Life evaluation device, program thereof, and method for evaluating life

Abstract

To more easily perform life evaluation of a material in long-term durability due to repeated fatigue by nondestructive inspection.SOLUTION: A life evaluation device 10 includes: a model creation unit 12 for creating a life evaluation model which derives life information relating to prediction of the remaining life of a material and estimation of a fracture position when analysis image data representing the distribution of the physical quantity corresponding to fatigue of the material used under a predetermined environment is input as physical quantity data; and a prediction unit 13 that outputs the life information relating to repeated fatigue of the material at the time of acquisition of the analysis image data in a non-destructive state by using the life evaluation model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a life evaluation device, a program therefor, and a life evaluation method for evaluating the life of a material in terms of long-term durability under a predetermined environment. [Background technology]

[0002] Fiber-reinforced plastics (FRPs) are a type of composite material that combines plastic materials with fibers. Short-fiber FRPs, in particular, have excellent specific strength and specific stiffness. They are used as lightweight structural materials, and demand for them has been increasing in recent years. However, compared to long-fiber FRPs, short-fiber FRPs have randomly arranged fibers, which means that their strength depends on the fiber direction, affecting fatigue properties. This can lead to unexpected fractures during long-term use as structural materials. Therefore, for structural materials such as short-fiber FRPs, the complex damage behavior caused by cyclic fatigue requires nondestructive testing to detect the onset of damage early and track its progression.

[0003] Incidentally, various mechanical models have been proposed as a method for evaluating the life of a material by non-destructive testing, such as the method disclosed in Patent Document 1. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2023-157601 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the material life assessment method in Patent Document 1 is a pre-use assessment method under a predetermined assumed use environment, and is not intended to predict long-term durability at the stage of actual use as a structural member. Furthermore, for example, in aircraft and other devices that are designed for damage tolerance, non-destructive testing is periodically performed to ensure safety, but since it is difficult to determine when and where a member will break, the testing requires a great deal of effort and cost.

[0006] The present invention has been devised with a focus on these problems, and its purpose is to provide a life assessment device, a program therefor, and a life assessment method that can more easily assess the life of a material in terms of its long-term durability due to repeated fatigue through non-destructive testing. [Means for solving the problem]

[0007] In order to achieve the above-mentioned object, the present invention is primarily an apparatus for evaluating the lifespan of a material used in a specified environment using a machine learning model, and is configured to predict the remaining lifespan of the material due to repeated fatigue at the time of acquisition of the analysis image data in a non-destructive state, and / or estimate the future location of failure of the material, based on analysis image data representing the distribution of physical quantities corresponding to fatigue of the material.

[0008] Furthermore, the present invention is primarily a method for evaluating the lifespan of a material used in a specified environment using a machine learning model, and employs the following technique: a lifespan evaluation model is created that derives lifespan information regarding the lifespan of the material using analytical image data representing the displacement distribution or strain distribution of the material obtained by a digital image correlation method as a physical quantity; the analytical image data in a non-destructive state is input into the lifespan evaluation model, thereby predicting the remaining lifespan due to cyclic fatigue of the material at the time the analytical image data is obtained, and / or estimating the location of failure of the material. [Effects of the Invention]

[0009] According to the present invention, unlike prediction of damage areas under static loads on materials using predetermined simulations, analysis image data actually obtained through fatigue testing of materials can be used to predict the remaining life of materials at the time of acquisition and estimate future fracture locations. This allows for easy fatigue life prediction taking damage morphology into account at a relatively early stage. Furthermore, by using analysis image data representing the displacement or strain distribution of materials obtained by digital image correlation as input data for machine learning models, this method can more effectively estimate fracture locations when assessing life spans based on damage accumulation (strain history) in cyclic fatigue, which is a very complex phenomenon, and can further improve the accuracy of remaining life predictions focusing on the fracture locations. Furthermore, by intermittently using analysis image data corresponding to the maximum load, where data features are prominent, from cyclic load data obtained in prior fatigue testing of materials, as training data, it is possible to effectively reduce the data volume of analysis image data used for machine learning. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a schematic configuration diagram of a life evaluation device according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram for explaining an image showing the strain distribution of a material obtained by a digital image correlation method. [Figure 3] 1 is a graph showing the relationship between time and stress in a fatigue test. [Figure 4] FIG. 10 is a diagram for explaining an image that visualizes the fracture position on the material surface. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0012] Figure 1 shows a schematic diagram of a life assessment device according to this embodiment. In this figure, the life assessment device 10 is a device that uses a machine learning model to assess the lifespan of various materials, such as mechanical structural members, used in a specified environment. The device is configured with a computer including a processor such as a CPU and storage devices such as a memory and a hard disk, and has installed thereon programs that cause the computer to function as the following components. As will be described later, the life assessment device 10 is configured to predict the remaining lifespan of materials in the long-term durability test and estimate the location of failure.

[0013] The material to be evaluated for life span in the present invention is not particularly limited, and any material such as resin or metal can be applied. In this embodiment, as an example, aramid fiber reinforced plastic material used for gear materials etc. is evaluated, and experiments are conducted to demonstrate the effects etc.

[0014] The life assessment device 10 includes a memory unit 11 that stores various information, processing results, and other data; a model creation unit 12 that creates a life assessment model consisting of a machine learning model that derives life information related to predicting the remaining life of a material and estimating the location of failure when analytical image data representing the distribution of physical quantities corresponding to the fatigue of the material is input as physical quantity data; and a prediction unit 13 that outputs life information due to repeated fatigue of the material at the time of acquiring the analytical image data of the material in a non-destructive state using the life assessment model.

[0015] The analysis image data used is the image data shown in Figure 2, which is obtained by a digital image correlation method in which the speckle pattern on the material surface is captured with a camera and the displacement or strain distribution is measured. The image data is an image that represents the strain distribution on the material surface using multiple shades of color. Note that any image data that represents a physical quantity corresponding to material fatigue can be used as the analysis image data of the present invention. For example, image data representing the displacement distribution obtained by digital image correlation or image data representing the temperature distribution obtained by thermography can also be used.

[0016] The model creation unit 12 performs a fatigue test on the target material over many cycles (for example, 900 to 70,000 cycles) until fracture, and creates the life assessment model by learning analytical image data obtained by a digital image correlation method at predetermined times during the fatigue test. The fatigue test here is a test under any loading mode, such as tension or compression.

[0017] The machine learning model used here is not particularly limited, and various models can be adopted as long as they achieve the same functions and effects as those described below. In this embodiment, a convolutional neural network (CNN) model is used. This CNN is a type of machine learning model mainly used as an image model for image recognition, pattern recognition, etc., and is capable of understanding complex patterns through hierarchical learning by combining layers such as convolutional layers and pooling layers. In particular, in this embodiment, a deep convolutional neural network (DCNN) that combines many of these layers is used, and a regression model with an output layer value of 1 is used. In DCNN, model learning is performed to optimize a loss function with respect to a target variable. In the present invention, the target variable is the fatigue life ratio (flr) defined by the following formula, and the loss function is the well-known root mean square error (RMSE).

number

[0018] The procedure for creating a life evaluation model in the model creation unit 12 will be described below.

[0019] First, a fatigue test is conducted using a predetermined test material, and analytical image data is acquired for learning and evaluation using a machine learning model. In this fatigue test, a black random pattern is applied to the test piece before the test to acquire a strain distribution image through analysis using digital image correlation, and image data is acquired at any number of cycles during the test, for example, at a timing of 100 fps (20 images per cycle). Here, the image data acquired at each timing is calculated based on the number of cycles N at the time of data acquisition and the number of cycles to fracture N according to the experimental results. f The fatigue life ratio flr calculated from the above is stored in the storage unit 11 in a corresponding state.

[0020] The analysis image data used to create the life assessment model may be all analysis image data acquired during the fatigue test. However, from the perspective of reducing the data volume during the model creation process, it is preferable to perform extraction processing using a device including a program for intermittently acquiring effective analysis image data. In this extraction process, as shown in Figure 3, only the strain distribution image where the maximum strain is maximized corresponds to the maximum load (maximum stress portion: the portion indicated by the dashed dotted line in the figure) where the data feature values ​​are prominently displayed among the cyclic load data is extracted as analysis image data. In other words, here, the image data when the absolute value of the cyclic load during the fatigue test is maximized is extracted as analysis image data. Subsequently, image processing such as trimming to remove white space is performed, and the image data is used to create the life assessment model. This image extraction process allows the creation of the life assessment model to be performed more efficiently.

[0021] The analysis image data acquired in advance in this manner is randomly divided into training data and test data consisting of training data and validation data at a predetermined ratio. While the ratio is not particularly limited, for example, the ratio of training data to test data is 80%:20%, and the ratio of training data to validation data in the training data is 80%:20%. The DCNN model is then trained using the training data, and the model is evaluated using the validation data based on the RMSE. Training of the DCNN model is then terminated by minimizing the RMSE, and the accuracy of the model is confirmed using test data. A suitable life assessment model is identified and stored in the memory unit 11. If the accuracy of the model is found to be outside the acceptable range based on the test data, the accuracy can be improved by increasing the number of test specimens used in the fatigue test and retraining the model, and / or by ensemble training that combines multiple CNN models.

[0022] The prediction unit 13 uses the life evaluation model created by the model creation unit 12, and outputs the life information at the time of acquisition of the analysis image data, as follows, using the analysis image data acquired by the digital image correlation method for the non-destructive material to be evaluated for life, as a physical quantity.

[0023] First, the analytical image data of the material to be evaluated is input into the life assessment model. The life assessment model then calculates and outputs the predicted value of the fatigue life ratio (flr) at the time the analytical image data was acquired as the remaining life. Furthermore, using the well-known Grad-CAM, a method that provides a visual explanation for the CNN model's decisions, the estimated future fracture location of the material at the time the analytical image data was acquired is visualized and output. For example, as shown in Figure 4, the estimated fracture location is visualized using a color map in which the region of interest on the material surface (shown by the dashed-dotted line in the figure) used to derive the fatigue life ratio (flr) is highlighted, and the region or its vicinity is calculated as the estimated future fracture location of the material. Note that the remaining life in this invention can also be calculated from the life assessment model using other life information corresponding to the remaining life of the material.

[0024] According to the results of the inventors' experimental research, the life assessment model created by this invention takes into account the average strain of the entire material and places importance on the maximum strain value, and it has been demonstrated that it is possible to accurately predict the remaining life due to cyclic fatigue of the material and estimate the location of failure in the material at a relatively low stage when the fatigue life ratio flr is around 20%, that is, at a relatively early stage that cannot be determined by visual inspection of the analysis image data using the digital image correlation method.

[0025] The prediction unit 13 may be configured to obtain, as the life information, either a prediction of the remaining life due to repeated fatigue of the material or an estimation of the future fracture position of the material.

[0026] Furthermore, the configuration of each part of the device in the present invention is not limited to the illustrated configuration example, and various modifications are possible as long as they provide substantially the same effect. [Explanation of symbols]

[0027] 10 Life evaluation device 11 Storage section 12 Model Creation Department 13 Prediction Department

Claims

1. An apparatus for evaluating the lifespan of a material used in a predetermined environment using a machine learning model, A life assessment device characterized by predicting the remaining life due to repeated fatigue of the material at the time of acquisition of the analysis image data in a non-destructive state, and / or estimating the future location of failure of the material, based on analysis image data representing the distribution of physical quantities corresponding to fatigue of the material.

2. An apparatus for evaluating the lifespan of a material used in a predetermined environment using a machine learning model, A life assessment device characterized by comprising: a model creation unit that creates a life assessment model that derives life information related to predicting the remaining life of the material and / or estimating the location of failure when analytical image data representing the distribution of physical quantities corresponding to fatigue of the material is input as physical quantity data; and a prediction unit that outputs the life information due to repeated fatigue of the material in a non-destructive state at the time the analytical image data is acquired using the life assessment model.

3. 3. The life evaluation device according to claim 2, wherein the analysis image data is image data representing a displacement distribution or a strain distribution of the material, obtained by a digital image correlation method.

4. 4. The life evaluation device according to claim 3, wherein the model creation unit creates the life evaluation model using the analysis image data obtained at the maximum absolute load of the repeated load in a fatigue test of the material as learning data.

5. A program that causes a device to function that evaluates the lifespan of a material used in a specified environment using a machine learning model, A program for a life assessment device that causes a computer to function as: a model creation unit that creates a life assessment model that derives life information related to predicting the remaining life of the material and / or estimating the location of failure when analytical image data representing the distribution of physical quantities corresponding to fatigue of the material is input as physical quantity data; and a prediction unit that outputs the life information due to repeated fatigue of the material in a non-destructive state at the time the analytical image data is acquired using the life assessment model.

6. A method for evaluating the lifespan of a material used in a predetermined environment using a machine learning model, comprising: A life assessment method characterized by creating a life assessment model that derives life information regarding the life of the material using analytical image data representing the displacement distribution or strain distribution of the material obtained by a digital image correlation method as a physical quantity, and inputting the analytical image data in a non-destructive state into the life assessment model to predict the remaining life due to repeated fatigue of the material at the time the analytical image data was obtained and / or estimate the location of failure of the material.

7. 7. The life assessment method according to claim 6, wherein the analytical image data obtained in a fatigue test of the material conducted in advance is extracted and used as learning data for the life assessment model, and the analytical image data obtained at the maximum absolute load of the repeated load is used.

Images