Adhesive force estimation system and adhesive force estimation method

Abstract

The present invention proposes to estimate the strength of adhesive bonding of an individual adhesive member using both information about the adherend and physical property values of the adhesive. An adhesive force estimation system according to the present invention comprises: an input interface that receives an input of mechanical property values of a solidified adhesive and information about the surface roughness of an adherend; a storage unit that stores a trained model that has been trained using training data including, as inputs, combinations of test results of mechanical property values of solidified adhesive test pieces and information about the surface roughness of adherends, and as outputs, adhesive strengths; and an adhesive strength prediction unit that estimates an adhesive strength using the trained model on the basis of the input to the input interface (see Fig. 4).

Description

Adhesion strength estimation system and adhesion strength estimation method

[0001] The present invention relates to an adhesive strength estimation system, an adhesive strength estimation method, and a computer.

[0002] The increased use of lightweight structural materials aimed at improving the energy efficiency of mobility, along with the accompanying widespread adoption of adhesive bonding, is expected.

[0003] Strength reliability is an important factor for structural materials used in vehicles, etc., and accurate assessment of adhesive bond strength is necessary to determine the soundness of adhesive bonds.

[0004] The strength of adhesive bonds can be determined by measuring the strength using tensile tests, etc. However, this method destroys the adhesive bonded components, so a non-destructive method of strength determination is required.

[0005] Furthermore, from the viewpoint of industrial application, it is necessary to estimate the strength of each adhesive bond, rather than just the average strength.

[0006] Non-Patent Document 1 attempts to estimate adhesive strength using strength evaluation data of adhesive members by the finite element method. The numerical simulation technique using the finite element method can non-destructively give the average strength of adhesive members.

[0007] Patent Document 1 proposes a method for non-destructively estimating the fracture mode by accumulating data on adhesive bonding. This method estimates the fracture mode by providing measured adhesive strength as an input value.

[0008] Non-Patent Document 2 claims that, based on the technology of Patent Document 1, adhesive strength can be estimated non-destructively by inputting the type of adhesive, the material of the adherend, surface treatment conditions, etc.

[0009] Japanese Patent Application Laid-Open No. 2022-37646

[0010] Z. Gu, Composites Part B 217 (2021) 108894, pp.2-3 Panasonic Holdings Co., Ltd., "AI Diagnosis of Adhesion", [online], [Retrieved February 9, 2023], Internet, URL: https: / / holdings.panasonic / jp / corporate / pac / reliability / durability / adhesiveai.html

[0011] The numerical simulation method described in Non-Patent Document 1 does not allow evaluation of the adhesive strength of individual adhesive bonds.

[0012] According to the estimated strength shown in Non-Patent Document 2, the actual measured value is about 4 MPa, but the estimated strength is sometimes about 12 MPa, which is an error of three times the actual measured value. This corresponds to, for example, an adhesive bond predicted to withstand a load of 30 kg actually only withstanding a load of 10 kg. In order to further improve the reliability of adhesive bonds, it is necessary to improve the accuracy of prediction of adhesive strength.

[0013] In order to solve the above problems, the present inventors have devised a method for estimating the strength of adhesive bonds for individual adhesive members by using information on the adherend and physical properties of the adhesive in combination.

[0014] An example of an adhesive strength estimation system according to the present invention comprises: an input interface that accepts input of mechanical property values ​​of the solidified adhesive and information on the surface roughness of the adherend; a memory unit that stores a trained model trained using training data that takes as input a combination of test results of the mechanical property values ​​of a test piece of the solidified adhesive and information on the surface roughness of the adherend, and outputs adhesive strength; and an adhesive strength estimation unit that uses the trained model to estimate adhesive strength based on the input to the input interface.

[0015] In one example of the adhesive strength estimation method according to the present invention, the adhesive strength estimation system described above estimates the adhesive strength.

[0016] An example of a computer according to the present invention stores data correlating the test results of mechanical property values ​​of test pieces of solidified adhesive with information on the surface roughness of the adherend, and adhesive strength.

[0017] According to the present invention, it is possible to estimate the strength of an adhesive bond with high accuracy.

[0018] FIG. 1 is a diagram showing the basic configuration of an adhesive strength estimation system in embodiment 1 of the present invention. FIG. 2 is a procedure for estimating adhesive strength using the adhesive strength estimation system. FIG. 3 is a procedure for performing only estimation of adhesive strength. FIG. 4 is a configuration diagram that specifies each element more specifically. FIG. 5 is an example of the configuration of training data. FIG. 6 is a diagram showing contact of an adhesive with an adherend having a gentle slope in the surface unevenness. FIG. 7 is a diagram showing contact of an adhesive with an adherend having a steep slope in the surface unevenness. FIG. 8 is a diagram of adhesive bonding. FIG. 9 is a diagram showing measurement positions of line roughness. FIG. 10 is an estimation result of adhesive strength. FIG. 11 is an explanatory diagram of thickness measurement points. FIG. 12 is an estimation result of adhesive strength using dataset 4. FIG. 13 is an estimation result of adhesive strength using dataset 1. FIG. 14 is a relationship between the usage rate of information related to RΔq and estimation accuracy. FIG. 15 is a data format of dataset 7. FIG. 16 is a data format of dataset 8. FIG. 17 is a data processing procedure.

[0019] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Advances in information science technology, such as machine learning, have led to an active movement to apply it to materials development. The movement to apply machine learning and other technologies to materials development is generally called materials informatics (MI). The present invention was born from research into MI.

[0020] Typically, one embodiment of MI applies machine learning to create a mathematical model to estimate the physical properties of a material based on its composition, process, etc. This method is known to be able to estimate material properties, and can also be used to estimate the strength of adhesive bonds.

[0021] The concept of adhesive bond strength is generally understood in terms of the work of adhesion. The work of adhesion is the energy required to bond the adhesive and adherend, and is equal to the work required to separate the adhesive and adherend and create two new surfaces. This relationship is expressed by a formula known as the Dupre equation. However, the work of adhesion is purely a thermodynamic concept, and is the total work required to bring about complete separation of the adhesive interface.

[0022] In order to discuss the strength of practical adhesive bonds, it is necessary to consider the microscopic conditions that occur when the adhesive interface breaks down. That is, in the case of actual adhesive bond fracture due to tension, if there are areas where the bonding strength of the adhesive interface is locally weakened due to, for example, surface irregularities, application of stress first causes microscale fracture in these areas. These cracks then gradually grow and progress in response to the stress, forming delamination regions, and continued stress eventually leads to total fracture.

[0023] In the fracture process of such microscopic adhesive bonds, the propagation of peeling due to stress is the main process, and the energy required to propagate the peeling is much smaller than the work of adhesion, meaning that the work of adhesion is not appropriate for practical purposes.

[0024] Numerical simulation using the finite element method (FEM) is a method for estimating adhesive strength that takes into account the peeling process at an actual adhesive interface. This method can be used to simulate crack propagation and calculate the adhesive strength in a pseudo-manner. However, the adhesive strength calculated using the FEM is not sufficiently accurate, and it cannot be used to estimate the adhesive strength of individual components.

[0025] Therefore, the inventors have devised a means for achieving high-precision estimation based on machine learning methods in order to estimate with high accuracy the adhesive strength of adhesively bonded individual components, taking into account the peeling process at the actual adhesive interface.

[0026] The adhesive strength estimation method is explained in detail below. The adhesive strength estimation method is divided into two steps. One is to build an estimation model, and the other is to estimate the adhesive strength using the estimation model.

[0027] 1 is a diagram showing the basic configuration of an adhesive strength estimation system 109 according to embodiment 1. The adhesive strength estimation system 109 estimates adhesive strength by the processing described in this specification. Note that in this specification, "estimation" can be replaced with "prediction."

[0028] The adhesive force estimation system 109 has a hardware configuration as a known computer, and includes, for example, a calculation means and a storage means. The calculation means includes, for example, a processor, and the storage means includes, for example, a storage medium such as a semiconductor memory device or a magnetic disk device. A part or all of the storage medium may be a non-transitory storage medium.

[0029] The storage means may store a program, and the processor may execute the program, causing the computer to function as the adhesive strength estimation system 109.

[0030] The adhesive strength estimation system 109 is a device including a memory unit 104, an estimation model construction unit 105, an adhesive strength estimation unit 106, an input interface 103, and an output interface 107. These components are realized by the cooperation of the calculation means and memory means of a computer.

[0031] The input interface 103 is a device for inputting training data 101 to be used for training the estimation model and experimental data 102 for determining the adhesive strength 108 of an adhesive bond whose adhesive strength is unknown. There are no limitations on the method for inputting the data, and data stored in a file may be read from a reading device, or the data may be input from an input device consisting of a monitor and keyboard.

[0032] The storage unit 104 can store the teacher data 101, the experimental data 102, and the estimation model constructed by the estimation model construction unit 105.

[0033] The estimation model construction unit 105 constructs an estimation model of adhesive strength by machine learning based on the training data 101 .

[0034] The adhesive strength estimation unit 106 calculates the adhesive strength 108 corresponding to the experimental data 102 using the adhesive strength estimation model constructed by the estimation model construction unit 105. The adhesive strength 108 is output from an output interface 107. There are no limitations on the method of outputting the data, and the data may be written directly to a file on a storage medium, displayed on a monitor, or transmitted via a communication network.

[0035] The procedure for estimating adhesive strength using the adhesive strength estimation system 109 is shown in Figure 2. First, training data necessary for constructing an estimation model is prepared (step S1). Next, the training data is input into the adhesive strength estimation system 109 (step S2), and an estimation model is constructed based on the training data (step S3). After that, experimental data is prepared (step S4) for determining the adhesive strength of an adhesive bond whose adhesive strength is unknown. The experimental data is input (step S5), and the constructed estimation model is used to calculate and estimate the adhesive strength corresponding to the experimental data (step S6). Finally, the adhesive strength is output (step S7), completing the process.

[0036] When estimating adhesive strength, it is not necessary to reconstruct an estimation model each time. In other words, once an estimation model is constructed, it can be used to estimate adhesive strength any number of times. The procedure for this case is shown in FIG. 3. Since an estimation model has already been constructed, it is sufficient to start by preparing experimental data (step S11). As shown in the figure, the experimental data is input (step S12), and then the adhesive strength corresponding to the experimental data is estimated by calculation using the estimation model stored in the memory unit (step S13), and the adhesive strength is output (step S14), completing the process.

[0037] A more specific configuration diagram of each element is shown in Figure 4. First, training data 401 is prepared in which measurement data 402 of the mechanical properties (mechanical property values) of the solidified adhesive and measurement data 403 of the surface line roughness of the adherend are used as explanatory variables, and measurement data 404 of the adhesive strength of the corresponding adhesive bond is used as the objective variable.

[0038] The estimation model construction device 405 has a hardware configuration as a known computer, and includes, for example, a calculation means and a storage means. The calculation means includes, for example, a processor, and the storage means includes, for example, a storage medium such as a semiconductor memory device or a magnetic disk device.

[0039] The input interface 406 of the estimation model construction device 405 inputs this teacher data 401 and sends it to the estimation model construction unit 407. The teacher data 401 is stored in a memory unit (not shown) of the estimation model construction device 405. At this point, the estimation model construction device 405 stores the teacher data 401, i.e., data in which a combination of test results for mechanical property values ​​using test pieces of the solidified adhesive, information on the surface roughness of the adherend, and adhesive strength are associated with each other. Using data with such a structure makes it possible to appropriately perform machine learning, which will be described later.

[0040] The estimation model construction unit 407 constructs an estimation model based on machine learning, and outputs a trained model 409 from an output interface 408 .

[0041] Although not shown in Fig. 4, the trained model 409 is stored in the memory unit 104 (see Fig. 1) of the adhesive strength estimation system 109. That is, the memory unit 104 stores the trained model 409 trained with training data 401 that takes as input a combination of test results of mechanical property values ​​using test pieces of solidified adhesive and information on the surface roughness of the adherend, and outputs adhesive strength.

[0042] In order to estimate the adhesive strength of an unknown adhesive bond, experimental data 410 consisting of mechanical property data 411 (mechanical property values) obtained by solidifying the same type of adhesive as the adhesive used in the adhesive bond and measuring its mechanical property values, and measurement data 412 of the surface line roughness of the adherend, is used by adhesive strength estimation system 109 to obtain an estimated adhesive strength 417.

[0043] Specifically, experimental data 410 is input from the input interface 103, and the adhesive strength is estimated by processing the experimental data 410 with the adhesive strength estimation unit 106 using the trained model 409 constructed based on the machine learning processing, and the adhesive strength 417 is output from the output interface 107. Here, the input interface 103 accepts input of the mechanical property values ​​of the solidified adhesive and information on the surface roughness of the adherend, and the adhesive strength estimation unit 106 estimates the adhesive strength using the trained model 409 based on the input to the input interface 103.

[0044] In the example of FIG. 4, line roughness is used as the surface roughness, but area roughness may be used instead of or in addition to line roughness.

[0045] There are various methods for solidifying adhesives to measure their mechanical properties. For example, it is preferable to use a silicone or polyethylene mold to obtain JIS K 7139 Type A1 test specimen dimensions, pour in an adhesive mixture of the base agent and curing agent, and then use a desiccator or similar to mold the adhesive into a dumbbell test specimen at the specified curing temperature and time.

[0046] It is preferable that the adhesive solidification method be consistent across the training data and experimental data. Mixing data solidified using different methods is not preferable, as it reduces estimation accuracy. This also applies to all data below, and it is preferable that the test method be consistent across the training data and experimental data.

[0047] There are various methods for measuring the mechanical properties of solidified adhesives. For example, it is recommended to measure the mechanical properties according to JIS K-7161. From the stress-strain diagram measured in this way, mechanical properties such as tensile stress, tensile stress at yield, tensile strength, tensile modulus, and elongation at break can be obtained.

[0048] It is preferable to prepare multiple samples of the solidified adhesive and measure the mechanical properties of each. The number of repeated preparations is preferably 5 or more. It is preferable to remove the maximum and minimum values ​​of the obtained mechanical properties and adopt the average value of the mechanical properties of the remaining samples as the mechanical property value of the adhesive.

[0049] There are various methods for measuring the surface roughness of adherends. One example is a one-shot 3D profilometer. Specifically, the adherend surface is degreased by wiping with hexane or acetone, and then the surface roughness is measured. The following equations can be used to define surface roughness: maximum peak height, maximum valley depth, maximum height, average height, arithmetic mean roughness, maximum cross-sectional height, root-mean-square height, skewness, kurtosis, mean element length, root-mean-square slope, and ten-point mean roughness. Adhesive roughness is measured on both surfaces of the two adherends that form an adhesive bond.

[0050] The line used to measure line roughness is preferably a straight line within the adhesive area on the surface of the adherend. The end points of the line should preferably be the boundary of the adhesive area. The line should also preferably pass through the center of the adhesive area, and if tensile shear strength is considered as adhesive strength, it should preferably be a straight line parallel to the shear direction. However, these are just examples, and it is not necessary to meet these conditions.

[0051] There are various methods for preparing adhesive joint samples to obtain strength data for adhesive bonds. Types of joints used to join two adherends include lap joints, butt joints, double butt joints, ripped lap joints, ripped joints, stud joints, stepped lap joints, and so on. While any type of joint can be selected, it is preferable to use strength values ​​for a single type of joint within the same data set. Mixing strength values ​​for different types of joints within the same data set is not recommended, as it reduces estimation accuracy.

[0052] In this embodiment, a regression method based on supervised machine learning is used to construct an estimation model. That is, a means is used that can construct a regression equation that calculates a target variable using explanatory variables. There are various regression algorithms, such as linear multiple regression, support vector regression, random forest regression, regression using a neural network, and regression based on a gradient boosting method, and any method may be used.

[0053] When constructing an estimation model, it is preferable to take care to prevent the model from over-learning the training data. One such technique is cross-validation. In cross-validation, K is set to be greater than or equal to 2 and less than the number of records in the training data, and the training data is divided into K parts. One of the divided parts is used as validation data, and the remaining (K-1) parts are used as training data for training. This results in K models with different estimation accuracies being constructed, and the average of these estimation accuracies is recorded. Next, the hyperparameters are changed, and K models are constructed for the K divided data in the same way, and the average estimation accuracy is calculated. If this value is higher than the average estimation accuracy from the previous time, this model is adopted as a candidate for the generalization model. By repeating this process, an optimized generalization model is obtained.

[0054] Hyperparameters are parameters that are determined in advance before training a machine learning model. In the case of random forest regression, the hyperparameters include the number of decision trees, which are weak learners, and the number of branches. There are various methods for searching for hyperparameters, such as grid search and Bayesian inference, but any method can be used.

[0055] The constructed estimation model is used to estimate the adhesive strength. That is, experimental data is input to the estimation model to calculate an estimated value of the adhesive strength.

[0056] According to this embodiment, since the adhesive strength is estimated from the explanatory variables included in the experimental data, it is not necessary to measure the adhesive strength of the adhesive bond included in the experimental data for the purpose of estimation. If an experimental value of the adhesive strength is available for the adhesive bond included in the experimental data, the estimation accuracy can be calculated by comparing it with the estimated value of the adhesive strength.

[0057] The estimation accuracy is calculated by applying the constructed estimation model to the experimental data and calculating the difference between the calculated adhesive strength and the experimental value of the adhesive strength of the adhesive bond included in the experimental data. When the difference is evaluated using a correlation coefficient, it can be evaluated dimensionlessly, and when the difference is evaluated using the root mean square error, it can be evaluated in units of MPa.

[0058] Figure 5 shows an example of the structure of the training data. In this example, the training data is structured in a matrix format, but any format is acceptable as long as the objective variable is linked to the explanatory variables. For ease of understanding, the first column lists the sample numbers of the adhesively bonded samples, but these sample numbers are not used in model construction.

[0059] The second column contains the adhesive strength. For example, adhesive strength can be expressed as tensile shear strength. This is the column where the measured adhesive strength values ​​corresponding to the sample numbers are entered, and this is the target variable.

[0060] All columns from the third onwards are used as explanatory variables. Here, breaking strength, breaking elongation, elastic modulus, maximum peak height, root mean square slope, and ten-point average roughness are entered as examples. Breaking strength may be expressed in tensile stress or tensile yield stress, breaking elongation may be expressed in tensile elongation, and elastic modulus may be expressed in tensile modulus. These values ​​correspond to the sample number specified in the corresponding row, and are measured values ​​of the mechanical properties (breaking strength, breaking elongation, elastic modulus) of the adhesive used for that sample when solidified, and the line roughness (maximum peak height, root mean square slope, ten-point average roughness) of the adherend surface used for that sample.

[0061] As explanatory variables, in addition to the mechanical properties of the solidified adhesive and the measured values ​​of the line roughness of the adherend surface, any explanatory variables can be used. For example, since it is known that the thickness of the adhesive layer affects the adhesive strength, the thickness of the adhesive layer can also be added as an explanatory variable.

[0062] In the example of Figure 5, each training data (one record) includes data on a single adherend as an explanatory variable, i.e., the adhesive strength is expressed as the strength when two adherends of the same type are bonded together. However, each training data may also include data on two adherends as explanatory variables. For example, the breaking strength may include the breaking strength of a first adherend and the breaking strength of a second adherend. In such a configuration, the adhesive strength is expressed as the strength when two adherends of different types are bonded together.

[0063] Although not shown in FIG. 5, the training data includes test results of the mechanical properties of the adhesive.

[0064] The inventors investigated the accuracy of estimating adhesive strength using the above-mentioned explanatory variables. As a result, they found that explanatory variables related to the mechanical property values ​​(e.g., mechanical strength) of the adhesive are effective in expressing differences in adhesive strength due to adhesives. However, while methods for estimating adhesive strength that use only the mechanical property values ​​of the adhesive can determine the average strength of adhesive members, it is fundamentally impossible to determine the strength of each individual adhesive member.

[0065] It is known that there is variation in the adhesive strength of individual adhesive materials. Even if the same worker performs the adhesive work according to a predetermined work procedure, the adhesive strength will not be completely equal. For example, when polishing the surface of an adherend, particles scraped from the adherend surface may get between the adherend and the polishing material, and even if the polishing state appears the same to the naked eye, there will be differences on the order of micrometers, such as the surface not being identical.

[0066] As a result of studying adhesive strength, the inventors found that both the surface roughness and the line roughness of the surface can affect the adhesive strength of individual adhesive members, but that the factor that has the greatest influence is the line roughness, not the surface roughness. Among the line roughness factors, they found that factors related to the inclination of the slopes of the surface irregularities have a large influence.

[0067] To investigate why the inclination of the slope has such a large effect on adhesive strength, the inventors considered the inclination of the slope of the surface unevenness of the adherend and the state of contact with the adhesive. Figure 6 shows the state of contact between an adhesive and an adherend with a surface unevenness that has a gentle slope. Adherend 602 has an uneven surface, and the slope of the unevenness is gentle. When applied to the adherend surface, adhesive 601 comes into contact in a liquid state. Because the slope of the unevenness is gentle, the adhesive reaches near the bottom of the valleys, increasing the contact area between the adhesive and the adherend. For this reason, it is thought that adhesive strength increases when the slope of the unevenness is gentle.

[0068] On the other hand, Figure 7 shows the contact state between an adherend and an adhesive when the surface unevenness of the adherend has a steep gradient. It shows the state when adhesive 701 is applied to the surface of adherend 702. Because the surface unevenness of the adherend has a steep gradient, the adhesive has difficulty reaching the bottom of the valleys, resulting in large voids being formed between the adherend and the adhesive, resulting in a smaller adhesive surface area. It is believed that this is how the adhesive strength decreases.

[0069] A representative example of such factors is the root-mean-square slope. If the surface roughness information includes the root-mean-square slope, the adhesive strength can be more appropriately estimated. Factors whose absolute value of correlation with the root-mean-square slope is 0.6 or greater are also effective. An example of such a factor is the ten-point mean roughness. If the surface roughness information includes the ten-point mean roughness, the adhesive strength can be more appropriately estimated.

[0070] The ten-point mean roughness is a quantity that represents the sum of the average of the five largest peaks on a profile curve and the average of the five largest valleys on a reference length. Therefore, if the ten-point mean roughness is large, it means that the valleys are deep when viewed from the highest peaks, which means that it is difficult for adhesive to reach the bottom of the valleys, and there are cases where it is strongly correlated with the root-mean-square slope.

[0071] The inventors' investigations revealed that this effect is particularly pronounced when the adherend is carbon fiber reinforced plastic (CFRP). CFRP is a composite material made by weaving bundles of carbon fibers, forming them into a sheet-like shape, and then impregnating the sheet with resin, thereby achieving a lightweight yet significantly improved resin strength. CFRP is produced by impregnating the carbon fibers with liquid resin, and then solidifying the resin. When the inventors investigated the surface irregularities of CFRP, they found that irregularities formed on the CFRP surface reflect the weaving state of the carbon fibers inside. It is believed that such irregularities affect the adhesive strength of adhesive members.

[0072] In this way, the adhesive strength estimation system 109 according to this embodiment can estimate the strength of an adhesive bond with high accuracy by using information on the adherend and the physical properties of the adhesive in combination.

[0073] The present invention will be described in more detail below with reference to various examples, although the present invention is not limited to the configurations and / or structures described in these examples.

[0074] Example 1 In this example, the adhesive strength between an aluminum plate and a CFRP plate using an epoxy adhesive A is estimated.

[0075] As shown in Figure 8, a single lap joint structure was fabricated by bonding an aluminum alloy plate 801 and a CFRP plate 802 with adhesive 803. The fabrication method conformed to JIS K 6850, and aluminum alloy and CFRP plates measuring 100 mm x 25 mm x 2 mm were bonded together with adhesive. A standard test piece (type: JISK6850AJAA01) was used as the CFRP, and A5083P and ADC-12 were used as the aluminum alloys.

[0076] For bonding, the adherend surfaces were first degreased (wiped with hexane or acetone) to clean them, and then roughened with sandpaper to control the breaking strength. The sandpaper used was #80, #400, #1000, and #2000, and five types of samples were prepared, including an untreated sample. In addition, some samples were UV-treated (for 5 minutes) to improve hydroxyl group bonding on the Al surface.

[0077] The adhesive used was a two-component epoxy adhesive A, which provides sufficient strength. After mixing thoroughly, it was applied to the bonding surfaces and bonded using a Teflon (registered trademark) spacer to ensure an adhesive layer thickness of 0.1 mm. After bonding, it was left to cure at room temperature for 24 hours in a desiccator.

[0078] 74 single lap joint specimens were prepared as specimens to be used in constructing the strength estimation model.

[0079] Before bonding the adherends, the surface roughness was measured. The surface roughness and line roughness were measured using a one-shot 3D shape measuring instrument (Keyence VR3000).

[0080] Seven types of surface roughness indexes were measured for a 12.5 mm x 25 mm adhesive area: root mean square height, maximum peak height, maximum valley depth, arithmetic mean height, maximum height, skewness, and kurtosis. There were seven measurements of the surface roughness of one adherend, and since there were two adherends, there were 7 x 2 = 14 measurements of the surface roughness of one adherend pair.

[0081] The measurement positions for line roughness are shown in Figure 9. The adherend 909 is divided into an adhesive region 907 and a non-adhesive region 908. The line roughness was measured in the adhesive region 907 at a total of three locations on 10 mm long line segments A 901, B 902, and C 903 in the longitudinal direction of the plate, and at a total of three locations on 22 mm long line segments D 904, E 905, and F 906 perpendicular to the longitudinal direction of the plate. This resulted in measurements on a total of six line segments.

[0082] Line segments B902 and E905 were set so that they passed through the center of the adhesive area. The other line segments were set at a sufficient distance from the center line segment, since if they were close to the center line segment, the measurements would likely be nearly identical. However, if they were close to the edge of the adhesive area, they would be affected by deformation caused by cutting during the preparation of the adherend. Therefore, line segments A901, C903, D904, and F906, which do not pass through the center of the adhesive area, were measured at a position 2.25 mm away from the edge of the adhesive area.

[0083] Twelve types of line roughness indices were measured: maximum peak height, maximum valley depth, maximum height, average height, arithmetic mean roughness, maximum cross-sectional height, root mean square height, skewness, kurtosis, average element length, root mean square slope, and ten-point mean roughness. Since there are 12 types of line roughness values ​​for a total of six line segments, there are 6 x 12 = 72 different measurements of the line roughness of one adherend. Since there are two adherends, there are 72 x 2 = 144 different measurements of the line roughness of one pair of adherends.

[0084] To estimate the adhesive strength of adhesively bonded components, a strength estimation model was constructed using 158 explanatory variables related to surface roughness (14 types of surface roughness and 144 types of line roughness). The model was constructed using the Random Forest method, a type of regression technique in machine learning. The Random Forest method constructs a regression model using a large number of decision trees. This method was used in this study because the sandpaper numbers used to roughen the sample surfaces were discrete and many factors were determined by the presence or absence of treatments such as UV treatment. Furthermore, this method has the advantages of being able to model nonlinear behavior and being less prone to overfitting, making it suitable for estimating phenomena where multiple factors are intertwined and complex behavior is expected, as in this example.

[0085] Of the 74 records in the created data, 66 were used as training data and 8 records were used as experimental data, and an estimation model for adhesive strength was then constructed. Figure 10 shows the results of estimation of adhesive strength using the estimation model. The horizontal axis represents the experimental adhesive strength value, and the vertical axis represents the estimated adhesive strength value. If the estimated value matches the experimental value, the plot will lie on the dotted line. In this estimation result, the plot is close to the dashed line, and the eight estimated points generally gave strengths similar to the experimental values. The maximum difference between the experimental value and the theoretical value was less than 2 MPa. The correlation coefficient between the experimental value and the estimated value was R = 0.88, indicating that the estimation was highly accurate.

[0086] Table 1 shows the importance of the variables used in the estimation model for the top 10 variables using the out-of-bag estimation method.

[0087]

[0088] A higher variable importance indicates a greater influence on the objective variable. The variable with the greatest influence on adhesive strength was the root-mean-square slope of line segment B902 on the CFRP surface. Similarly, the variable importance of the root-mean-square slopes for line segments A901, D904, and E905 was also high, so it can be said that the root-mean-square slope determines the majority of adhesive strength. Furthermore, the importance of line segment B902 accounts for more than half of the variables, indicating that line segment B902, which is parallel to the tensile direction and passes through the center of the adhesive region, is the most important.

[0089] As an index of line roughness, not only the root mean square slope but also the average 10-point roughness had importance, although its importance was much lower than that of the root mean square slope.

[0090] Example 2 In this example, the adhesive strength between an aluminum plate and a CFRP plate when various types of adhesives are used is estimated.

[0091] In this example, 12 types of adhesives were selected. All adhesives are two-component, and can be hardened and bonded by mixing the main component with a hardener in a specified ratio. Generally, epoxy adhesives have high tensile shear strength but are brittle, urethane adhesives have low tensile shear strength but good elongation, and acrylic adhesives have properties intermediate between epoxy and urethane adhesives. For this reason, acrylic, epoxy, and urethane adhesives were used without bias.

[0092] The types of adhesives used are shown in Table 2.

[0093]

[0094] Seven types of epoxy adhesives, A to G, were used. Two types of acrylic adhesives, H and I, were used. Three types of urethane adhesives, J, K, and L, were used.

[0095] The mechanical properties of the solidified adhesive obtained for each adhesive A to L were the modulus of elasticity, breaking strength, and breaking elongation. In this way, by including at least one of the modulus of elasticity, breaking strength, and breaking elongation in the mechanical property values ​​of the solidified adhesive, the dependency of adhesive strength on adhesive properties can be reflected in the estimation model by using the property values ​​of many different adhesives that vary over a wide range, enabling more accurate estimation of adhesive strength. Additionally, adhesive viscosity was used to describe the affinity of the adhesive with the adherend surface during application. A total of four parameters were obtained.

[0096] The elastic modulus, breaking strength, and breaking elongation were measured by tensile testing dumbbell test specimens prepared from the adhesive. Dumbbell test specimens were prepared using silicone or polyethylene molds to conform to the JIS K 7139 Type A1 test specimen dimensions. A mixture of the base resin and curing agent was poured into the mold. Five specimens (N=5) were prepared for each adhesive type, and the mechanical properties of each specimen were measured and averaged. Dumbbell tensile tests were performed using a tensile testing machine (Autograph AG-100kNX, Shimadzu Corporation). The tensile speed was 1 mm / min. The breaking strength was measured as the test force at break, and the elongation at break was also measured. The slope of the plot of the stress-strain curve from 0.05% to 0.25% strain was used to calculate the elastic modulus.

[0097] Regarding viscosity, since hardening begins immediately when the base resin and hardener are mixed, experimental measurement of the viscosity of the two-liquid mixture is difficult. Therefore, the viscosity of the mixture was estimated using the Kendahl equation, which shows that the viscosity of the mixture is additivity between the viscosity of the base resin and the viscosity of the hardener.

[0098] Adhesive bonded samples were prepared in the same manner as in Example 1. That is, the single lap joint structure shown in Fig. 8 was prepared. However, instead of using one type of adhesive, adhesive bonded samples were prepared for each of 12 types of adhesive.

[0099] For bonding, the adherend surfaces were first degreased (wiped with hexane or acetone) to clean them, and then roughened with sandpaper to control the breaking strength. The sandpaper used was #2000, and two types of samples were prepared, including an untreated one.

[0100] The surface irregularities were measured before bonding the adherends. Surface roughness and line roughness were measured using a one-shot 3D shape measuring instrument (Keyence VR3000). Seven types of surface roughness indices were measured for a 12.5 mm × 25 mm bonding area: root mean square height, maximum peak height, maximum valley depth, arithmetic mean height, maximum height, skewness, and kurtosis. Linear roughness was measured in the same manner as in Example 1, as shown in Figure 9. Twelve types of line roughness indices were measured: maximum peak height, maximum valley depth, maximum height, average height, arithmetic mean roughness, maximum cross-sectional height, root mean square height, skewness, kurtosis, average element length, root mean square slope, and ten-point average roughness.

[0101] The adhesive was applied to the bonding surface after thoroughly mixing the base agent and curing agent, and a Teflon spacer was used to bond the adhesive so that the adhesive layer was 0.1 mm thick. To examine the effect of adhesive layer thickness on the epoxy adhesives B and C, additional samples were prepared with an adhesive layer thickness of 1.0 mm. After bonding, the samples were allowed to harden at room temperature for 24 hours in a desiccator.

[0102] The number of adhesive bonding samples produced was 73. The breakdown was 11 x 3 x 2 = 66 samples (N = 3 for each of the 11 adhesives excluding epoxy adhesive A, with two types, one with and one without surface polishing treatment), one sample produced for comparison using epoxy adhesive A without roughening treatment (N = 1), and 2 x 3 = 6 samples with increased adhesive layer thickness for epoxy adhesive B and epoxy adhesive C, for a total of 66 + 1 + 6 = 73 samples.

[0103] The adhesive layer thickness and adhesive layer slope were used as explanatory variables. To account for the variability in adhesive layer thickness in single-lap joint structures, the sample thickness was measured with a micrometer at three points on each of the aluminum plate and the CFRP plate, and at six points on the adhesive area. The adhesive layer thickness was calculated by subtracting the average thickness of the aluminum plate and the average thickness of the CFRP plate from the average thickness of the adhesive area.

[0104] The thickness measurement points will be explained using Figure 11. Figure 11(a) shows a side view and Figure 11(b) shows a top view of a sample in which an aluminum plate 1101 and a CFRP plate 1102 are bonded together by an adhesive layer 1103. In Figure 11(b), three measurement points, measurement point A 1107, measurement point B 1108, and measurement point C 1109, are taken on a line perpendicular to the long side of the aluminum plate 1101. The average value of these three points is calculated to obtain the average thickness of the aluminum plate 1101.

[0105] Similarly, the average thickness of the CFRP plate is obtained by measuring the thickness at three measurement points, measurement point J1116, measurement point K1117, and measurement point L1118, and calculating the average value.

[0106] The average thickness of the adhesive region is measured at six measurement points: measurement point D 1110, measurement point E 1111, measurement point F 1112, measurement point G 1113, measurement point H 1114, and measurement point I 1115, and the average value is calculated. The adhesive thickness is obtained by subtracting the average thickness of the aluminum plate 1101 and the average thickness of the CFRP plate 1102 from the average thickness of the adhesive region.

[0107] The explanatory variables used to estimate the adhesive strength of adhesively bonded members were 158 explanatory variables related to surface roughness (14 types of surface roughness and 144 types of line roughness), three types of mechanical properties of the adhesive, adhesive viscosity, and adhesive layer thickness, for a total of 163 types. The number of records used was 147, including the 74 records prepared in Example 1 and the 73 records prepared in this example. A random forest method, a type of regression method in machine learning, was used to build the model. A strength model was constructed and estimated using 146 of the 147 records as training data and the remaining 1 record as experimental data, and this process was repeated for all data to calculate the average estimation accuracy.

[0108] To consider the extent to which the root-mean-square slope RΔq of line segment B902 on the CFRP surface and related information, among the line roughness indices, influences adhesive strength, a data set was created by removing line roughness indices with a high correlation coefficient with RΔq from all data. Specifically, the correlation coefficient between RΔq and each line roughness indices was calculated, and a data set was created by removing indices whose value exceeded φ. This φ is referred to as the usage rate of RΔq-related information. Here, four data sets were created with φ = 1.0, 0.6, 0.4, and 0.3, and are referred to as Data Set 1, Data Set 2, Data Set 3, and Data Set 4, respectively. φ = 1.0 represents a complete data set with nothing removed, while φ = 0.3 represents a data set with the lowest usage rate of RΔq-related information.

[0109] Figure 12 shows the results of constructing an estimation model and estimating adhesive strength using Dataset 4, and evaluating its accuracy. The horizontal axis represents the experimental adhesive strength value, and the vertical axis represents the estimated adhesive strength value. If the adhesive strength could be accurately estimated, the plot of the estimated value would fall on the dashed line. Here, the plot points vary due to error, and the correlation coefficient between the experimental and estimated adhesive strength values ​​was R = 0.66. This shows that the estimation accuracy is low when the usage rate of related information for RΔq is only 0.3.

[0110] An estimation model was constructed and its accuracy was evaluated using Dataset 1. The results of the bond strength estimation are shown in Figure 13. The horizontal axis represents the experimental bond strength value, and the vertical axis represents the estimated bond strength value. In this case, the correlation coefficient between the experimental and estimated bond strength values ​​was R = 0.80. It can be seen that the usage rate of related information for RΔq was 1.0, which improved the estimation accuracy compared to when Dataset 4 was used.

[0111] When adhesive strength estimation was performed using Dataset 2 in the same manner as in the verification of estimation accuracy for Dataset 1 and Dataset 4, the correlation coefficient between the experimental and estimated adhesive strength values ​​was R = 0.79. Similarly, when adhesive strength estimation was performed using Dataset 3, the correlation coefficient between the experimental and estimated adhesive strength values ​​was R = 0.76. The relationship between the usage rate of information related to RΔq and estimation accuracy was plotted as shown in Figure 14. It can be seen that accuracy improves rapidly when the usage rate of information related to RΔq exceeds 0.4.

[0112] From this, it can be said that more accurate estimation is possible if the surface roughness information is configured to include at least one of the following indicators: - Line roughness having a correlation coefficient of 0.4 or more with the root mean square slope; and - Areal roughness having a correlation coefficient of 0.4 or more with the root mean square slope.

[0113] [Example 3] In this example, the adhesive strength between an aluminum plate and a CFRP plate was estimated when various types of adhesives were used, as in Example 2, but the estimation accuracy was improved by dividing the data set. In the adhesive strength estimation method according to this example, the adhesive strength estimation system estimates adhesive strength by using an adhesive strength estimator for one type of adhesive based on surface roughness information of the adherend as part of an estimator for adhesive strength of other types of adhesive.

[0114] In this example, the data set used in Example 2 was utilized. However, in this example, the data set was divided to improve the accuracy of estimating the adhesive strength.

[0115] Below, an example of an improved method for estimating adhesive strength will be described. The overall picture of the method for improving prediction accuracy implemented in this example consists of three parts. The first part is a part for constructing an estimation model for adhesive strength based on adhesive A. The second part is a part for constructing a model for estimating adhesive strength when an arbitrary adhesive is used, using the estimation model for adhesive strength based on adhesive A. The last part is a part for predicting the unknown adhesive strength of an adhesive bond, using the estimation model for adhesive strength based on adhesive A and the model for estimating adhesive strength when an arbitrary adhesive is used.

[0116] The datasets used in the above method will be described below. The dataset used in Example 2 had 147 records, which were separated into 74 records prepared in Example 1 and 73 records prepared in Example 2. The separated datasets are called Dataset 5 and Dataset 6, respectively.

[0117] Two new datasets were created from Dataset 5. The first dataset was created using information on adhesive strength, surface roughness, and line roughness from Dataset 5. This dataset is called Dataset 7. The data format of Dataset 7 was designed as shown in Figure 15. The other dataset was created using information on adhesive strength, adhesive breaking strength, breaking elongation, elastic modulus, adhesive viscosity, and adhesive layer thickness from Dataset 5. This dataset is called Dataset 8. The data format of Dataset 8 was designed as shown in Figure 16. Datasets 9 and 10 were created from Dataset 6 in a similar manner. The data format of Dataset 9 is as shown in Figure 15, and the data format of Dataset 10 was designed as shown in Figure 16.

[0118] The part of constructing an estimation model of adhesive strength based on adhesive A will be described below. The data processing procedure is shown in FIG. 17 . First, a trained model of adhesive strength is constructed using dataset 7 1701. As explained above, dataset 7 consists of surface line roughness measurement values ​​1702 and adhesive strength measurement values ​​1703. Dataset 7 is input to an estimation model construction device 1704. Since all data in dataset 7 is based on adhesive A, the estimation model constructed here is a model that predicts adhesive strength when adhesive A is used for bonding. This is called the A-reference adhesive strength estimation model. Dataset 7 is passed via input interface 1705 to an A-reference adhesive strength estimation model construction unit 1706, which constructs a trained model of adhesive strength. The constructed model is output via output interface 1707 and becomes dataset 7 trained model 1708.

[0119] The following describes the process of constructing a model for estimating the adhesive strength when an arbitrary adhesive is used, using an adhesive strength estimation model based on adhesive A. Here, a prediction model for other types of adhesives is constructed using dataset 9 and dataset 10. Records related to the adhesive to be predicted are extracted from dataset 9 and dataset 10, and are referred to as dataset 9 (experimental data) 1724 and dataset 10 (experimental data) 1731, respectively. Dataset 10 (experimental data) 1731 includes mechanical property measurements 1732 and adhesive strength measurements 1733. The remaining data are referred to as dataset 9 (training data) 1709 and dataset 10 (training data) 1716, respectively, and are used to construct an adhesive strength trained model 1723. Dataset 10 (training data) 1716 includes mechanical property measurements 1717 and adhesive strength measurements 1718.

[0120] Data set 9 (teaching data) 1709 consists of measured values ​​1710 of surface line roughness, and is processed by an A-standard bond strength estimation device 1711. Specifically, data set 9 (teaching data) 1709 is input using an input interface 1712 and sent to an A-standard bond strength estimation unit 1713. The A-standard bond strength estimation unit 1713 is an estimation unit created using a data set 7 trained model 1708. This calculates an A-standard bond strength 1715 and outputs it from an output interface 1714.

[0121] The reference adhesive strength A 1715 is sent to an estimation model construction device 1719 and processed in an adhesive strength estimation model construction unit 1721 via an input interface 1720. Similarly, data set 10 (teacher data) 1716 is also sent to the adhesive strength estimation model construction unit 1721 via the input interface 1720. In this way, an adhesive strength trained model 1723 is constructed and output from an output interface 1722.

[0122] The following describes the process of predicting the unknown adhesive strength of an adhesive bond using an adhesive strength estimation model based on adhesive A and a model for estimating the adhesive strength when an arbitrary adhesive is used. Data set 9 (experimental data) 1724 is input to an A-reference adhesive strength estimation device 1726. Specifically, the data is processed in an A-reference adhesive strength estimation unit 1728 via an input interface 1727. The A-reference adhesive strength estimation unit 1728 is an estimation unit created using the Dataset 7 trained model 1708, and has the same function as the A-reference adhesive strength estimation unit 1713. Here, an A-reference adhesive strength 1730 is calculated and output from an output interface 1729.

[0123] The output A-standard adhesive strength 1730 is sent to an adhesive strength estimation device 1734, and processed in an adhesive strength estimation unit 1736 via an input interface 1735. Similarly, data set 10 (experimental data) 1731 is also sent to the adhesive strength estimation unit 1736 via the input interface 1735. The adhesive strength estimation unit 1736 is an estimation unit created using the adhesive strength trained model 1723. In this way, an adhesive strength 1738 is calculated and output from an output interface 1737.

[0124] The predicted accuracy of the adhesive strength of the adhesive bond thus obtained was R = 0.82. This value exceeds the predicted accuracy R = 0.80 when φ = 1.0 in Example 2, and therefore the effectiveness of the prediction method in this example is confirmed.

[0125] REFERENCE SIGNS LIST 101 Teacher data 102 Experimental data 103 Input interface 104 Memory unit 105 Estimation model construction unit 106 Adhesive strength estimation unit 107 Output interface 108 Adhesive strength of adhesive bond 109 Adhesive force estimation system 401 Teacher data 402 Measured mechanical property data of solidified adhesive 403 Measured surface line roughness data of adherend 404 Measured adhesive strength data of adhesive bond 405 Estimation model construction device 406 Input interface 407 Estimation model construction unit 408 Output interface 409 Trained model 410 Experimental data 411 Mechanical property data 412 Measured surface line roughness data of adherend 417 Adhesive strength 601 Adhesive 602 Adherend 701 Adhesive 702 Adherend 801 Aluminum alloy plate 802 CFRP plate 803 Adhesive 901 Line segment A 902 Line segment B 903 Line segment C 904 Line segment D 905 Line segment E 906 Line segment F 907 Adhesion area 908 Non-adhesion area 909 Adherend 1101 Aluminum plate 1102 CFRP plate 1103 Adhesion layer 1107 Measurement point A 1108 Measurement point B 1109 Measurement point C 1110 Measurement point D 1111 Measurement point E 1112 Measurement point F 1113 Measurement point G 1114 Measurement point H 1115 Measurement point I 1116 Measurement point J 1117 Measurement point K 1118 Measurement point L 1701 Data set 7 1702 Surface line roughness measurement value 1703 Adhesion strength measurement value 1704 Estimation model construction device 1705 Input interface 1706 A-standard adhesive strength estimation model construction unit 1707 Output interface 1708 Data set 7 trained model 1709 Data set 9 (teaching data) 1710 Measured surface line roughness value 1711 A-standard adhesive strength estimation device 1712 Input interface 1713 A-standard adhesive strength estimation unit 1714 Output interface 1715 A-standard adhesive strength 1716 Data set 10 (teaching data) 1717 Measured mechanical property value 1718 Measured adhesive strength value 1719 Estimation model construction device 1720 Input interface 1721 Adhesive strength estimation model construction unit 1722 Output interface 1723 Adhesive strength trained model 1724 Data set 9 (experimental data)1725 Measured surface line roughness 1726 A-standard bond strength estimator 1727 Input interface 1728 A-standard bond strength estimator 1729 Output interface 1730 A-standard bond strength 1731 Data set 10 (experimental data) 1732 Measured mechanical property value 1733 Measured bond strength 1734 Bond strength estimator 1735 Input interface 1736 Bond strength estimator 1737 Output interface 1738 Bond strength

Claims

1. An adhesive strength estimation system comprising: an input interface that accepts input of mechanical property values ​​of a solidified adhesive and information on the surface roughness of an adherend; a memory unit that stores a trained model trained using training data that takes as input a combination of test results of the mechanical property values ​​of a test piece of the solidified adhesive and information on the surface roughness of the adherend, and outputs adhesive strength; and an adhesive strength estimation unit that uses the trained model to estimate adhesive strength based on the input to the input interface.

2. The adhesive strength estimation system according to claim 1, wherein the surface roughness information includes a root mean square slope.

3. An adhesive strength estimation system according to claim 1, wherein the surface roughness information includes ten-point average roughness.

4. An adhesive strength estimation system as claimed in claim 1, wherein the surface roughness information includes at least one of the following indicators: a line roughness having a correlation coefficient of 0.4 or more with the root mean square slope; and a surface roughness having a correlation coefficient of 0.4 or more with the root mean square slope.

5. An adhesive strength estimation system according to claim 1, wherein the mechanical property value of the solidified adhesive includes at least one of elastic modulus, breaking strength, and breaking elongation.

6. An adhesive strength estimation method, in which the adhesive strength estimation system according to claim 1 estimates adhesive strength.

7. An adhesive strength estimation method as claimed in claim 6, wherein the adhesive strength estimation system estimates adhesive strength by using an adhesive strength estimator for one type of adhesive based on surface roughness information of the adherend as part of an estimator for the adhesive strength of another type of adhesive.

8. A computer that stores data relating to a combination of test results of mechanical properties of test pieces of solidified adhesive, information on the surface roughness of the adherend, and adhesive strength.

Images