperature change unit 11 changed the surface temperature of the CFRP laminate. The feature data may be, for example, a surface temperature change distribution or a surface principal stress sum distribution. The feature amount calculation unit 13 transmits the calculated feature data to the defect estimation device 2.

[0058] <Functional configuration of the defect estimation device> As shown in FIG. 5, the defect estimation device 2 includes a learning data generation unit 21, a pre-learning unit 22, a feature reception unit 23, an additional data generation unit 24, a transfer learning unit 25, a model storage unit 26, a defect position estimation unit 27, and a result output unit 28.

[0059] The learning data generation unit 21, the pre-learning unit 22, the feature receiving unit 23, the additional data generation unit 24, the transfer learning unit 25, the defect position estimation unit 27, and the result output unit 28 are realized by processing executed by the CPU 501 in accordance with a program loaded from the HDD 504 onto the RAM 503 shown in Fig. 3. The model storage unit 26 is realized by the HDD 504 shown in Fig. 3.

[0060] The learning data generation unit 21 generates learning data. The learning data includes feature data based on temperature changes of the CFRP laminate and position information of internal defects in the laminate. The learning data generation unit 21 calculates the feature data by numerical analysis based on the finite element method (FEM). The learning data generation unit 21 accepts input of position information of internal defects in response to user operation. The learning data generation unit 21 generates learning data by associating the feature data with the position information of internal defects.

[0061] The training data generation unit 21 may construct finite element models representing CFRP laminates having various internal defects and calculate feature data based on each of the multiple finite element models. The training data generation unit 21 may construct finite element models representing CFRP laminates having no internal defects and calculate feature data based on the finite element models. The training data generation unit 21 may construct finite element models of CFRP laminates having internal defects and finite element models of CFRP laminates having no internal defects in a predetermined ratio. The predetermined ratio may be determined based on the past internal defect detection rate, etc., or may be adjusted based on prediction accuracy.

[0062] The training data generation unit 21 may construct a finite element model in which physical property values ​​corresponding to the internal defect are set in a region where the internal defect exists. As an example, if the internal defect is a foreign matter, the training data generation unit 21 may set an elastic modulus similar to that of the foreign matter in a region corresponding to the foreign matter. As an example, the elastic modulus may include at least one of Young's modulus and Poisson's ratio.

[0063] The training data generation unit 21 may perform data expansion by adding noise to the calculated feature data. The training data generation unit 21 may generate multiple feature data from one piece of feature data by adding multiple different noises. The training data generation unit 21 may add noise with different densities depending on the region to the feature data. The training data generation unit 21 may add noise to regions corresponding to reinforcing fibers contained in the CFRP laminate. The training data generation unit 21 may add noise with a relatively high density to regions corresponding to reinforcing fibers contained in the CFRP laminate.

[0064] The position information of the internal defect includes the three-dimensional coordinates of the internal defect in the CFRP laminate and the size of the internal defect, where the size of the internal defect is the range in each layer of the CFRP laminate where the internal defect exists.

[0065] Specifically, the position information of internal defects is information obtained by dividing each layer of a CFRP laminate into a predetermined number of meshes, and setting a channel for each mesh to indicate whether or not an internal defect exists. For example, the channel is set to 1 if an internal defect exists, and 0 if it does not. In other words, the position of each mesh represents a three-dimensional coordinate obtained by adding the depth of the layer to a two-dimensional coordinate as viewed from the surface of the CFRP laminate. Furthermore, the range in each layer where the channel value is 1 represents the size of the internal defect.

[0066] The mesh may divide each layer of the CFRP laminate at equal orthogonal intervals. The mesh may divide each layer of the CFRP laminate into a predetermined number of layers in the depth direction. The size of the mesh may be determined arbitrarily depending on the size of the CFRP laminate or the resolution for detecting internal defects. However, if the locations where defects are likely to occur are known, the area near the location may be divided finely and other areas may be divided coarsely.

[0067] Fig. 6 is a diagram showing an example of position information. As shown in Fig. 6, position information 1500 includes a plurality of layers 1501, 1502, and 1503 corresponding to the layers of the CFRP laminate. Each of the layers 1501, 1502, and 1503 is configured with a mesh divided at equal intervals in the vertical direction (X axis), horizontal direction (Y axis), and depth direction (Z axis). In Fig. 6, each of the layers 1501, 1502, and 1503 is divided into three in the depth direction, but this number of divisions is just an example and may be two or less, or four or more. The number of divisions for each layer may be the same or different.

[0068] In the position information 1500, a channel indicating whether or not an internal defect exists is set in each mesh. In Fig. 6, meshes in which a channel in which an internal defect exists is set are shown shaded. An internal defect 1504 indicates an internal defect in which a foreign object has been mixed between layers 1502 and 1503. As shown in Fig. 6, in the case of an internal defect caused by the mixing of a foreign object, a channel in which an internal defect exists is set in each of the meshes adjacent to each other across the boundary between the layers.

[0069] 5, the pre-learning unit 22 pre-learns an estimation model based on training data, and stores the pre-learned estimation model in the model storage unit .

[0070] The feature amount receiving unit 23 receives feature data from the temperature measuring device 1. When the feature amount receiving unit 23 receives the feature data related to the CFRP laminate to be learned, the feature amount receiving unit 23 sends the feature data to the additional data generating unit 24. When the feature amount receiving unit 23 receives the feature data related to the CFRP laminate to be estimated, the feature amount receiving unit 23 sends the feature data to the defect position estimating unit 27.

[0071] The additional data generation unit 24 generates additional data. The additional data includes feature data based on temperature changes in the CFRP laminate and position information of internal defects in the laminate. The additional data generation unit 24 receives the feature data from the feature amount receiving unit 23. The additional data generation unit 24 receives input of the position information of the internal defects in response to a user operation. The additional data generation unit 24 generates additional data by associating the received feature data with the position information of the internal defects.

[0072] Based on the additional data, the transfer learning unit 25 performs transfer learning on the estimation model pre-trained by the pre-learning unit 22. The transfer learning unit 25 updates the estimation model stored in the model storage unit 26 with the estimation model obtained through transfer learning.

[0073] A trained estimation model is stored in the model storage unit 26. The trained estimation model is an estimation model generated by the pre-learning unit 22 and updated by the transfer learning unit 25.

[0074] The defect position estimation unit 27 inputs the feature data received from the feature receiving unit 23 into the estimation model read out from the model storage unit 26, thereby calculating an estimate of the position information of the internal defect in the CFRP laminate to be estimated.

[0075] The result output unit 28 outputs the estimation result of the internal defect to the display device 506 etc. The estimation result includes the estimated value of the position information of the internal defect calculated by the defect position estimation unit 27.

[0076] <Processing Procedure> A defect estimation method executed by the defect estimation system 100 will be described with reference to Fig. 7. Fig. 7 is a flowchart showing an example of the defect estimation method.

[0077] In step S1, the defect estimation system 100 executes a pre-learning process. In the pre-learning process, the defect estimation device 2 calculates feature data by numerical analysis based on the finite element method and generates learning data including the feature data. The defect estimation device 2 pre-learns an estimation model based on the learning data. As a result, the pre-learned estimation model is stored in the model storage unit 26 of the defect estimation device 2.

[0078] In step S2, the defect estimation system 100 executes a transfer learning process. In the transfer learning process, the temperature measurement device 1 measures temperature changes in the CFRP laminate to be learned and calculates feature data based on the temperature changes. The defect estimation device 2 generates additional data including the feature data calculated by the temperature measurement device 1. The defect estimation device 2 transfer learns the pre-calculated estimation model based on the additional data. As a result, the transfer-learned estimation model is stored in the model storage unit 26 of the defect estimation device 2.

[0079] In step S3, the defect estimation system 100 executes an estimation process. In the estimation process, the temperature measurement device 1 measures temperature changes in the CFRP laminate to be estimated, and calculates feature data based on the temperature changes. The defect estimation device 2 inputs the feature data calculated by the temperature measurement device 1 into an estimation model read from the model storage unit 26. The estimation model outputs an estimated value of position information of an internal defect in the CFRP laminate to be estimated. The defect estimation device 2 outputs an estimation result including the position information of the internal defect output by the estimation model.

[0080] <Pre-learning process> The pre-learning process (step S1 in FIG. 7) will be described in more detail with reference to FIG. 8. FIG. 8 is a flowchart showing an example of the pre-learning process.

[0081] In step S1-1, the learning data generation unit 21 of the defect estimation device 2 acquires position information of internal defects in the CFRP laminate. The learning data generation unit 21 may acquire a plurality of pieces of position information indicating various internal defects. The learning data generation unit 21 may accept input of internal defect position information in response to a user operation. The learning data generation unit 21 may read out internal defect position information stored in advance in a storage device.

[0082] In step S1-2, the learning data generation unit 21 of the defect estimation device 2 constructs a finite element model representing a CFRP laminate. The learning data generation unit 21 constructs finite element models representing CFRP laminates having various internal defects based on the position information acquired in step S1-1. The learning data generation unit 21 may also construct finite element models representing CFRP laminates having no internal defects at a predetermined rate.

[0083] In step S1-3, the learning data generation unit 21 of the defect estimation device 2 calculates feature data based on each of the finite element models constructed in step S1-2. The learning data generation unit 21 may calculate the surface principal stress sum distribution of the CFRP laminate by numerical analysis based on the finite element method.

[0084] In step S1-4, the learning data generation unit 21 of the defect estimation device 2 adds noise to the feature data calculated in step S1-3. The learning data generation unit 21 may add noise to regions corresponding to reinforcing fibers contained in the CFRP laminate. The learning data generation unit 21 may generate multiple pieces of feature data from one piece of feature data by adding multiple different types of noise. In this embodiment, the learning data generation unit 21 generates five pieces of feature data from one piece of feature data by adding five types of noise.

[0085] In step S1-5, the learning data generation unit 21 of the defect estimation device 2 generates learning data by associating the feature data to which noise was added in step S1-4 with the position information of the internal defect acquired in step S1-1. The learning data generation unit 21 may standardize the feature data included in the learning data with respect to the entire learning data. The learning data generation unit 21 sends the learning data to the pre-learning unit 22.

[0086] In this embodiment, as an example, training data including 6095 pieces of feature data is generated, of which 3000 pieces are feature data based on CFRP laminates without internal defects, and the remaining 3095 pieces are feature data based on CFRP laminates with internal defects.

[0087] In step S1-6, the pre-learning unit 22 of the defect estimation device 2 receives the learning data from the learning data generation unit 21. The pre-learning unit 22 pre-learns an estimation model based on the received learning data.

[0088] In this embodiment, the estimation model is a convolutional neural network that takes the surface principal stress sum distribution of a CFRP laminate as input and outputs estimated values ​​of position information of internal defects in the laminate. A convolutional neural network is a special neural network used to process data with a lattice structure. Examples of data with a lattice structure include time-series data in which samples acquired at equal time intervals are arranged one-dimensionally, or image data in which pixels are arranged two-dimensionally.

[0089] As an example, the pre-learning unit 22 may train a convolutional neural network as follows: The loss function uses binary intersection error; the optimization method uses Adam; the learning rate is set to 0.00007; the mini-batch size is 64; and the number of training iterations is 800. Note that regularization methods such as L2 regularization or batch normalization do not need to be used.

[0090] In step S1-7, the pre-learning unit 22 of the defect estimation device 2 stores the estimation model pre-learned in step S1-6 in the model storage unit .

[0091] <Transfer learning processing> The transfer learning process (step S2 in FIG. 7) will be described in more detail with reference to FIG. 9. FIG. 9 is a sequence diagram showing an example of the transfer learning process.

[0092] In step S2-1, the temperature measurement unit 12 of the temperature measurement device 1 measures the surface temperature of the CFRP laminate to be learned before changing the surface temperature of the laminate. The temperature measurement unit 12 generates a surface temperature image (hereinafter also referred to as a "pre-temperature change image") that represents the distribution of the measured surface temperature. The temperature measurement unit 12 sends the pre-temperature change image to the feature calculation unit 13.

[0093] In step S2-2, the temperature change unit 11 of the temperature measurement device 1 changes the temperature of the CFRP laminate to be learned. Specifically, a predetermined fixed external force is applied to the CFRP laminate to be learned.

[0094] In step S2-3, the temperature measurement unit 12 of the temperature measurement device 1 changes the surface temperature of the CFRP laminate to be learned, and then measures the surface temperature of the laminate. The temperature measurement unit 12 generates a surface temperature image (hereinafter also referred to as a "post-temperature change image") that represents the distribution of the measured surface temperature. The temperature measurement unit 12 sends the post-temperature change image to the feature calculation unit 13.

[0095] In step S2-4, the feature amount calculation unit 13 of the temperature measurement device 1 receives the pre-temperature change image and the post-temperature change image from the temperature measurement unit 12. The feature amount calculation unit 13 calculates feature data of the CFRP laminate to be learned based on the pre-temperature change image and the post-temperature change image. The feature amount calculation unit 13 transmits the calculated feature data to the defect estimation device 2.

[0096] Specifically, the feature amount calculation unit 13 generates a temperature change image representing a change in the surface temperature of the CFRP laminate to be learned, based on the before-temperature change image and the after-temperature change image. The feature amount calculation unit 13 may calculate the difference between the before-temperature change image and the after-temperature change image. That is, the feature amount calculation unit 13 may subtract the surface temperature represented by the before-temperature change image from the surface temperature represented by the after-temperature change image for each pixel included in each image.

[0097] The feature amount calculation unit 13 may calculate feature data based on the temperature change image. The feature amount calculation unit 13 may calculate a surface principal stress sum distribution as feature data based on the temperature change image. The feature amount calculation unit 13 may obtain the surface temperature change distribution by dividing the temperature change image into a predetermined number of meshes and calculating the temperature change amount corresponding to each mesh. The temperature change amount of a mesh may be the temperature change amount at a representative point (such as the center) of the mesh, or may be the average temperature change amount within the mesh. The method of dividing the mesh is the same as for the position information of the internal defect.

[0098] The feature calculation unit 13 may calculate the surface principal stress sum distribution by applying Kelvin's theory to the surface temperature change distribution. Kelvin's theory gives the relationship between the temperature change ΔT due to the thermoelastic effect and the change Δσ in the principal stress sum as ΔT = -kTΔσ, where k is the thermoelastic coefficient and T is the absolute temperature.

[0099] In step S2-5, the feature amount receiving unit 23 of the defect estimation device 2 receives the feature data of the CFRP laminate to be learned from the temperature measuring device 1. The feature amount receiving unit 23 sends the feature data received from the temperature measuring device 1 to the additional data generating unit 24.

[0100] In step S2-6, the additional data generation unit 24 of the defect estimation device 2 acquires position information of internal defects in the CFRP laminate to be learned. The additional data generation unit 24 may accept input of the position information of the internal defects in response to a user operation. The additional data generation unit 24 may read out position information of the internal defects stored in advance in a storage device.

[0101] Steps S2-1 to S2-6 are repeatedly executed for each CFRP laminate to be learned, thereby inputting feature data and internal defect position information corresponding to each CFRP laminate to be learned into the additional data generator 24.

[0102] In step S2-7, the additional data generation unit 24 of the defect estimation device 2 generates additional data by associating the feature data received in step S2-5 with the position information of the internal defect acquired in step S2-6. The additional data generation unit 24 may standardize the feature data included in the additional data with respect to the entire additional data. The additional data generation unit 24 sends the additional data to the transfer learning unit 25.

[0103] In this embodiment, as an example, training data including 22 pieces of feature data is generated. Of these, 10 pieces of feature data are based on CFRP laminates without internal defects. The remaining 12 pieces of feature data are based on CFRP laminates with internal defects. Therefore, the amount of additional data may be significantly smaller than that of the training data.

[0104] In step S2-8, the transfer learning unit 25 of the defect estimation device 2 receives the additional data from the additional data generation unit 24. The transfer learning unit 25 performs transfer learning on the estimation model based on the received additional data.

[0105] As an example, the transfer learning unit 25 may train a convolutional neural network as follows: The loss function uses binary intersection error; the optimization method uses Adam; the learning rate is set to 0.00007; the mini-batch size is set to 1; and the number of training iterations is set to 1000. Note that regularization methods such as L2 regularization or batch normalization do not need to be used.

[0106] In step S2-9, the transfer learning unit 25 of the defect estimation device 2 updates the estimation model stored in the model storage unit 26 with the estimation model obtained through transfer learning in step S2-8.

[0107] <<Estimation process>> The estimation process (step S3 in FIG. 7) will be described in more detail with reference to FIG. 10. FIG. 10 is a sequence diagram showing an example of the estimation process.

[0108] In step S3-1, the temperature measurement unit 12 of the temperature measurement device 1 measures the surface temperature of the CFRP laminate to be estimated before changing the surface temperature of the laminate. The temperature measurement unit 12 generates a pre-temperature change image that represents the distribution of the measured surface temperature. The temperature measurement unit 12 sends the pre-temperature change image to the feature calculation unit 13.

[0109] In step S3-2, the temperature change unit 11 of the temperature measurement device 1 changes the temperature of the CFRP laminate that is the estimation target. Specifically, a predetermined fixed external force is applied to the CFRP laminate that is the estimation target.

[0110] In step S3-3, the temperature measurement unit 12 of the temperature measurement device 1 changes the surface temperature of the CFRP laminate to be estimated, and then measures the surface temperature of the laminate. The temperature measurement unit 12 generates a post-temperature change image that represents the distribution of the measured surface temperatures. The temperature measurement unit 12 sends the post-temperature change image to the feature calculation unit 13.

[0111] In step S3-4, the feature amount calculation unit 13 of the temperature measurement device 1 receives the pre-temperature change image and the post-temperature change image from the temperature measurement unit 12. The feature amount calculation unit 13 calculates feature data of the CFRP laminate to be estimated based on the pre-temperature change image and the post-temperature change image. The feature amount calculation unit 13 transmits the calculated feature data to the defect estimation device 2.

[0112] In step S3-5, the feature amount receiving unit 23 of the defect estimation device 2 receives the feature data of the CFRP laminate to be estimated from the temperature measurement device 1. The feature amount receiving unit 23 sends the feature data received from the temperature measurement device 1 to the defect position estimation unit 27.

[0113] In step S3-6, the defect position estimation unit 27 of the defect estimation device 2 receives the feature data from the feature amount receiving unit 23. The defect position estimation unit 27 reads out an estimation model from the model storage unit 26. The defect position estimation unit 27 inputs the received feature data into the read out estimation model, thereby calculating an estimate of the position information of the internal defect in the CFRP laminate that is the estimation target. The defect position estimation unit 27 sends the estimate of the position information of the internal defect to the result output unit 28.

[0114] In step S3-7, the result output unit 28 of the defect estimation device 2 receives the estimated value of the position information of the internal defect from the defect position estimation unit 27. The result output unit 28 outputs the estimated result of the position information of the internal defect to the display device 506 or the like. The estimation result includes the estimated value of the position information of the internal defect calculated by the defect position estimation unit 27.

[0115] <Test Results> An evaluation test was conducted to evaluate the performance of the defect estimation system 100. The evaluation results of the defect estimation system 100 will be described with reference to FIGS.

[0116] In the evaluation tests, a rectangular CFRP laminate measuring 25mm x 75mm was used. Internal defects were simulated by inserting a 0.2mm thick fluororesin sheet between any of the sixth to tenth layers. The fluororesin sheets were prepared in several sizes and inserted in different positions. The sizes of the fluororesin sheets prepared were 3mm x 3mm, 3mm x 5mm, 5mm x 10mm, 10mm x 10mm, 10mm x 15mm, and 15mm x 15mm.

[0117] Fig. 11 is a diagram showing a first example of test results. Fig. 11 shows the estimation results when a CFRP laminate C1 having an internal defect D is used as the estimation target. The internal defect D is a foreign object mixed in between the 7th and 8th layers. Fig. 11 compares the correct value (Ground Truth) of the position information with the estimated value (Prediction) of the position information between each of the 7th to 10th layers. The filled-in area of ​​each position information is a channel indicating the presence of an internal defect.

[0118] As shown in Fig. 11, the ground truth of the position information indicates that an internal defect exists in the region corresponding to the internal defect D between the 7th and 8th layers (8-7 layer). The predicted value of the position information (Prediction) indicates that an internal defect exists in the region roughly corresponding to the ground truth.

[0119] Fig. 12 is a diagram showing a second example of test results. Fig. 12 shows the estimation results when a CFRP laminate C2 having no internal defects was used as the estimation target. As with Fig. 11, Fig. 12 compares the ground truth of the position information with the prediction of the position information between each of the seventh to tenth layers.

[0120] As shown in Fig. 12, the ground truth of the position information indicates that no internal defects exist between any of the layers. The predicted value of the position information, like the ground truth, also indicates that no internal defects exist between any of the layers.

[0121] 11 and 12 show that it is possible to accurately determine whether or not the CFRP laminate being estimated has an internal defect. Also, Fig. 11 and 12 show that it is possible to accurately estimate the position information of the internal defect.

[0122] Fig. 13 is a diagram showing an example of estimation accuracy. In this evaluation test, the prediction results were quantitatively evaluated using the following two indices.

[0123] The first index is the coincidence rate (R) of defect plane position prediction. The plane position coincidence rate R is the proportion of areas where the ground truth of the position information indicates the presence of an internal defect (hereinafter referred to as "defect area") where the estimated value (Prediction) of the position information indicates the presence of an internal defect. Specifically, the plane position coincidence rate R is the value obtained by dividing the number of pixels in the defect area where it is estimated that an internal defect exists by the number of pixels in the defect area. In this evaluation test, the number of pixels where it is estimated that an internal defect exists was counted in an area extending 2 mm square from the defect area.

[0124] The second indicator is the defect insertion layer prediction (P). The defect insertion layer prediction P is the ratio of pixels estimated to have internal defects in each layer to pixels estimated to have internal defects in the entire stack. Specifically, the defect insertion layer prediction P is the value obtained by dividing the number of pixels estimated to have internal defects in each layer by the sum of the number of pixels estimated to have internal defects in each layer.

[0125] Figure 13 shows the case with the highest and lowest planar positional agreement rates R among the CFRP laminates used in the test. The case with the highest planar positional agreement rate R was a case using a CFRP laminate with a 10 mm x 15 mm foreign object mixed in between the 7th and 8th layers. The case with the lowest planar positional agreement rate R was a case using a CFRP laminate with a 3 mm x 3 mm foreign object mixed in between the 9th and 10th layers.

[0126] As shown in Figure 13, the planar position coincidence rate R is numerically inferior when the size of the internal defect (foreign matter) is small, but it is generally within the size of the defect area. Also, the defect layer prediction P is highest in all layers where internal defects exist. Figure 13 shows that the position information of internal defects can be estimated with high quantitative accuracy.

[0127] <Effects of the embodiment> A defect estimation device 2 according to an embodiment of the present disclosure pre-trains an estimation model based on training data including feature data obtained by analyzing temperature changes occurring on the surface of a CFRP laminate based on a finite element model, transfer-trains the estimation model based on additional data including feature data obtained by measuring temperature changes occurring on the surface of the laminate, and inputs the feature data based on the temperature changes occurring on the surface of the laminate to the estimation model, thereby estimating position information of internal defects in the laminate. Because the defect estimation device 2 transfer-trains the estimation model pre-trained with training data including feature data based on a finite element model using additional data including feature data measured using real samples, it is possible to train a highly accurate estimation model even with a small number of real samples. According to one aspect, this embodiment enables accurate estimation of internal defects in a fiber-reinforced plastic laminate.

[0128] The feature data included in the training data may have noise added to positions corresponding to the reinforcing fibers included in the CFRP laminate. Feature data acquired from real samples often contains noise at positions corresponding to the reinforcing fibers. According to this embodiment, feature data similar to the feature data acquired from real samples can be generated.

[0129] The position information may include information indicating whether or not an internal defect exists in each mesh obtained by dividing each layer of the CFRP laminate into a predetermined number of meshes. The meshes may be obtained by dividing each layer of the CFRP laminate in the depth direction. The finite element model may have an internal defect indicating a foreign object between layers of the CFRP laminate. The position information may indicate the presence of an internal defect in each of adjacent meshes across the boundary between layers of the laminate. According to this embodiment, the position information of a foreign object mixed between layers can be estimated with high accuracy.

[0130] The finite element model may have an internal defect representing a foreign object, and a predetermined elastic modulus may be set for a region corresponding to the internal defect. According to this embodiment, a surface principal stress sum distribution similar to the surface principal stress sum distribution obtained from an actual sample can be calculated by numerical analysis.

[0131] The training data may include a predetermined proportion of feature data based on a finite element model that represents a CFRP laminate that has no internal defects. According to this embodiment, the estimation model learns the features of a CFRP laminate that has no internal defects, and therefore, the probability of erroneously estimating a CFRP laminate that does not have an internal defect as having an internal defect can be reduced.

[0132] [supplement] Each function of the above-described embodiments can be realized by one or more processing circuits. Here, the term "processing circuit" in this specification includes a processor programmed to perform each function by software, such as a processor implemented by an electronic circuit, as well as devices such as an ASIC (Application Specific Integrated Circuit), a DSP (Digital Signal Processor), an FPGA (Field Programmable Gate Array), and conventional circuit modules designed to perform each of the above-described functions.

[0133] Although the embodiments of the present invention have been described in detail above, the present invention is not limited to these embodiments, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims. [Explanation of symbols]

[0134] 100: Defect estimation system 1:Temperature measuring device 2: Defect estimation device 11: Temperature change section 12:Temperature measurement part 13: Feature calculation unit 21: Learning data generation unit 22: Pre-learning section 23: Feature reception unit 24: Additional data generation unit 25: Transfer learning unit 26: Model memory section 27: Defect location estimation unit 28: Result output section

Claims

1. a pre-learning unit configured to learn a model that outputs position information when feature data is input, based on learning data including feature data generated by analyzing temperature changes occurring on the surface of a fiber-reinforced plastic laminate based on a finite element model representing the laminate and position information of internal defects in the laminate; and a transfer learning unit configured to update the model trained by the pre-learning unit based on additional data including feature data generated by measuring temperature changes occurring on the surface of the laminate and position information of internal defects of the laminate; a feature amount receiving unit configured to receive input of feature data based on a temperature change occurring on the surface of the laminate; a defect position estimation unit configured to estimate position information of the internal defect of the laminate by inputting the feature data received by the feature amount receiving unit into the model; A defect estimation device comprising:

2. 2. The defect estimation device according to claim 1, The feature data included in the learning data has noise added to positions corresponding to the reinforcing fibers included in the laminate. Defect estimation device.

3. 2. The defect estimation device according to claim 1, the position information includes information indicating whether or not the internal defect exists in each of the meshes obtained by dividing each layer of the laminate into a predetermined number of meshes; Defect estimation device.

4. 4. The defect estimation device according to claim 3, The mesh is obtained by dividing each layer of the laminate in the depth direction. Defect estimation device.

5. 5. The defect estimation device according to claim 4, the finite element model has the internal defect indicative of an inclusion between layers of the laminate; The position information indicates that the internal defect exists in each of the meshes adjacent to each other across a boundary between layers of the laminate. Defect estimation device.

6. 6. A defect estimation device according to claim 1, the finite element model has the internal defect indicating a foreign object, and a predetermined elastic modulus is set in a region corresponding to the internal defect; Defect estimation device.

7. 6. A defect estimation device according to claim 1, the learning data includes a predetermined proportion of the feature data based on the finite element model that represents the laminate that does not have the internal defect; Defect estimation device.

8. A defect estimation system including a temperature measurement device and a defect estimation device, The temperature measuring device is a temperature change unit configured to cause a temperature change in the fiber reinforced plastic laminate; a temperature measuring unit configured to measure a temperature change occurring on the surface of the laminate; a feature amount calculation unit configured to calculate feature data based on the temperature change; Equipped with the defect estimation device, a pre-learning unit configured to learn a model that outputs position information when inputting feature data based on learning data including feature data generated by analyzing temperature changes occurring on the surface of the laminate based on a finite element model representing the laminate and position information of internal defects of the laminate; and a transfer learning unit configured to update the model trained by the pre-learning unit based on additional data including the feature data generated by the temperature measuring device and position information of internal defects in the laminate; a feature amount receiving unit configured to receive input of the feature data generated by the temperature measuring device; a defect position estimation unit configured to estimate position information of the internal defect of the laminate by inputting the feature data received by the feature amount receiving unit into the model; A defect estimation system comprising:

9. The computer a pre-learning procedure for learning a model that outputs position information when feature data is input, based on learning data including feature data generated by analyzing temperature changes occurring on the surface of a fiber-reinforced plastic laminate based on a finite element model representing the laminate, and position information of internal defects in the laminate; a transfer learning procedure for updating the model trained in the pre-learning procedure based on additional data including feature data generated by measuring temperature changes occurring on the surface of the laminate and position information of internal defects of the laminate; a receiving step of receiving input of characteristic data based on a temperature change occurring on the surface of the laminate; an estimation step of estimating position information of the internal defect of the laminate by inputting the feature data received in the receiving step into the model; A defect estimation method that performs

10. On the computer, a pre-learning procedure for learning a model that outputs position information when feature data is input, based on learning data including feature data generated by analyzing temperature changes occurring on the surface of a fiber-reinforced plastic laminate based on a finite element model representing the laminate, and position information of internal defects in the laminate; a transfer learning procedure for updating the model trained in the pre-learning procedure based on additional data including feature data generated by measuring temperature changes occurring on the surface of the laminate and position information of internal defects of the laminate; a receiving step of receiving input of characteristic data based on a temperature change occurring on the surface of the laminate; an estimation step of estimating position information of the internal defect of the laminate by inputting the feature data received in the receiving step into the model; A program to execute.

Images