Performance prediction system and performance prediction method

Abstract

To provide a performance prediction system which can perform advanced performance prediction in consideration of an influence of components of various panels for an automobile.SOLUTION: In a performance prediction system, a storage unit stores datasets generated by associating, for each load point, objective variables which are load displacement information on depressed load points with explanatory variables for each component, obtained by associating point cloud coordinate data of a point cloud formed in the vicinity of the load points arranged uniformly in a region having a yield point on a panel with materials and plate thicknesses of the points constituting the point cloud. In a control unit, a machine learning unit generates, by machine learning, a performance prediction model which predicts load displacement information of a performance prediction point using the datasets, and a performance prediction unit receives input data, which is a combination of point cloud coordinate data of a point cloud formed in the vicinity of the performance prediction point and materials and plate thicknesses of the points constituting the point cloud, and predicts load displacement information of the performance prediction point using the trained performance prediction model.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a performance prediction system and a performance prediction method for predicting a time-series load-displacement curve (F-S diagram) when pressing various metal panels such as an engine hood and a door of an automobile.

Background Art

[0002] For example, an engine hood of an automobile needs to have a rigidity such that it does not dent or deform under a certain impact, and high dent performance is required. In order to satisfy such requirements, performance prediction of the engine hood is indispensable.

[0003] For example, as an example of performance prediction of an engine hood or the like, Patent Document 1 below describes a dent rigidity prediction method for calculating the displacement when a load is applied to a measurement point of an engine hood from the information of a design drawing.

[0004] Specifically, when a load is applied to a measurement point, the area of a stress influence region defined around the measurement point is calculated as a warping area, and the curvature of the measurement point of the plate member, the plate thickness of the measurement point of the plate member, the material of the plate member, and the warping area are used as factors representing the rigidity of the plate member, and a predetermined relational expression is obtained by multiple regression analysis or the like. Then, by taking out the curvature of the measurement point of the plate member, the plate thickness of the measurement point of the plate member, the material of the plate member, and the warping area from the design drawing, the displacement (dent rigidity) in the load direction of the measurement point is appropriately calculated at the design stage based on the predetermined relational expression.

[0005] In addition, Patent Document 2 below describes a performance estimation device capable of accurately estimating the performance of an object having a three-dimensional shape.

[0006] For example, the performance estimation device described in Patent Document 2 below uses a pre-trained convolutional neural network to estimate the pedestrian protection performance of a vehicle's hood, taking an image showing the cross-sectional shape of the vehicle's hood as input data. Specifically, the performance estimation device obtains the following (values ​​A) to (values ​​D) as information relating to the vehicle's pedestrian protection performance (more specifically, head protection performance).

[0007] [Value A] Head injury threshold value [Value B] A value corresponding to the amount of intrusion of the head impactor into the vehicle hood for pedestrian protection testing. [Value C] A value corresponding to the average acceleration of the head impactor during the initial stage (1 millisecond) of a collision with the vehicle hood. [Value D] This value corresponds to the average acceleration of the head impactor during the initial stage (2 milliseconds) of a collision with the vehicle hood. [Prior art documents] [Patent Documents]

[0008] [Patent Document 1] Japanese Patent Publication No. 2007-33067 [Patent Document 2] Japanese Patent Publication No. 2020-166512 [Overview of the project] [Problems that the invention aims to solve]

[0009] In the dent stiffness prediction method described in Patent Document 1 above, the derivation of the correlation formula (a predetermined relational formula) and the calculation of the amount of deflection (dent stiffness) require measuring the curvature, plate thickness, material, and area of ​​the stress-influenced region defined around the measurement point when a load is applied to the measurement point each time, which has the problem of requiring a great deal of time.

[0010] Furthermore, the performance prediction methods described in Patent Documents 1 and 2 had a problem in that, for example, when internal components such as the engine hood affect performance, the performance prediction could not take that influence into account. In particular, when there are components that are highly correlated with the performance to be predicted and components that are not highly correlated, it is difficult to make a sophisticated performance prediction that takes their influence into account.

[0011] The present invention has been made in view of the above-mentioned problems, and aims to provide a performance prediction system and performance prediction method that can significantly reduce the time required to predict the performance of various panels for automobiles, and that can perform advanced performance predictions that take into account the influence of the components of various panels. [Means for solving the problem]

[0012] The performance prediction system according to the present invention is a performance prediction system that predicts the performance of metal panels constituting the body of an automobile, for example, Region on the panel where buckling occurs under a constant load. The system is characterized by comprising: a storage means for storing a dataset in which the explanatory variables are data relating point cloud coordinate data of a point cloud formed around uniformly provided load points and information on the material and plate thickness of each point constituting the point cloud, further classified into component units constituting the panel, and the explanatory variables are linked to a target variable which is information on the FS diagram when the load point is pressed (corresponding to the load displacement information in the embodiment described later); a model generation means for generating a performance prediction model by machine learning that predicts information on the FS diagram when a performance prediction point, which is an arbitrary position within the above region, is pressed using the dataset read from the storage means; a performance prediction means that receives input data relating point cloud coordinate data of a point cloud formed around the performance prediction point and information on the material and plate thickness of each point constituting the point cloud, further classified into component units constituting the panel, and predicts information on the FS diagram when the performance prediction point is pressed using the trained performance prediction model; and an FS diagram creation means for creating an FS diagram at the performance prediction point based on the predicted information on the FS diagram.

[0013] Furthermore, in the performance prediction system according to the present invention, the information regarding the FS diagram is extracted from the FS diagram of the load point, which has been acquired in advance by a known method. Descending It is desirable to use load-displacement information that represents the load and displacement amount corresponding to each of the multiple inflection points, including the inflection point.

[0014] Furthermore, in the performance prediction system according to the present invention, it is desirable to employ a convolutional neural network as the machine learning algorithm, to have an intermediate layer that extracts features for each component constituting the panel, and to determine the depth of the intermediate layer for each component according to the correlation between the predicted value and each component.

[0015] According to the performance prediction system of the present invention, by introducing a pre-trained performance prediction model, the time required to calculate the FS diagram when a performance prediction point on the panel is pressed can be significantly reduced. Furthermore, the distinctive configuration of the convolutional neural network enables advanced performance prediction that takes into account the correlation between each component constituting the panel.

[0016] Furthermore, the performance prediction method according to the present invention is a performance prediction method for predicting the performance of metal panels that constitute the body of an automobile, Region on the panel where buckling occurs under a constant load.The system includes: a dataset storage step in which data relating point cloud coordinate data of a point cloud formed around uniformly provided load points and information on the material and thickness of each point constituting the point cloud, further classified into component units constituting the panel, are used as explanatory variables, and a target variable, which is information on the FS diagram when a load point is pressed, is linked to each load point, and a dataset is stored in memory; a model generation step in which a performance prediction model is generated by machine learning using the dataset read from memory to predict information on the FS diagram when a performance prediction point, which is an arbitrary position within the above region, is pressed; a performance prediction step in which input data relating point cloud coordinate data of a point cloud formed around a performance prediction point and information on the material and thickness of each point constituting the point cloud, further classified into component units constituting the panel, is received, and the trained performance prediction model is used to predict information on the FS diagram when a performance prediction point is pressed; an FS diagram creation step in which an FS diagram at the performance prediction point is created based on the predicted information on the FS diagram; and a display step in which the FS diagram created in the FS diagram creation step is displayed on a display.

[0017] Furthermore, in the performance prediction method according to the present invention, it is desirable that the information related to the FS diagram be load-displacement information, which is the load and displacement amount corresponding to each of the multiple inflection points, including the yield point, extracted from the FS diagram of the load point that has been acquired in advance by a known method.

[0018] In addition, in the performance prediction method according to the present invention, the 3D design information of the panel and the information on the material and thickness of the panel associated with the 3D design information are stored in advance in the memory. Then, as preprocessing for generating a performance prediction model, based on the 3D design information of the panel, a shape arrangement step of arranging the panel in a 3D coordinate space (X coordinate, Y coordinate, Z coordinate), a cross-sectional shape acquisition step of acquiring a plurality of cross-sectional shapes of the panel centered on the load point, a point group generation step of generating the above-mentioned point group which is a collection of the center points by dividing the obtained cross-sectional shape with a square of a predetermined angle, a point group coordinate data generation step of acquiring the coordinates (X coordinate, Y coordinate, Z coordinate) of each point forming the point group to generate the above-mentioned point group coordinate data, and based on the point group coordinate data generated in the point group coordinate data generation step, the 3D design information of the panel, the information on the material and thickness of the panel associated with the 3D design information, and the load displacement information of each load point, a data set generation step of generating a data set for model generation, are included.

[0019] Furthermore, in the performance prediction method according to the present invention, as preprocessing for the performance prediction step, the above-mentioned shape arrangement step, the above-mentioned cross-sectional shape acquisition step of acquiring a plurality of cross-sectional shapes of the panel centered on the performance prediction point, the above-mentioned point group generation step, the above-mentioned point group coordinate data generation step, and based on the point group coordinate data generated in the point group coordinate data generation step, the 3D design information of the panel, and the information on the material and thickness of the panel associated with the 3D design information, an input data generation step of generating input data for performance prediction, are included.

[0020] According to the performance prediction system of the present invention, by introducing a learned performance prediction model, the time required for calculating the F-S diagram when pressing the performance prediction point on the panel can be significantly shortened. In addition, due to the characteristic configuration in the convolutional neural network, advanced performance prediction considering the correlation relationship of each component constituting the panel becomes possible.

Effects of the Invention

[0021] According to the performance prediction system and performance prediction method of the present invention, the time required for performance prediction of various panels for automobiles can be significantly shortened, and furthermore, advanced performance prediction considering the influence degree of the components of various panels can be realized.

Brief Description of Drawings

[0022] [Figure 1] FIG. 1 is a diagram showing an example of the hardware configuration of a computer operating as a performance prediction system according to the present invention. [Figure 2] FIG. 2 is a functional block diagram showing the functions of a control unit that performs learning model generation processing. [Figure 3] FIG. 3 is a flowchart showing an example of learning model generation processing. [Figure 4] FIG. 4 is a diagram showing an example of an F-S diagram. [Figure 5] FIG. 5 is a diagram showing an example of a 3D image of an engine hood. [Figure 6] FIG. 6 is a diagram showing the acquisition position of a cross-sectional shape. [Figure 7] FIG. 7 is a diagram showing an example of a cross-sectional shape of an engine hood and an image of the state where the cross-sectional shape is grouped into point clouds. [Figure 8] FIG. 8 is a diagram showing an example of point cloud coordinate data. [Figure 9] FIG. 9 is a diagram showing an example of a data set for generating a performance prediction model. [Figure 10] FIG. 10 is a diagram showing an example of the configuration of a machine learning unit. [Figure 11] FIG. 11 is a diagram showing the evaluation result of a model by machine learning. [Figure 12] FIG. 12 is a diagram showing the prediction result by a learned performance prediction model using test data. [Figure 13] FIG. 13 is a functional block diagram showing the functions of a control unit that performs performance prediction processing. [Figure 14] FIG. 14 is a flowchart showing an example of performance prediction processing. [Modes for carrying out the invention]

[0023] Hereinafter, embodiments of the performance prediction system and performance prediction method according to the present invention will be described in detail with reference to the drawings. However, the present invention is not limited by these embodiments. Furthermore, in the specification and drawings of this application, elements that can be similarly described are denoted by the same reference numerals, thereby omitting redundant explanations.

[0024] <System Configuration> Figure 1 shows an example of the hardware configuration of a computer operating as a performance prediction system according to the present invention. The performance prediction system 1 of this embodiment operates as a host computer that performs the process of creating an FS diagram (time-series load displacement curve) when a predetermined position on various metal panels for automobiles is pressed (hereinafter referred to as the performance prediction process), and the process of generating a learning model that performs prediction of information related to the FS diagram (hereinafter referred to as the load displacement information) (hereinafter referred to as the learning model generation process).

[0025] In Figure 1, the performance prediction system 1 comprises a control unit 11 consisting of a CPU (Central Processing Unit) and an FPGA (Field Programmable Gate Array), a storage unit 12 including various types of memory such as ROM (Read Only Memory) and RAM (Random Access Memory), an input unit 13 including a user interface such as a keyboard and mouse, an interface (I / F) unit 14 that performs input / output processing such as printing and scanning, a display unit 15 which is a display, and a communication unit 16 that communicates with the outside via a predetermined network. In Figure 1, the system is shown to include an input unit 13 including a user interface such as a keyboard and mouse, but the performance prediction system 1 of this embodiment is not limited to this, and the system may be configured without an input unit 13, or in combination with an input unit 13, by giving the display unit 15 a touch panel function.

[0026] In Figure 1, the control unit 11 executes, for example, a performance prediction program that creates an FS diagram when a predetermined position on various automobile panels is pressed, and a learning model generation program that generates a learning model for predicting load displacement information, in order to realize the performance prediction processing and learning model generation processing by the performance prediction system 1 of this embodiment. The storage unit 12 stores programs related to the performance prediction processing and learning model generation processing of this embodiment (performance prediction program, learning model generation program), various information (3D shape data showing the shape of various panels, material and thickness, and datasets for machine learning, etc.), and various data obtained during the processing (point cloud coordinate data, etc.). The control unit 11 executes the performance prediction processing and learning model generation processing of this embodiment by reading the various programs stored in the storage unit 12. Note that the storage unit 12 is not limited to internal memory, but may be an external storage medium such as a DVD (Digital Versatile Disc) or SD memory, or it may consist of both internal memory and an external storage medium (DVD or SD memory, etc.). Furthermore, for the sake of explanation, the hardware configuration of the performance prediction system 1 in this embodiment is a list of the configurations related to the performance prediction processing and learning model generation processing in this embodiment, and does not represent all the functions of the computer that constitutes the performance prediction system 1.

[0027] Furthermore, while the performance prediction system 1 of this embodiment is intended for general-purpose PCs such as desktop computers and notebook computers, it is not limited to these, and may also be a mobile device such as a smartphone or tablet.

[0028] <Learning model generation process> Next, before describing the performance prediction process in the performance prediction system 1 of this embodiment, we will explain the learning model generation process that is a prerequisite for it. Figure 2 is a functional block diagram showing the functions of the control unit 11 that performs the learning model generation process, and Figure 3 is a flowchart showing an example of the learning model generation process.

[0029] In Figure 2, the control unit 11 includes a point cloud coordinate data generation unit 21, a dataset generation unit 22, a machine learning unit 23 that generates a learning model using a machine learning algorithm, and a known CAD (computer-aided design) system 24 and a CAE (computer-aided engineering) system 25. In this embodiment, as an example of various metal panels, the case of generating a learning model using 3D design information of an automobile engine hood will be described. That is, in this embodiment, as an example of a learning model, a performance prediction model is generated that predicts load displacement information (load and displacement amount corresponding to the six inflection points described later) when a predetermined position on the automobile engine hood is pressed. This performance prediction model is generated using a known machine learning algorithm, such as a neural network (CNN: Convolutional Neural Network, RNN: Recurrent Neural Network, etc.).

[0030] Furthermore, the engine hood of the automobile is designed using, for example, the CAD system 24 shown in Figure 2, and 3D design information, including 3D shape data, is pre-stored in the storage unit 12. In addition, information such as the material and thickness of each component constituting the engine hood is also pre-stored in the storage unit 12 in association with the 3D design information (3D shape data).

[0031] Furthermore, in this embodiment, as a prerequisite for performing the learning model generation process, FS diagrams for each load point (for example, 100 locations) on the engine hood are acquired in advance when a specific position (load point: for example, 100 locations) on the engine hood is pressed, and these FS diagrams are stored in the storage unit 12 in association with each load point. Specifically, for example, the CAE system 25 shown in Figure 2 reads the 3D design information of the engine hood from the storage unit 12, performs CAE analysis using the finite element method based on this 3D design information, and obtains FS diagrams (performance information) corresponding to each load point as an analysis result. In CAE analysis using the finite element method, the 3D shape of the engine hood is divided into a mesh, and by applying loads to the load points, the displacement of each part of the divided mesh is analyzed. Figure 4 shows an example of an FS diagram obtained by the prerequisite process.

[0032] Furthermore, since the 3D design using the CAD system 24 and the CAE analysis using the CAE system 25 are performed using well-known and common methods, a detailed explanation will be omitted. In this embodiment, the FS diagrams at each load point (100 locations) are obtained in advance by performing CAE analysis using the finite element method based on the 3D design information, but this is not the only method, and any known method may be used to obtain the FS diagrams at each load point.

[0033] Furthermore, in this embodiment, as an example, the number of load points is set to 100, but this is not limited to this, and it is desirable to have, for example, 100 or more load points depending on the required accuracy of the performance prediction processing. In this embodiment, each load point is positioned as a specific location on the engine hood as described above, for example, so as to be uniformly (evenly distributed) within the region on the engine hood that has a yield point (the region where buckling occurs under a certain load). Specifically, each load point placed on the outer part of the engine hood is appropriately placed within a region directly below which there are no seating surfaces (for example, internal parts such as dent R / F or lock R / F) that are joined by deflection (displacement). In other words, in the prerequisite processing described above, FS diagrams corresponding to the 100 load points within the region on the engine hood that has a yield point are obtained in advance (see Figure 4).

[0034] Furthermore, in the functional block shown in Figure 2, the CAD system 24 and the CAE system 25 are included within the control unit 11. However, this configuration is not limited to this, and for example, the CAD system 24 and the CAE system 25 may be configured separately and connected to the performance prediction system 1 via a network.

[0035] In this embodiment, assuming that various types of information are pre-stored in the storage unit 12 as described above, for example, as shown in Figure 3, the control unit 11 performs a process to generate a performance prediction model (learning model generation process) using the 3D design information of the automobile engine hood, the material and thickness information of the engine hood, and the FS diagrams at each load point that have been acquired in advance.

[0036] In Figure 3, first, when the operator instructs the generation of a performance prediction model by operating the input unit 13, the point cloud coordinate data generation unit 21 of the control unit 11 reads the 3D design information of the engine hood from the storage unit 12 and places the engine hood (shape) in 3D coordinate space (X coordinate, Y coordinate, Z coordinate) based on the 3D shape data obtained from the 3D design information (step S1). At this time, a 3D image of the engine hood is displayed on the display unit 15. Figure 5 shows an example of a 3D image of the engine hood.

[0037] Next, the point cloud coordinate data generation unit 21 acquires the cross-sectional shape of the engine hood around the load points (100 locations) from the 3D design information (step S2). Specifically, it acquires the cross-sectional shape of the engine hood with a width of 150 mm in the front, rear, left, and right directions, centered on the load point. Figure 6 shows the acquisition locations of the cross-sectional shapes. In this embodiment, the cross-sectional shape of the engine hood with a width of 150 mm in the left-right direction centered on the load point (see the A-A' line (a line parallel to the X-axis) centered on load point C shown in Figure 6) and the cross-sectional shape of the engine hood with a width of 150 mm in the front-rear direction centered on the load point (see the B-B' line (a line parallel to the Y-axis) centered on load point C shown in Figure 6) are acquired for each load point.

[0038] Figures 7(a) and (b) show examples of cross-sectional shapes of the engine hood. Specifically, (a) shows the A-A' cross-section at load point C, and (b) shows the B-B' cross-section at load point C. In this embodiment, the width of the cross-sectional shape to be acquired is set to 150 mm as an example, but it is not limited to this. It is sufficient if the cross-sectional shape is acquired uniformly (evenly distributed) from the entire panel. For example, it can be set appropriately depending on the shape of the panel, the material and thickness, the stress range when a load is applied, the required accuracy of performance prediction, etc. In this embodiment, two orthogonal cross-sectional shapes are acquired, such as the A-A' and B-B' cross-sections shown in Figure 6, but it is not limited to this. For example, in order to improve the accuracy of performance prediction, cross-sectional shapes in other directions centered on the load point may also be acquired.

[0039] Furthermore, the point cloud coordinate data generation unit 21 divides the cross-sectional shape (150 mm wide) acquired for each load point into, for example, 5 mm intervals (5 mm x 5 mm squares), and generates a point cloud of the cross-sectional shape (a collection of 5 mm square center points) for each load point (step S3). The spacing between each point constituting the point cloud was set to 5 mm based on the balance between the shape representation accuracy (fillet radius, etc.) and the number of features when the cross-sectional shape is converted into a point cloud. Then, the point cloud coordinate data generation unit 21 acquires the coordinates (X coordinate, Y coordinate, Z coordinate) of each point forming the point cloud based on the 3D shape data of the engine hood placed in the 3D coordinate space, generates point cloud coordinate data of the cross-sectional shape for each load point (step S4), and stores it in the storage unit 12. Figures 7(c) and 7(d) are schematic diagrams showing, for example, an image of the cross-sectional shape acquired at load point C in Figure 6, which has been converted into a point cloud. Specifically, (c) shows the point cloud image of the A-A' cross section of load point C, and (b) shows the point cloud image of the B-B' cross section of load point C. Figure 8 is a diagram showing an example of point cloud coordinate data. In this embodiment, the spacing between points in the point cloud was set to 5 mm as an example, but this is not limited to this, and can be appropriately set according to, for example, the shape of the panel, the material and thickness of the plate, the stress range when a load is applied, the required accuracy of the performance prediction, etc.

[0040] As described above, after the point cloud coordinate data generation unit 21 generates point cloud coordinate data for each load point, the dataset generation unit 22 then reads the point cloud coordinate data, 3D design information (including information on the material and thickness of each component constituting the engine hood), and pre-stored FS diagrams (performance information) for each load point from the storage unit 12. Then, it links the coordinates of each point forming the point cloud coordinate data, the material and thickness of each point, and the load displacement information obtained from the FS diagram for each load point to generate a dataset for machine learning (step S5). In this embodiment, the dataset generation unit 22 extracts, for example, six inflection points (for example, in order of decreasing curvature) including the yield point from each of the FS diagrams for each load point (see Figure 4), and generates a dataset for machine learning using the load and displacement amounts corresponding to each of these six inflection points as load displacement information. In this embodiment, six inflection points are extracted as an example, but the number of extracted points is arbitrary, as long as the FS diagram can be derived from the extracted inflection points.

[0041] In other words, the dataset used in the machine learning algorithm consists of explanatory variables (input data) that associate the coordinates of each point forming point cloud coordinate data corresponding to a single load point with the material and plate thickness of each point, and load displacement information corresponding to that load point as the target variable (training data). The training data corresponding to each load point is individually linked to the input data prepared for each load point.

[0042] Furthermore, this dataset is generated by classifying (subdividing) the input data prepared for each load point into individual components that make up the engine hood, as determined by the dataset generation unit 22. Here, the components that make up the engine hood are, for example, the outer hood, inner hood, dent reinforcement, lock reinforcement, etc. The dataset generation unit 22 then stores the datasets for each load point (C-1, C-2, C-3, C-4, C-5…) classified into individual components in the storage unit 12. Figure 9 shows an example of a dataset for generating a performance prediction model.

[0043] The machine learning unit 23 of the control unit 11 reads 89 datasets from the memory unit 12 out of the dataset of 100 load points shown in Figure 9, and generates a performance prediction model that predicts load displacement information when a predetermined position on the engine hood of an automobile is pressed by performing supervised learning on the dataset, for example using a convolutional neural network (CNN), which is one of the machine learning algorithms (step S6).

[0044] In other words, in this embodiment, a performance prediction model is generated that predicts load displacement information when a predetermined position on the engine hood of an automobile is pressed, by using a machine learning algorithm to learn the correlation between the input data (explanatory variables) and training data (dependent variables) described above. This provides a performance prediction model that can predict load displacement information at a predetermined position on the engine hood that has a yield point.

[0045] <An example of a machine learning algorithm> Figure 10 shows an example of the configuration of a machine learning unit 23, which corresponds to a convolutional neural network (CNN) that performs machine learning. The machine learning algorithm used in this embodiment is, for example, a convolutional neural network, and comprises an input unit 31 that forms an input layer for input data of individual parts constituting the engine hood, a feature extraction unit 32 that learns to extract different features for each layer by forming multiple intermediate layers that extract features using convolutional operations with filters (weights), and an output unit 33 that has multiple intermediate layers that are fully connected after smoothing the features extracted by the feature extraction unit 32 and an output layer that outputs the load displacement information described above. The output unit 33 corresponds to the intermediate layers and output layer of a typical fully connected neural network.

[0046] In the machine learning unit 23 of this embodiment, the input unit 31 receives input data (explanatory variables) for each component constituting the engine hood, for each load point. The machine learning unit 23 then calculates the mean squared error between the predicted value (load and displacement corresponding to the six inflection points), which is the output of the output layer in the output unit 33, and the training data (load and displacement information). It then repeatedly performs learning over a predetermined number of epochs to minimize this error. This trains a performance prediction model.

[0047] Furthermore, in the machine learning unit 23 of this embodiment, the feature extraction unit 32 is configured to extract feature quantities for each component that makes up the engine hood (outer, inner, dent R / F, lock R / F, etc.), and further determines the depth (size) of the intermediate layer for each component according to the correlation (magnitude of influence) between the predicted value and each component. For example, the outer, which is directly subjected to load and has the greatest correlation with the predicted value, has the deepest intermediate layer, the inner, which has the next greatest correlation, has a shallower intermediate layer than the outer, and for components with an even smaller correlation with the predicted value than these components (dent R / F, lock R / F), the shallowest intermediate layer is set. In this embodiment, as an example, the depth of the intermediate layer for the outer is set to 3, and the feature quantities obtained from each intermediate layer are set to 64, 32, and 16, respectively. Also, the depth of the intermediate layer for the inner is set to 2, and the feature quantities obtained from each intermediate layer are set to 32 and 16. Also, the depth of the intermediate layer for the dent R / F is set to 1, and the feature quantity obtained from this intermediate layer is set to 32. The depth of the intermediate layer of the lock R / F was set to 1, and the number of features obtained from this intermediate layer was set to 16. This makes it possible to perform advanced performance predictions that take into account the correlation between each component that makes up the engine hood.

[0048] In this embodiment, as described above, a dataset of 100 points (100 load points) was prepared for machine learning. Of these, 89 points were used as the dataset for generating the performance prediction model; that is, 80 points were used as training data and 9 points as evaluation data. The remaining 11 points were used as test data. Here, the training data is the dataset used to train the performance prediction model, the evaluation data is the dataset used to verify the model's performance, and the test data is the dataset used to perform the final evaluation of the model by comparing it with the training data. Furthermore, the training data is prepared in a quantity sufficient to obtain the desired prediction accuracy (for example, a correlation coefficient R of 0.8 or higher) without causing overfitting or underfitting. In addition, for performance evaluation of the performance prediction model in this embodiment, the loss function and mean absolute error (MAE), which are examples of evaluation metrics, are used. That is, the smaller these values ​​are, the less error the model has. Note that the evaluation metrics are not limited to these.

[0049] Figure 11 shows the evaluation results of a machine learning model. (a) is the learning curve for load (the load corresponding to each of the six inflection points: vertical axis of the FS diagram), which is one of the load-displacement information, and (b) is the learning curve for displacement (the displacement corresponding to each of the six inflection points: horizontal axis of the FS diagram), which is another of the load-displacement information. Here, for example, the changes in loss (loss function) and mean absolute error with respect to the degree of learning (epochs) are shown. As shown in Figure 11, the performance prediction model of this embodiment shows that the loss and mean absolute error become smaller as learning progresses, confirming that learning is progressing smoothly. Furthermore, from around 30 epochs, the loss and mean absolute error become sufficiently small, and there is little change thereafter, so it can be said that around 30 epochs is suitable for learning the performance prediction model in this embodiment.

[0050] Figure 12 shows the prediction results from a trained performance prediction model using 11 test data points. Here, the predicted values ​​(load displacement information) output from the performance prediction model and the training data (measured values: load displacement information obtained from pre-acquired FS diagrams) were plotted, and the results were observed. In Figure 12, the range in which the correlation coefficient R = 0.8 or higher with respect to the straight line of measured values ​​connecting the plotted data points marked with circles is shown by a dotted line. As a result, the approximation curve of the predicted values ​​from the performance prediction model (66 load displacement information points (= 11 (number of data points) × 6 (number of inflection points)) indicated by triangles) satisfies the correlation coefficient R = 0.8 or higher, indicating that good prediction results were obtained.

[0051] In this embodiment, a neural network such as CNN or RNN was used as an example of a machine learning algorithm to generate the performance prediction model described above. However, the machine learning algorithm used to generate the performance prediction model is not limited to this. For example, other machine learning algorithms such as random forests or boosted decision trees can also be used.

[0052] <Performance prediction processing> Next, the performance prediction process in the performance prediction system 1 of this embodiment will be described. Figure 13 is a functional block diagram showing the functions of the control unit 11 that performs the performance prediction process, and Figure 14 is a flowchart showing an example of the performance prediction process.

[0053] In Figure 13, the control unit 11 includes the aforementioned point cloud coordinate data generation unit 21, input data generation unit 26, performance prediction unit 27 which operates as a performance prediction model, and FS diagram creation unit 28 which creates an FS diagram from predicted values ​​(load displacement information). In this embodiment, as with the above, 3D design information of an automobile engine hood is used as an example of various metal panels, and the load displacement information of the performance prediction point (the position where the FS diagram is to be created: corresponding to the predetermined position described above) on the engine hood is predicted using the performance prediction model generated above, and then an FS diagram is created based on the obtained load displacement information. That is, in this embodiment, the load displacement information when the performance prediction point on the automobile engine hood is pressed is predicted using the learned performance prediction model, and an FS diagram of the performance prediction point is created from this load displacement information.

[0054] In this embodiment, the performance prediction point is, for example, any position within the region on the engine hood that has a yield point (a position where buckling occurs under a certain load). Specifically, the performance prediction point is any position in the region directly below which there are no seating surfaces (for example, parts such as dent R / F or lock R / F) that are joined by deflection (displacement).

[0055] In Figure 14, first, when the operator instructs the execution of the performance prediction process by operating the input unit 13, the point cloud coordinate data generation unit 21 of the control unit 11 reads the 3D design information of the engine hood from the storage unit 12 and places the engine hood (shape) in the 3D coordinate space (X coordinate, Y coordinate, Z coordinate) based on the 3D shape data obtained from the 3D design information (step S11). At this time, a 3D image of the engine hood is displayed on the display unit 15 (see Figure 5).

[0056] Next, when a performance prediction point on the engine hood is set by the operator using the input unit 13, the point cloud coordinate data generation unit 21 obtains the cross-sectional shape of the engine hood around the performance prediction point from the 3D design information (step S12). Specifically, it obtains the cross-sectional shape of the engine hood with a width of 150 mm in the front, rear, left, and right directions, centered on the performance prediction point. In other words, in this embodiment, the cross-sectional shape of the engine hood with a width of 150 mm in the left-right direction centered on the performance prediction point and the cross-sectional shape of the engine hood with a width of 150 mm in the front-rear direction centered on the performance prediction point are obtained (see Figures 6 and 7). The width dimension of the cross-sectional shape obtained here is the same as the width dimension of the cross-sectional shape obtained when the performance prediction model was generated (see step S2 in Figure 3).

[0057] Furthermore, the point cloud coordinate data generation unit 21 divides the cross-sectional shape (150 mm wide) in the front-to-back and left-to-right directions centered on the performance prediction point into, for example, 5 mm intervals (5 mm x 5 mm squares), and generates a point cloud of the cross-sectional shape at the performance prediction point (a collection of 5 mm square center points) (step S13). Then, based on the 3D shape data of the engine hood placed in 3D coordinate space, the point cloud coordinate data generation unit 21 acquires the coordinates (X coordinate, Y coordinate, Z coordinate) of each point forming the point cloud, generates point cloud coordinate data of the cross-sectional shape at the performance prediction point (step S14) (see Figures 7(c), (d) and 8), and stores it in the storage unit 12. The interval between each point constituting the point cloud of the cross-sectional shape at the performance prediction point is the same as the interval between each point constituting the point cloud of the cross-sectional shape acquired for each load point when generating the performance prediction model (see step S3 in Figure 3).

[0058] As described above, after the point cloud coordinate data generation unit 21 generates point cloud coordinate data at the performance prediction points, the input data generation unit 26 then reads the point cloud coordinate data and 3D design information (including information on the material and thickness of each component constituting the engine hood) from the storage unit 12, and links the coordinates of each point forming the point cloud coordinate data with the material and thickness of each point to generate input data to be input to the performance prediction model (step S15). At this time, the input data generation unit 26 further classifies (subdivides) the above input data into component units that constitute the engine hood. Here, the components that constitute the engine hood are, for example, the outer, inner, dent R / F, lock R / F, etc. The input data generation unit 26 then inputs the generated input data to the performance prediction unit 27, which operates as a performance prediction model.

[0059] The performance prediction unit 27 receives input data classified into component units that make up the engine hood, and uses a performance prediction model to predict load displacement information at the performance prediction points on the engine hood (load and displacement amount corresponding to each of the six inflection points mentioned above) (step S16). In other words, in this embodiment, a learned performance prediction model is used to predict load displacement information when the performance prediction points on the engine hood of the automobile are pressed.

[0060] Subsequently, the FS diagram creation unit 28 uses the predicted values ​​(load displacement information) from the performance prediction unit 27 to approximate the data using approximation functions such as first-order and second-order approximations, thereby creating an FS diagram at the performance prediction point (step S17). The created FS diagram is then displayed as the performance prediction result on the display unit 15.

[0061] <Effects> As described above, the performance prediction system 1 of this embodiment predicts the FS diagram (time-series load-displacement curve), which is one of the performance characteristics of metal panels (for example, the engine hood of an automobile) that make up the body of an automobile.

[0062] Specifically, the system includes a storage unit 12 that stores a dataset (see Figure 9) in which the explanatory variables are the point cloud coordinate data (see Figure 8) of a point cloud (see Figure 7) formed around load points uniformly provided in the region of the engine hood that has a yield point, and data relating the material and plate thickness information of each point constituting that point cloud, and further classified into component units that make up the engine hood. The dataset is then linked for each load point. The machine learning unit 23 uses the dataset read from the storage unit 12 to generate a performance prediction model using machine learning to predict the load displacement information when a performance prediction point is pressed. Subsequently, the performance prediction unit 27 receives input data that associates point cloud coordinate data (see Figure 8) of the point cloud (see Figure 7) formed around the performance prediction point, with information on the material and plate thickness of each point constituting the point cloud, and further classifies this data into component units that make up the engine hood. Using a trained performance prediction model, it predicts the load displacement information when the performance prediction point is pressed. Furthermore, the FS diagram creation unit 28 creates an FS diagram at the performance prediction point based on the load displacement information.

[0063] The memory unit 12 is pre-stored with 3D design information of the engine hood and information on the material and plate thickness of the engine hood associated with the 3D design information. The point cloud coordinate data generation unit 21 then places the engine hood in 3D coordinate space (X coordinate, Y coordinate, Z coordinate) based on the 3D design information of the engine hood. Subsequently, the point cloud coordinate data generation unit 21 acquires multiple cross-sectional shapes of the engine hood centered on load points or performance prediction points, divides the acquired cross-sectional shapes into squares of predetermined angles to generate a point cloud which is a collection of the center points, and acquires the coordinates (X coordinate, Y coordinate, Z coordinate) of each point forming the point cloud to generate point cloud coordinate data.

[0064] As a result, the dataset generation unit 22 can generate a dataset for generating a learning model based on the point cloud coordinate data generated by the point cloud coordinate data generation unit 21, the 3D design information of the engine hood, the material and thickness information of the engine hood associated with the 3D design information, and the FS diagrams of each load point that have been acquired in advance. In addition, the input data generation unit 26 can generate input data for performance prediction based on the point cloud coordinate data generated by the point cloud coordinate data generation unit 21, the 3D design information of the engine hood, and the material and thickness information of the engine hood associated with the 3D design information.

[0065] Furthermore, the performance prediction system 1 of this embodiment employs a convolutional neural network as the machine learning algorithm. In the machine learning process, the system has an intermediate layer that extracts features from each component that makes up the engine hood, and the depth of the intermediate layer is determined for each component according to the correlation between the predicted value (load displacement information) and each component.

[0066] According to the performance prediction system 1 of this embodiment, configured as described above, the time required to calculate the FS diagram when a performance prediction point on the engine hood is pressed can be significantly reduced by introducing a pre-trained performance prediction model. Furthermore, the distinctive configuration of the convolutional neural network in this embodiment enables advanced performance prediction that takes into account the correlation between each component constituting the engine hood.

[0067] In this embodiment, an automobile engine hood is used as an example of various metal panels, but it is not limited to this. The learning model generation process and performance prediction process of this embodiment can be applied to all metal panels that make up an automobile body, such as front fenders, roof panels, doors, etc., in addition to the engine hood. [Explanation of symbols]

[0068] 1. Performance prediction system 11 Control Unit 12 Storage section 13 Input section 14 Interface (I / F) section 15 Display 16 Communications Department 21 Point cloud coordinate data generation unit 22 Dataset Generation Unit 23 Machine Learning Department 24 CAD systems 25 CAE Systems 26 Input Data Generation Unit 27 Performance prediction unit 28 FS Diagram Creation Section 31 Input section 32 Feature Extraction Unit 33 Output section C-1,C-2,C-3,C-4,C-5,… Load point

Claims

1. In a performance prediction system that predicts the performance of metal panels that make up an automobile body, A storage means for storing a dataset in which the explanatory variables are data relating point cloud coordinate data of point clouds formed around load points uniformly provided in a region on the panel where buckling occurs under a constant load, and information on the material and thickness of each point constituting the point cloud, and further classified into component units constituting the panel, and the explanatory variables are linked to a target variable which is information on the F-S diagram when the load point is pressed, for each load point, A model generation means generates a performance prediction model using machine learning to predict information regarding the F-S diagram when a performance prediction point, which is an arbitrary position within the region, is pressed, using a dataset read from the storage means. A performance prediction means receives input data that associates point cloud coordinate data of a point cloud formed around a performance prediction point with information on the material and thickness of each point constituting the point cloud, and further classifies this data into component units that make up the panel, and uses a trained performance prediction model to predict information regarding the F-S diagram when the performance prediction point is pressed. An F-S diagram creation means creates an F-S diagram at a performance prediction point based on the information regarding the F-S diagram, which is the prediction result. Equipped with, The information regarding the F-S diagram is provided below. The load-displacement information consists of the load and displacement amounts corresponding to multiple inflection points, including the yield point, extracted from the F-S diagram of the load point, which was previously obtained using a known method. A performance prediction system characterized by the following features.

2. We have decided to adopt a convolutional neural network as the machine learning algorithm. The panel has an intermediate layer for extracting feature quantities for each component, and further determines the depth of the intermediate layer for each component according to the correlation between the predicted value and each component. The performance prediction system according to feature 1.

3. A performance prediction method for predicting the performance of metal panels that make up the body of an automobile, A dataset storage step involves storing in memory a dataset in which the explanatory variables are data obtained by associating point cloud coordinate data of point clouds formed around load points uniformly provided in a region on the panel where buckling occurs under a constant load, with information on the material and thickness of each point constituting the point cloud, and further classifying this data into component units constituting the panel, and linking these explanatory variables with a target variable, which is information on the F-S diagram when the load point is pressed, for each load point, and storing the dataset in memory. A model generation step involves generating a performance prediction model by machine learning that uses a dataset read from the memory to predict information about the F-S diagram when a performance prediction point, which is an arbitrary position within the region, is pressed. The performance prediction step involves receiving input data that associates point cloud coordinate data of a point cloud formed around a performance prediction point with information on the material and thickness of each point constituting the point cloud, and further classifying this data into component units that make up the panel, and using a trained performance prediction model to predict information regarding the F-S diagram when the performance prediction point is pressed, An F-S diagram creation step is performed to create an F-S diagram at a performance prediction point based on the information regarding the F-S diagram, which is the prediction result. A display step in which the F-S diagram created in the F-S diagram creation step is displayed on a display, Includes, The information regarding the F-S diagram is provided below. The load-displacement information consists of the load and displacement amounts corresponding to multiple inflection points, including the yield point, extracted from the F-S diagram of the load point, which was previously obtained using a known method. A performance prediction method characterized by the following.

4. The memory is pre-stored with 3D design information of the panel and information on the material and thickness of the panel associated with the 3D design information. Furthermore, as a preprocessing step for generating the performance prediction model, A shape placement step in which the panel is placed in a 3D coordinate space (X coordinate, Y coordinate, Z coordinate) based on the 3D design information of the panel, A step of acquiring a cross-sectional shape of the panel centered on the load point, A point cloud generation step is to generate the point cloud, which is a collection of center points, by dividing the obtained cross-sectional shape into squares of predetermined angles, A point cloud coordinate data generation step involves obtaining the coordinates (X coordinate, Y coordinate, Z coordinate) of each point forming the point cloud and generating the point cloud coordinate data, A dataset generation step that generates the dataset for model generation based on the point cloud coordinate data generated in the point cloud coordinate data generation step, the 3D design information of the panel, the material and thickness information of the panel associated with the 3D design information, and the load displacement information of each load point. including, The performance prediction method according to feature 3.

5. Furthermore, as a preprocessing step for the performance prediction step, The shape arrangement step, The cross-sectional shape acquisition step involves acquiring multiple cross-sectional shapes of the panel centered on the performance prediction point, The point cloud generation step, The point cloud coordinate data generation step, An input data generation step generates input data for performance prediction based on the point cloud coordinate data generated in the point cloud coordinate data generation step, the 3D design information of the panel, and the material and thickness information of the panel associated with the 3D design information. including, The performance prediction method according to feature 4.

Images