Method and apparatus for estimating fracture life

Abstract

To provide an estimation method and device of a rupture life which can obtain estimation with excellent accuracy by performing estimation of an S-N curve (relation of stress amplitude-rupture life) by machine learning.SOLUTION: An estimation device of a rupture life comprises: a function unit (310) which reads designation of a type of fatigue limit estimation and designation of mechanical characteristics used in a decision tree of machine learning; a function unit (320) which reads fatigue data of the designated steel type from a fatigue data sheet; and a machine learning arithmetic unit (330) which uses the designated mechanical characteristics to perform machine learning with machine-learned teacher to the read fatigue data of the steel type. The machine learning arithmetic unit that has performed the machine learning with teacher uses the designated mechanical characteristics to output an arithmetic result of fatigue limit estimation according to the designated type of fatigue limit estimation.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a method and apparatus for estimating the fracture life of structures and structural materials, and in particular to a method and apparatus for estimating the fracture life of various structural materials using machine learning methods. 7 This invention relates to a method and device for estimating fracture life using measurement data of cycle fatigue limit. [Background technology]

[0002] Steel materials are the main components of machines and structures, and fatigue properties are one of the important issues in terms of maintenance and management, and dealing with damage and defects. 7 The fatigue limits of various cycles have been compiled as NIMS Fatigue Data Sheets (FDS). From these FDS, it is empirically known that there is a correlation between the fatigue limit and other mechanical properties (e.g., Vickers hardness, tensile strength, etc.) (Fig. 11). In addition to the fatigue limit, we are also attempting to estimate the fracture life (SN curve) by normalizing the stress amplitude. Patent Document 1 discloses a method for estimating the fatigue properties of steel materials. Patent Document 2 discloses a method for analyzing rubber materials to improve the wear resistance of tires. Patent Document 3 discloses a highly accurate classification and analysis method for measurement data using machine learning software.

[0003] Figure 12 shows the index property proposed by the applicant for the fatigue characteristics of materials. For simplicity, it will be referred to as the index property below. An index property is defined as a value that "approximately reveals the fatigue characteristics of a material" by referring to the index property of fatigue, just as tensile strength is used as an index to evaluate the strength of a material. In Figure 12, fatigue is first divided into high-cycle fatigue and low-cycle fatigue according to the life range. High-cycle fatigue life characteristics are generally expressed by the relationship curve σ between stress amplitude and life. a In this case, the index to be referred to is the strength characteristic, and the static index is the tensile strength σ B , and the dynamic index is the repeated yield stress σ yc The reason for this will be explained later.

[0004] On the other hand, the low cycle fatigue life characteristics are shown by the strain-life relationship curve ε a Therefore, the index to be referred to is the deformation characteristics. In this case, the fracture ductility ε f The dynamic index is the exponent n' of the cyclic stress-strain curve (Reference 1). Tensile strength σ B and fatigue limit σ w It is empirically known that there is a good correlation between the yield stress σ y (or 0.2% yield strength σ 0.2 ) and σ w The correlation between has also been investigated, but σ B -σ w This is not as strong as the relationship between σ y However, the repeated yield stress σ yc and σ w A linear relationship is established between the cyclic yield stress σ yc This is thought to be because the internal structure has reached a steady state after repeated plastic straining. B , and the dynamic index is the repeated yield stress σ yc It is considered appropriate to adopt the following.

[0005] Since fatigue occurs due to repeated plastic strain, it is thought that a dynamic index should essentially be adopted. yc There are barriers to the adoption of σ. yc Measurement of σ requires strain control testing using the companion specimen method or the incremental step method, and there has been a problem in that the number of measurements is not always sufficient. As shown in Figure 13, the tensile strength σ B and cyclic yield stress σ yc Since the two index characteristics are proportional to each other, it seems that it is acceptable to use the static index in practice. Bσ normalized by a / σ B When the fatigue properties of the material were evaluated using -Nf, the results shown in Figure 14 were obtained. B σ normalized by a / σ B The overall -Nf band was wide, which meant that the estimation was never very accurate. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2017-187408 [Patent Document 2] Patent Publication No. 2021-071803 [Patent Document 3] International Publication No. 2018-207524 [Non-patent literature]

[0007] [Non-Patent Document 1] S. Matsuoka, N,Nagashima and S. Nishijima, “Index property for the fatigue of engineering alloys”, NIMS Materials Strength Data Sheet Technical Document, No.17(1997) Summary of the Invention [Problem to be solved by the invention]

[0008] The present invention solves the problems of the conventional technology described above, and aims to provide a method and device for estimating fracture life that can obtain accurate estimations by estimating the SN curve (relationship between stress amplitude and fracture life) using machine learning. [Means for solving the problem]

[0009] The inventors of the present invention came up with the idea that the relationship between stress amplitude and rupture life can be estimated accurately by analyzing experimental data available from the NIMS fatigue data sheet using a machine learning method. 7 The accuracy of estimating the fatigue limit and the estimated value of the SN curve are 6 This is a verification of the probability that the difference between the estimated fracture life of less than 100 times and the actual measured value falls within the acceptable range.

[0010] [1] As shown in FIG. 3, the fracture life estimation device of the present invention includes a functional unit (310) that reads the specification of the type of fatigue limit estimation and the specification of the mechanical properties used in the machine learning decision tree; a functional unit (320) for reading fatigue data of a specified steel type from a fatigue data sheet; a machine learning calculation unit (330) that performs supervised machine learning on fatigue data of the loaded steel type using the specified mechanical properties, The machine learning calculation unit that performed the supervised machine learning uses the specified mechanical characteristics to output the calculation result of the fatigue limit estimation according to the specified type of fatigue limit estimation. [2] In the fracture life estimation device [1] of the present invention, it is preferable that the machine learning calculation unit (330) has a pre-processing unit (325) that performs machine learning on the fatigue data of the loaded steel type using the specified mechanical properties, and performs the supervised machine learning. [3] In the device for estimating fracture life of the present invention [2], the model used for the machine learning preferably includes a model using a random forest method, deep learning, a neural network, or LASSO regression.

[0011] [4] In the fracture life estimation device [1] to [3] of the present invention, preferably, the type of fatigue limit estimation is 10 7 Estimation of fatigue limit, 10 6 This may include estimation of the fatigue limit, estimation of the SN curve, and estimation of the low cycle fatigue life. [5] In the fracture life estimation device [1] to [4] of the present invention, preferably, the mechanical properties used in the machine learning decision tree include Vickers hardness, tensile strength, fracture elongation, fracture reduction, stress ratio, or fatigue test method. [6] In the device for estimating fracture life of the present invention [5], preferably, the fatigue test method includes rotating bending test data, axial load test data, and torsion test data. [7] In the fracture life estimation device [1] to [6] of the present invention, the steel types preferably include carbon steel, nickel chromium molybdenum steel, manganese steel, and stainless steel. [8] In the fracture life estimation device [7] of the present invention, the steel types preferably include, for example, S25C, S35C, and S55C for carbon steel materials, SNCM439 for nickel-chromium-molybdenum steel materials, SMn438 and SMn443 for manganese steel materials, and SUS403 and SUS304 for stainless steel materials.

[0012] [9] In the fracture life estimation device [1] to [8] of the present invention, it is preferable to have a machine learning evaluation unit (335) that evaluates the accuracy of the fatigue limit estimation of the machine learning calculation unit (330) that performed supervised machine learning using the calculation result of the fatigue limit estimation according to the type of fatigue limit estimation used for the calibration object.

[10] The fracture life estimation device [9] of the present invention preferably has an assist function (340) that recommends, based on the evaluation results of the machine learning evaluation unit (335), the types of mechanical properties to be used in the machine learning decision tree and the types of fatigue data of the steel type to be read in the fatigue data sheet, so as to improve the accuracy of the fatigue limit estimation by the machine learning calculation unit (330).

[0013]

[11] The method for estimating fracture life of the present invention includes, for example, as shown in FIG. 4, a step (S400) of specifying the type of fatigue limit estimation and the mechanical properties to be used in the machine learning decision tree; A step (S405) of reading fatigue data of the specified steel type from a fatigue data sheet; and a step (S415) of outputting the calculation result of fatigue limit estimation according to the specified type of fatigue limit estimation using the specified mechanical properties to a machine learning calculation unit (330) that has performed supervised machine learning on the fatigue data of the loaded steel type using the specified mechanical properties.

[12] In the fracture life estimation method

[11] of the present invention, it is preferable to have a step (S410) of performing supervised machine learning by performing machine learning on the fatigue data of the loaded steel type using specified mechanical properties.

[13] In the fracture life estimation method

[12] of the present invention, preferably, the model used for the machine learning includes a model using a random forest method, deep learning, a neural network, or LASSO regression.

[0014]

[14] In the fracture life estimation method

[11] to

[13] of the present invention, preferably, the type of fatigue limit estimation is 10 7 Estimation of fatigue limit, 10 6 This may include estimation of the fatigue limit, estimation of the SN curve, and estimation of the low cycle fatigue life.

[15] In the fracture life estimation methods

[11] to

[14] of the present invention, preferably, the mechanical properties used in the machine learning decision tree include Vickers hardness, tensile strength, fracture elongation, fracture reduction, stress ratio, or fatigue test method.

[16] In the fracture life estimation method

[15] of the present invention, preferably, the fatigue testing method includes rotating bending test data, axial load test data, and torsion test data.

[17] In the fracture life estimation method

[11] to

[16] of the present invention, the steel types preferably include carbon steel, nickel-chromium-molybdenum steel, manganese steel, and stainless steel.

[18] In the method

[17] for estimating fracture life of the present invention, the steel type preferably includes, for example, S25C, S35C, and S55C for carbon steel materials, SNCM439 for nickel-chromium-molybdenum steel materials, SMn438 and SMn443 for manganese steel materials, and SUS403 and SUS304 for stainless steel materials.

[0015]

[19] In the fracture life estimation method

[11] to

[18] of the present invention, it is preferable to include a step (S425) of evaluating the accuracy of the fatigue limit estimation of the machine learning calculation unit (330) that performed supervised machine learning using the calculation result of the fatigue limit estimation according to the type of fatigue limit estimation used for the calibration object.

[20] The fracture life estimation method

[19] of the present invention preferably includes a step (S430) of recommending, based on the evaluation results of the machine learning evaluation step (S425), the types of mechanical properties to be used in the machine learning decision tree and the types of fatigue data of the steel type to be read in the fatigue data sheet, so as to improve the accuracy of the fatigue limit estimation by the machine learning calculation unit (330).

[21] The fracture life estimation method

[20] of the present invention preferably includes a step (S435) of modifying the type of fatigue limit estimation specified for use in S400 of the fracture life estimation method

[11] and the mechanical properties used in the machine learning decision tree, taking into account the type of mechanical properties used in the recommended machine learning decision tree and the steel type read in the fatigue data sheet. [Effects of the Invention]

[0016] According to the method and apparatus for estimating fracture life of the present invention, data on various steel types included in fatigue data sheets are used, and a random forest method is used to estimate fracture life. 7 fatigue limit and 10 6 We attempted to estimate the fracture life of less than 1000 times and examined the possibility of estimating the SN curve, and obtained the following results. (1) Machine learning regression models that can associate multiple learning elements are excellent at estimating fatigue limits. (2) Machine learning estimation of the SN curve can be performed with high accuracy by separately estimating the fracture life and fatigue limit. [Brief explanation of the drawings]

[0017] [Figure 1] 1 is a block diagram illustrating an example of a schematic configuration of a fatigue limit estimation system according to an embodiment of the present disclosure. [Figure 2] FIG. 2 is a block diagram showing an exemplary computing device 200 in the case where a fatigue life estimation calculation processing unit of the device shown in FIG. 1 is configured using a computer. [Figure 3] FIG. 3 is a functional block diagram of software for a computer having the functional blocks shown in FIG. 2. [Figure 4] FIG. 4 is an explanatory diagram of a fatigue limit estimation algorithm for the device according to the software functional block diagram shown in FIG. 3. [Figure 5A] This figure shows the relationship between the fatigue limit predicted by AI using 107 consecutive data sets from rotating bending fatigue tests and torsional fatigue tests for S25C and S55C, and the fatigue limit determined experimentally. It also shows predictions made by a decision tree model for HV only. [Figure 5B] This figure shows the relationship between the fatigue limit predicted by AI using 107 consecutive data sets of rotating bending fatigue tests and torsional fatigue tests for S25C and S55C, and the fatigue limit determined experimentally. It also shows predictions made using a decision tree model for HV and test method. [Figure 6A] This is a prediction diagram showing the relationship between the fatigue limit predicted by AI using 107 consecutive data sets (306 tests in total) of axial load tests (R=0, -1) in rotating bending fatigue tests and torsional fatigue tests for S25C and S55C, and the fatigue limit determined experimentally, based on tensile strength and test method. [Figure 6B] This graph shows the relationship between the fatigue limit predicted by AI using 107 consecutive data sets (306 tests in total) of axial load tests (R=0, -1) in rotating bending fatigue tests and torsional fatigue tests for S25C and S55C, and the fatigue limit determined experimentally, based on tensile strength, test method, and stress ratio. [Figure 7] This is a graph showing the predicted fatigue limit for 107 cycles using the data for the steel types included in the fatigue data sheet. [Figure 8] This is the result of predicting the breakdown life using data from S25C and S55C (only data for breakdown life of 106 cycles or less was used). [Figure 9] This is the result of predicting fracture life using data on steel types included in fatigue data sheets (only data on fracture life of 106 cycles or less was used). [Figure 10] This is a prediction of the SN curve of fracture life using data on the steel type included in the fatigue data sheet (data on fracture life of 5 x 106 times or less, fatigue limit only considers hardness). [Figure 11A] 1 is a diagram showing the relationship between mechanical properties and fatigue limits, showing Vickers hardness. [Figure 11B] This is a diagram showing the relationship between mechanical properties and fatigue limit, and shows tensile strength. [Figure 12] FIG. 1 is a diagram showing an index property proposed by the present applicant that serves as an index for the fatigue properties of a material. [Figure 13] FIG. 1 is a diagram showing the relationship between two index characteristics, tensile strength σB and cyclic yield stress σyc. [Figure 14] This is a graph showing the SN curve normalized by tensile strength. DETAILED DESCRIPTION OF THE INVENTION

[0018] The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an example of a schematic configuration of a fatigue limit estimation system according to an embodiment of the present disclosure.

[0019] The fatigue limit estimation system 10 includes a fatigue testing machine 20 and a fatigue limit estimation device 30. The fatigue testing machine 20 includes a testing machine main body 22 and a controller. The fatigue limit estimation device 30 includes an estimation calculation processing unit 32 and an extensometer 34. The controller of the fatigue testing machine 20 and the extensometer 34 are connected to the estimation calculation processing unit 32 of the fatigue limit estimation device 30.

[0020] The fatigue testing machine 20 is fitted with a test piece SP whose fatigue limit is to be tested, and a predetermined load is repeatedly applied to perform the fatigue test. The fatigue limit estimation device 30 uses an extensometer 34 to measure the elongation of the test piece SP to which the load is applied by the fatigue testing machine 20, and estimates the fatigue limit.

[0021] Next, an example of the hardware configuration of the estimation calculation processing unit 32 that constitutes the fatigue limit estimation device 30 will be described. Figure 2 is a block diagram showing an exemplary computing device 200 in which the fatigue life estimation processing unit of the device shown in Figure 1 is configured using a computer. The fatigue life estimation processing unit 32 of Figure 1 can be implemented using all or part of the computing device 200. In a very basic configuration 901, computing device 200 typically includes one or more processors 210 and a system memory 220. A memory bus 930 may be used for communication between the processors 210 and the system memory 220.

[0022] Depending on the desired configuration, processor 210 may be of any type, including but not limited to a microprocessor (μP), a microcontroller (μC), a digital signal processor (DSP), or a combination thereof. Processor 210 may include another level of caching, such as a level 1 cache 211 and a level 2 cache 212, a processor core 213, and registers 214. An exemplary processor core 213 may include an arithmetic logic unit (ALU), a floating point unit (FPU), a digital signal processing core (DSP core), or any combination thereof. An exemplary memory controller 215 may also be used with processor 210, or in some implementations, memory controller 215 may be an internal part of processor 210.

[0023] Depending on the desired configuration, the system memory 220 can be of any type, including but not limited to, volatile memory (such as RAM), non-volatile memory (such as ROM, flash memory, etc.), or any combination thereof. The system memory 220 can include an operating system 221, one or more applications 222, and a decision tree model machine learning unit 232. The applications 222 can be configured to run on the system. 7 Estimation of fatigue limit 223, 10 6 The calculation may include a cycle fatigue limit estimation section 224, an SN curve estimation section 225, and a low cycle fatigue life estimation section 226.

[0024] The decision tree model machine learning unit 232 may include, as a decision tree model used in a random forest method, which is a type of machine learning, Vickers hardness 233, tensile strength 234, breaking elongation 235, breaking reduction 236, stress ratio 237, and fatigue test method 238. The fatigue test method 238 includes, for example, rotating bending test data, axial load test data, and torsion test data.

[0025] Computing device 200 may have additional features or functionality and additional interfaces to facilitate communication between basic configuration 201 and any necessary devices and interfaces. For example, bus / interface control 240 may be used to facilitate communication between basic configuration 201 and one or more data storage devices 250 via storage interface bus 241. Data storage device 250 may be removable storage device 251, non-removable storage device 252, or a combination thereof. Examples of removable and non-removable storage devices include magnetic disk drives such as floppy disk drives and hard disk drives (HODs), optical disk drives such as compact disk (CD) drives or digital versatile disk (DVD) drives, solid-state drives (SSDs), and tape drives. Exemplary computer storage media may include volatile and non-volatile, removable and fixed media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules, or other data.

[0026] System memory 220, removable storage 251, and non-removable storage 252 are all examples of computer storage media, including, but not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVDs) or other optical storage devices, magnetic cassettes, magnetic tape, magnetic disk storage devices or other magnetic storage devices. Any such computer storage media that can be used to store desired information and that can be accessed by computing device 200 can be part of device 900.

[0027] The computing device 200 may also include an interface bus 242 to facilitate communication from various interface devices (e.g., output interfaces, peripheral interfaces, and communication interfaces) to the basic configuration 201 via the bus / interface control unit 240. In the output device 260, the image processing unit 261 and the audio processing unit 262 may be configured to communicate with various external devices, such as a display device 291 or speakers, via one or more AV ports 263.

[0028] The exemplary peripheral interface 270 includes a serial interface control unit 271 or a parallel interface control unit 272 that can be configured to communicate with external devices such as input devices (e.g., keyboard, mouse, pen, voice input device, touch input device, etc.) The peripheral interface 270 can be configured to communicate with a fatigue testing machine 292 or an external database device storing fatigue data sheets 293 via an I / O port 273. Exemplary communications device 280 includes a network controller 281, which may be configured to facilitate communications with one or more other computing devices 290 over a network communications link via one or more communications ports 282. The fatigue data sheets 293 include a data sheet 294 for carbon steel materials, a data sheet 295 for nickel-chromium-molybdenum steel materials, a data sheet 296 for manganese steel materials, and a data sheet 297 for stainless steel materials. For carbon steel materials, there are fatigue data sheets for various steel types such as S25C, S35C, and S55C, which are carbon steel materials for machine structures specified in JIS G 4051, and these are available as FDS-No. 1 to 16 and the like provided by the present applicant.

[0029] A network communication link may be an example of a communication medium. Communication media may typically be embodied by computer-readable instructions, data structures, program modules, or other data in a modulated data signal, such as a carrier wave or other transport mechanism, and may include any information delivery media. A "modulated data signal" may be a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media may include wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, radio frequency (RF), microwave, infrared (IR) and other wireless media. As used herein, the term computer-readable media may include both storage media and communication media.

[0030] Computing device 200 may be implemented as part of a small form factor portable (or mobile) electronic device such as a mobile telephone, a personal data assistant (PDA), a personal media player device, a wireless web watch device, a personal computer, a headset device including any of the above functionality, a special purpose device, or a hybrid device. Computing device 200 may also be implemented as a personal computer, including both laptop and non-laptop computer configurations.

[0031] Figure 3 is a software functional block diagram for a computer having the functional blocks shown in Figure 2, illustrating an exemplary computer program product 300. The program bearing medium 302, which may be implemented as a computer readable medium 306, a recordable medium 308, a communication medium 309, or a combination thereof, has program instruction storage 304 that may be configured to perform all or a portion of the processing of a processing unit.

[0032] The program instructions stored in the program instruction storage unit 304 include, for example, a function (310) for reading the specification of the type of fatigue limit estimation and the specification of the mechanical properties to be used in the machine learning decision tree, a function (320) for reading fatigue data of the specified steel type from a fatigue data sheet, and a preprocessing unit (325) for performing supervised machine learning by performing machine learning using the random forest method on the fatigue data of the loaded steel type using the specified mechanical properties.Furthermore, the program includes a machine learning calculation unit (330) for outputting the calculation result of the fatigue limit estimation according to the specified type of fatigue limit estimation using the specified mechanical properties to the machine learning calculation unit (330) that performed the supervised machine learning. The random forest method in machine learning is preferably a model using the well-known random forest method, which uses multiple decision trees for "classification" or "regression." The model used in machine learning is not limited to a model using the random forest method, but also includes, for example, a model using a neural network, such as the well-known deep learning, and a model using LASSO regression. The regression equation for fatigue limit estimation is not limited to a linear function, but can also be a polynomial, kriging, or a nonlinear function using RBF (Radial Base Function).

[0033] The program instructions stored in the program instruction storage unit 304 also include a machine learning evaluation unit (335) that evaluates the accuracy of the fatigue limit estimation by the machine learning calculation unit (330) that performed supervised machine learning, using the calculation results of the fatigue limit estimation according to the type of fatigue limit estimation used for the calibration target. Furthermore, the program instructions include an assist function (340) that recommends, based on the evaluation results of the machine learning evaluation unit (335), the types of mechanical properties to be used in the machine learning decision tree and the types of fatigue data for the steel type to be read in the fatigue data sheet, so as to improve the accuracy of the fatigue limit estimation by the machine learning calculation unit (330).

[0034] FIG. 4 is an explanatory diagram of the fatigue limit estimation algorithm of the device based on the software functional block diagram shown in FIG. 3, where (A) is the basic algorithm and (B) is an additional algorithm. First, in the basic algorithm, the type of fatigue limit estimation and the mechanical properties to be used in the machine learning decision tree are specified (S400). 7 Estimation of fatigue limit, 10 6 These include estimating the cycle fatigue limit, SN curve, and low-cycle fatigue life. Mechanical properties used in machine learning decision trees include Vickers hardness, tensile strength, elongation at break, reduction of area at break, stress ratio, and fatigue test method.

[0035] Next, the fatigue data for the specified steel type is read from the fatigue data sheet (S405). There are data sheets for carbon steel, nickel-chromium-molybdenum steel, manganese steel, and stainless steel, and the subdivided steel types for each steel type include, for example, S25C, S35C, and S55C for carbon steel, SNCM439 for nickel-chromium-molybdenum steel, SMn438 and SMn443 for manganese steel, and SUS403 and SUS304 for stainless steel. Using the specified mechanical properties, machine learning is performed using the random forest method on the fatigue data of the loaded steel type, and supervised machine learning is performed (S410). Note that if supervised machine learning has already been performed, this step may be omitted. Next, the specified mechanical characteristics are used to output the calculation results of fatigue limit estimation according to the specified type of fatigue limit estimation to the machine learning calculation unit (330) that performed supervised machine learning (S415).

[0036] The additional algorithm added to the basic algorithm includes the following steps: Using the calculation results of the fatigue limit estimation according to the type of fatigue limit estimation used for the calibration target, the accuracy of the fatigue limit estimation by the machine learning calculation unit (330) that performed the supervised machine learning is evaluated (S425). In addition, based on the evaluation results of the machine learning evaluation unit (335), the type of mechanical properties to be used in the machine learning decision tree and the type of fatigue data of the steel type to be read in the fatigue data sheet are recommended (S430) so as to improve the accuracy of the fatigue limit estimation by the machine learning calculation unit (330). Considering the type of mechanical properties used in the recommended machine learning decision tree and the steel type loaded in the fatigue data sheet, the type of fatigue limit estimation specified for use in S400 and the mechanical properties used in the machine learning decision tree are modified (S435).

[0037] Next, we will explain the supervised machine learning used in the fatigue limit estimation algorithm mentioned above. The random forest method is a machine learning algorithm, an ensemble learning algorithm that improves generalization ability by integrating weak learners from multiple decision tree models, and is primarily used for classification (discrimination) and regression (estimation). What is important here is that, for the target data group, (i) a larger amount of accurate data can be sampled, and (ii) a decision tree model is created for each learning element. Previous regression models used least squares approximation to find the correlation between two target data groups, but machine learning can create regression models that link decision tree models of multiple learning elements, which is expected to lead to even more accurate estimations.

[0038] 10 7 The data set used to estimate the cycle fatigue limit was based on experimental data from rotating bending fatigue tests on S25C (FDS-No. 1) and S55C (FDS-No. 4), carbon steels for machine structures specified in JIS G 4051. The effect of each learning element was investigated using the random forest method. Next, fatigue limit data from torsional fatigue tests was added to the data set to investigate the influence of different fatigue test methods. Furthermore, fatigue test data for stress ratios R = -1 and R = 0 were added to investigate the influence of stress ratio. Based on the results of the studies conducted so far, we investigated the accuracy of estimation when using fatigue data for various steel materials, including carbon steel for mechanical structures S35C (FDS-No. 2), nickel-chromium-molybdenum steel SNCM439 (FDS-No. 25), manganese steel SMn438 (FDS-No. 16), manganese steel SMn443 (FDS-No. 17), martensitic stainless steel SUS403 (FDS-No. 30), which is used for stainless steel pipes for mechanical structures specified in JIS G 3446, and austenitic stainless steel SUS304 (FDS-No. 33), in addition to S25C and S55C.

[0039] Next, 10 types of steel 6 Finally, we narrowed it down to S45C (FDS-No.3) tempered at 550°C, Heat A, and performed a 10-cycle test for each stress amplitude. 6 We estimated the fracture life of less than 1000 times and attempted to estimate the SN curve. Until now, the estimation of fatigue limit has had a good correlation with tensile strength and hardness, so "fracture elongation" and "fracture reduction area" have not received much attention. However, since fracture ductility is an indicator of low-cycle fatigue in the finite life range of the SN curve, particularly in the short-life low-cycle region, a decision tree model linking tensile strength, hardness, fracture elongation, and fracture reduction area was adopted.

[0040] For the machine learning, a commercially available personal computer was used as the computing device 200, and a system having software functional blocks including the fatigue limit estimation algorithm shown in Figures 2 to 4 was constructed using Python (registered trademark) 3.6.1 and the external library Anaconda (registered trademark).

[0041] The data used were the fatigue test results listed on the fatigue data sheet. To ensure fairness in the analysis, 80% was used as training data and 20% as test data, which were randomly selected each time. Therefore, it is impossible to determine which data corresponds to the test data. In addition, as one way of evaluating the analysis results, the mean absolute percentage error (MAPE) was calculated from the analysis results using the test data.

number

[0042] The root mean square error (RMSE), mean squared error (RSE), and coefficient of determination R are used as indicators to evaluate the fit of a model obtained through regression analysis. 2 However, when calculated using the error functions RMSE and MSE, the positive and negative data are summed, resulting in an average error that cancels out. On the other hand, because MAPE is an absolute value, it can localize discrepancies in the predicted data. Problems with MAPE include cases where the actual measured value is 0 or the predicted value is too small. However, the actual measured and predicted values ​​targeted in this invention are fatigue limits, and these cases do not apply. Furthermore, while it is conceivable that biased conclusions could result if cross-validation and grid search are not performed, we believe that visually inspecting the relationship between experimental and predicted values, such as those shown in Figures 5A and 5B, can serve as a substitute for cross-validation and grid search. For these reasons, we believe it is appropriate to use MAPE rather than RMSE or MSE as the error function in this invention.

[0043] [Fatigue limit estimation using machine learning] Using the rotating bending fatigue test data (218 in total) of S25C and S55C, four decision tree models were created for the learning elements: Vickers hardness, tensile strength, breaking elongation, and breaking reduction. 7 The fatigue limit was estimated as 10 times the experimental value. 7The correlation between the fatigue limit and the number of rotations was determined. Table 1 shows the results of fatigue limit analysis using machine learning for rotating bending tests on S25C and S55C. The analysis results shown in Table 1 show that the MAPE for Vickers hardness and tensile strength was 2% or less, demonstrating high estimation accuracy. These results confirm the good correlation between hardness and tensile strength and fatigue limit in Data Sheet No. 5 (Figures 11A and 11B) using machine learning, but the estimation accuracy has improved dramatically. [Table 1]

[0044] Impact of test method [Table 2] Torsion test data was added to the rotating bending test data for S25C and S55C conducted in the previous section (279 data in total). Test method items were added as learning elements. The analysis results are shown in Table 2 and Figures 5A and 5B. In the following Figures 5A, 5B, 6A, 6B, 7, 8, and 9, training data is indicated by "+" and test data by "▲". The fatigue limit estimated using only Vickers hardness, as shown in Figure 5A, had a MAPE of approximately 12%. On the other hand, the MAPE of the fatigue limit estimated using a regression model linking Vickers hardness and a decision tree model for the testing method, as shown in Figure 5B, was 2.23%, a dramatic improvement in estimation accuracy. This result demonstrates that a machine learning regression model that can associate multiple learning elements is effective for estimating the fatigue limit. Note that axial load test data may be used instead of the rotating bending test data or torsion test data described above.

[0045] Effect of stress ratio A decision tree model was added (306 in total) to the rotary bending and torsion test results for S25C and S55C conducted in the previous section, using test data from axial load tests (R = 0 and -1) as the stress ratio. The analysis results are shown in Table 2 and Figures 6A and 6B. The MAPE for the fatigue limit estimated using only the tensile strength and test method shown in Figure 6A was 3.02%. The MAPE for the fatigue limit estimated using a regression model with three learning elements, linking the tensile strength, test method, and the decision tree model for the stress ratio shown in Figure 6B, was 2.35%, an improvement in estimation accuracy.

[0046] Impact of various steel material data In addition to the fatigue test results for S25C and S55C, the fatigue data sheets also include fatigue limit data (892 data in total) from rotating bending fatigue tests on S35C, SNCM439, SMn438, SMn443, SUS403, and SUS304. The analysis results are shown in Table 2 and Fig. 7. The MAPE of the fatigue limit estimated from a regression model linking a decision tree model for hardness and test method was 2.94%, demonstrating high estimation accuracy.

[0047] 10 by machine learning 6 Estimation of fracture life of less than 100 times S25C and S55C 10 6 The decision tree model for Vickers hardness, tensile strength, reduction of area at break, and elongation at break was created by limiting the data to fracture data of 515 or less. The learning elements of all the decision trees were linked to each other. 6 The analysis results are shown in Table 3 and Fig. 8. [Table 3]

[0048] Estimation using a regression model linking all decision tree models for Vickers hardness, tensile strength, fracture elongation, and fracture area showed a high estimation accuracy of 92.0% for training data, but decreased to 65.8% for randomly sampled test data, with a MAPE of 38.7%. This is thought to be because the training data distinguishes between fracture data for S25C and S55C, resulting in fracture life estimates close to the original data. On the other hand, because the test data is randomly sampled, it does not distinguish between S25C and S55C, which have different fatigue strengths, and therefore the estimated data is likely to vary. It is unclear which analyzed data for S25C and S55C correspond to (because the data was randomly sampled), but it is believed to be the bands indicated by the circles in the figure.

[0049] Next, we examined 10 types of steels (S25C, S35C, S55C, SNCM439, SMn438, SMn443, SUS403, SUS304) included in the fatigue data sheet. 6 Using fracture data (total of 24,784) for fractures less than 100 times, Vickers hardness, tensile strength, fracture elongation, and fracture area were estimated in conjunction with decision tree models. The analysis results are shown in Table 3 and Fig. 9. Compared to the analysis results for the two steel types S25C and S55C in Fig. 8, the MAPE shown in Fig. 9 was 29.8%, a slight improvement in estimation accuracy. This result is thought to be due to the fact that the total amount of data has increased five-fold compared to Fig. 8, and it is expected that estimation accuracy will continue to improve as more experimental data is collected in the future.

[0050] Estimation of the SN curve for S45C steel 5x10 of various steel materials (see Table 3) 6 Using machine learning with a combination of decision tree models for Vickers hardness, tensile strength, breaking elongation, and breaking reduction based on breaking data (total of 2834) of 5×10 6 The relationship between the Vickers hardness and the fracture life at each stress amplitude was calculated using the mechanical properties of S45C. 7 Due to the relationship between the cycle fatigue limit and the Vickers hardness of S45C, 7The fatigue limit was calculated using the Vickers hardness test. The analysis results are shown in Figure 10. Experimental data is indicated by △ and estimated data by ●. First, looking at the fatigue limit estimation, the fatigue limit estimated from the Vickers hardness was in excellent agreement with the Vickers hardness test, as shown in Table 1, with an estimation accuracy of 99% and a MAPE of 1.76. 6 The estimation of the fracture life of less than 1000 times also showed good agreement, despite the MAPE being 29.8%. This result is thought to be due to the fact that the test data did not include steel types with different strengths, as shown in Figures 8 and 9, so there was no variation in prediction accuracy. In this way, machine learning estimation of the SN curve can be performed with high accuracy by estimating the fracture life and fatigue limit separately. This result shows that it is possible to estimate the SN curve by utilizing the experimental data listed in the accumulated fatigue data sheets.

[0051] In the above embodiment, the various steel types included in the fatigue data sheet are shown to include carbon steel, nickel-chromium-molybdenum steel, manganese steel, and stainless steel, but the present invention is not limited to this, and aluminum alloys, titanium alloys, composite materials of plastic materials and various metals, and composite materials of various ceramic materials and various metals may also be used. [Industrial Applicability]

[0052] According to the method and apparatus for estimating fracture life of the present invention, by referring to the fatigue index characteristics, the fatigue characteristics of the relevant material can be roughly understood. By estimating the SN curve (stress amplitude-fracture life relationship) by machine learning, it is possible to estimate the fracture life of the material by 10 times the experimental value. 7 Therefore, it is possible to evaluate the long-term reliability of various machines and structures that use the relevant metallic materials or various composite materials, and it can also be used for safety factor design, and it can be used for the design and reliability evaluation of various machines and structures. [Explanation of symbols]

[0053] 300 Computer Program Products 302 Program-carrying media 304 Program instruction storage unit 306 Computer-readable medium 308 Recordable Media 309 Communication media 310 Specified reading function unit 320 Fatigue data sheet reading function 325 Preprocessing section for supervised machine learning 330 Machine Learning Calculation Unit 335 Machine Learning Evaluation Department 340 Assist function unit

Claims

1. a functional unit that reads the specification of the type of fatigue limit estimation and the specification of the mechanical properties used in the machine learning decision tree; A function to read fatigue data for a specified steel type from a fatigue data sheet; a machine learning calculation unit that performs supervised machine learning on fatigue data of the loaded steel type using the specified mechanical properties, a machine learning calculation unit that performs the supervised machine learning to output a calculation result of fatigue limit estimation according to a specified type of fatigue limit estimation using specified mechanical characteristics, a machine learning evaluation unit that evaluates the accuracy of the fatigue limit estimation performed by the machine learning calculation unit using a calculation result of the fatigue limit estimation according to the type of fatigue limit estimation used for the calibration target; and A fracture life estimation device having an assist function unit that recommends, based on the evaluation results of the machine learning evaluation unit, the types of mechanical properties to be used in the machine learning decision tree and the types of fatigue data for the steel type to be read in the fatigue data sheet, so as to improve the accuracy of the fatigue limit estimation of the machine learning calculation unit.

2. The fracture life estimation device according to claim 1, further comprising a pre-processing unit that performs machine learning on fatigue data of the loaded steel type using the specified mechanical properties, thereby performing the supervised machine learning.

3. The device for estimating a fracture life according to claim 2 , wherein the model used for the machine learning includes any one of a random forest method, deep learning, a neural network, and a model using LASSO regression.

4. The fatigue limit estimation types are: ７ Estimation of fatigue limit, 10 ６ 4. The device for estimating fracture life according to claim 1, wherein the device includes at least one of estimation of cycle fatigue limit, estimation of SN curve, and estimation of low cycle fatigue life.

5. The fracture life estimation device according to any one of claims 1 to 4, wherein the mechanical properties used in the machine learning decision tree include at least one of Vickers hardness, tensile strength, fracture elongation, fracture reduction, stress ratio, or fatigue test method.

6. The device for estimating fracture life according to claim 5 , wherein the fatigue testing method includes rotating bending test data, axial load test data, and torsion test data.

7. 7. The device for estimating fracture life according to claim 1, wherein the steel type includes at least one of carbon steel, nickel-chromium-molybdenum steel, manganese steel, and stainless steel.

8. The fracture life estimation device according to claim 7, wherein the steel types include at least one of S25C, S35C, and S55C for carbon steel materials, SNCM439 for nickel-chromium-molybdenum steel materials, SMn438 and SMn443 for manganese steel materials, and SUS403 and SUS304 for stainless steel materials.

9. a step of specifying the type of fatigue limit estimation and the mechanical properties to be used in the machine learning decision tree; reading fatigue data for a specified steel type from a fatigue data sheet; and outputting a calculation result of fatigue limit estimation according to a specified type of fatigue limit estimation using the specified mechanical properties to a machine learning calculation unit that has performed supervised machine learning on fatigue data of the loaded steel type using the specified mechanical properties. A method for estimating fracture life, comprising: Furthermore, a step of evaluating the accuracy of the fatigue limit estimation by the machine learning calculation unit that performed the supervised machine learning using the calculation result of the fatigue limit estimation according to the type of fatigue limit estimation used for the calibration object; A method for estimating fracture life, comprising a step of recommending, based on the evaluation results of the machine learning evaluation step, the types of mechanical properties to be used in the machine learning decision tree and the types of fatigue data for the steel type to be read into the fatigue data sheet, so as to improve the accuracy of the fatigue limit estimation of the machine learning calculation unit.

10. The method for estimating a fracture life according to claim 9, further comprising a step of performing supervised machine learning on the fatigue data of the steel type read in by the machine learning calculation unit using the specified mechanical properties.

11. The model used for the machine learning includes any of a random forest method, deep learning, neural network, or LASSO regression model. The method for estimating fracture life according to claim 10.

12. The fatigue limit estimation types are: ７ Estimation of fatigue limit, 10 ６ Estimation of the fatigue limit, estimation of the S-N curve, or estimation of the low cycle fatigue life, The method for estimating fracture life according to any one of claims 9 to 11.

13. 13. The method for estimating fracture life according to claim 9, wherein the mechanical properties used in the machine learning decision tree include at least one of Vickers hardness, tensile strength, fracture elongation, fracture reduction, stress ratio, or fatigue test method.

14. 14. The method of estimating rupture life of claim 13, wherein the fatigue testing methods include rotating bending test data, axial load test data, and torsion test data.

15. 15. The method for estimating a fracture life according to claim 9, wherein the steel type includes at least one of a carbon steel material, a nickel-chromium-molybdenum steel material, a manganese steel material, and a stainless steel material.

16. The method for estimating a fracture life according to claim 15, wherein the steel type includes at least one of S25C, S35C, and S55C for carbon steel materials, SNCM439 for nickel-chromium-molybdenum steel materials, SMn438 and SMn443 for manganese steel materials, and SUS403 and SUS304 for stainless steel materials.

17. A method for estimating a fracture life described in any one of claims 9 to 16, further comprising a step of modifying the type of fatigue limit estimation, the type of fatigue limit estimation specified for use in the step of specifying the mechanical properties to be used in the machine learning decision tree, and the mechanical properties to be used in the machine learning decision tree, taking into account the type of mechanical properties to be used in the recommended machine learning decision tree and the steel type read in the fatigue data sheet.

Images