Strain region estimation device, strain region estimation method, machine learning device, and machine learning method

Abstract

To provide a technique for detecting visual distortion caused by optical illusion, which may occur in a three-dimensional surface model.SOLUTION: A distortion region estimation device (1) includes a control unit (a processor 12) that executes an estimation step of estimating a distortion region that can be included in a three-dimensional surface model and is visually distorted due to optical illusion.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a distortion region estimation device and a distortion region estimation method for estimating a distortion region in a three-dimensional surface model. The present invention also relates to a machine learning device and a machine learning method for constructing a learned model for estimating such a distortion region.

Background Art

[0002] In the design of vehicle parts (including interior and exterior parts) represented by passenger cars and work vehicles, three-dimensional CAD (Computer Aided Design) is widely used (see, for example, Patent Document 1). Each part is designed with an appearance design using three-dimensional CAD and an internal structure that cannot be visually observed from the outside. Three-dimensional CAD generates data representing the three-dimensional model of each designed part and visualizes the three-dimensional model. Also, in the manufacturing process of each part, each part can be manufactured based on the data of the three-dimensional model.

[0003] Hereinafter, among the three-dimensional models, the model representing the surface constituting the appearance is called a surface model.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] As described above, when a part is designed based on a three-dimensional model of a part designed using three-dimensional CAD, a part of the surface constituting the appearance may appear unnaturally distorted, different from the intention at the time of design. Such distortion is considered to be caused by the fact that a part of it appears visually distorted due to an optical illusion.

[0006] To eliminate such distortions in parts, visual inspection of the surface model in 3D CAD is essential. This is because distortions observed after the manufacturing of parts also occur in the surface model, albeit to varying degrees.

[0007] Visual inspection of surface models is time-consuming because it relies on manual labor. Furthermore, the degree of distortion that can be detected visually is highly dependent on the skill level of the inspector and therefore prone to variability. Thus, visual inspection of surface models often leads to increased costs and a decline in the aesthetic appearance of parts and, consequently, products.

[0008] One aspect of the present invention has been made in view of the above-mentioned problems, and its object is to provide a technique for detecting distortions that may occur in a three-dimensional surface model, which are visual distortions caused by optical illusions. [Means for solving the problem]

[0009] To solve the above problems, the strain region estimation device according to the first aspect of the present invention includes a control unit that performs an estimation step of estimating strain regions that may be included in a three-dimensional surface model and that appear visually distorted due to optical illusions.

[0010] To solve the above problems, the distortion region estimation method according to the ninth aspect of the present invention includes an estimation step of estimating distortion regions that may be included in a three-dimensional surface model and that appear visually distorted due to optical illusions.

[0011] To solve the above problems, a machine learning apparatus according to a thirteenth aspect of the present invention includes a control unit that performs a preprocessing step and a construction step that constructs a trained model that estimates distortion regions that may be included in lines included in a three-dimensional surface model and that appear visually distorted due to optical illusions by supervised learning using a training dataset, wherein the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving feature quantities that indicate the shape characteristics of each unit interval, the input to the trained model is the feature quantities derived for each unit interval, and the output to the trained model is an estimation result of whether or not each unit interval includes the distortion region.

[0012] To solve the above problems, a machine learning method according to a 14th aspect of the present invention includes a preprocessing step and a construction step of constructing a trained model that estimates distortion regions that may be included in lines included in a 3D surface model and that appear visually distorted due to optical illusions, by supervised learning using a training dataset, wherein the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving feature quantities that indicate the shape characteristics of each unit interval, the input to the trained model is the feature quantities derived for each unit interval, and the output to the trained model is an estimation result of whether or not each unit interval includes the distortion region.

[0013] Each aspect of the present invention may be implemented by a computer. In this case, the strain region estimation program for the strain region estimation device, which enables the computer to implement the strain region estimation device by operating the computer as each part (software element) of the strain region estimation device, and a computer-readable recording medium on which the program is recorded, also fall within the scope of the present invention.

[0014] Furthermore, each aspect of the present invention may be implemented by a computer. In this case, a strained machine learning program for the machine learning device, which enables the implementation of the machine learning device by operating the computer as each part (software element) of the machine learning device, and a computer-readable recording medium on which the program is recorded, also fall within the scope of the present invention. [Effects of the Invention]

[0015] According to one aspect of the present invention, a technique can be provided for detecting distortions that may occur in a three-dimensional surface model, which are visual distortions caused by optical illusions. [Brief explanation of the drawing]

[0016] [Figure 1] This is a block diagram showing the configuration of a strain region estimation system according to one embodiment of the present invention. [Figure 2] This flowchart shows the flow of the strain region estimation method performed by the strain region estimation device included in the strain region estimation system shown in Figure 1. [Figure 3] Figure 2 is a flowchart showing the flow of preprocessing steps included in the strain region estimation method. [Figure 4] Figure 1 shows a perspective view of the surface model of a component whose presence or absence of strain is estimated by the strain region estimation device. [Figure 5] Figure 4 is a magnified perspective view of a portion of the surface model shown. [Figure 6] Figure 4 shows an example of a cross-sectional profile of the surface model. [Figure 7] Figure 4 is an enlarged perspective view of a portion of the surface model, illustrating the points determined during the feature derivation step included in the strain region estimation method shown in Figure 2. [Figure 8] This is a schematic diagram illustrating an example of a feature derived in the feature derivation step included in the distortion region estimation method shown in Figure 2. [Figure 9]Figure 1 is a flowchart showing the flow of the machine learning method executed by the machine learning device included in the strain region estimation system. [Figure 10] The top panel is a scatter plot of PCA features obtained in one variation of the estimation step, plotted on a two-dimensional plane. The bottom panel is a scatter plot that is an enlarged portion of the scatter plot in the top panel. [Modes for carrying out the invention]

[0017] [Strain Region Estimation System] A strain region estimation system S, including a strain region estimation device 1 according to one embodiment of the present invention, will be described with reference to Figure 1. Figure 1 is a diagram showing the configuration of the strain region estimation system S.

[0018] The distortion region estimation system S is a system for estimating the presence or absence of visual distortions that can occur in a three-dimensional surface model, such as those caused by optical illusions. As shown in Figure 1, the distortion region estimation system S comprises a distortion region estimation device 1, a 3D CAD (Computer Aided Design) system 2, and a machine learning device 3 (see Figure 9).

[0019] The strain region estimation system S, including the strain region estimation device 1, aims to detect visual distortions caused by optical illusions, which can occur in the three-dimensional surface model of vehicle parts (including interior and exterior parts) during the design phase of vehicle parts, such as passenger cars and work vehicles. Examples of vehicles include construction machinery and agricultural machinery in addition to passenger cars and work vehicles. Furthermore, the vehicle may be operated by a user or operated by autonomous driving. In this embodiment, the strain region estimation system S and strain region estimation device 1 will be described using an exterior part of a passenger car as an example of a part.

[0020] The parts are designed using 3D CAD2. Specifically, 3D CAD2 generates a three-dimensional surface model of the part and stores the data representing this surface model in the storage provided by 3D CAD2, or in storage located outside of 3D CAD2. In this embodiment, it is assumed that the data representing the surface model is stored in the storage of 3D CAD2. Note that a surface model is one aspect of a three-dimensional model, and is a model that represents the appearance (design) of the part using surfaces whose thickness is not defined.

[0021] The distortion region estimation device 1 includes a control unit (processor 12, described later) that performs an estimation step to estimate distortion regions that may be included in a three-dimensional surface model and that appear visually distorted due to optical illusions. This estimation step may be configured to estimate distortion regions that may be included in lines included in the surface model and that appear visually distorted due to optical illusions using a trained model constructed by supervised learning, or it may be configured to estimate distortion regions using a trained model constructed by unsupervised learning, or it may be configured to estimate distortion regions using a predetermined algorithm (for example, a principal component analysis (PCA) algorithm) executed according to a program.

[0022] <Overview of the strain region estimation device> The strain region estimation device 1 of this embodiment is a device for executing the strain region estimation method M1. The strain region estimation method M1 is a method for estimating strain regions that may be included in lines included in a surface model, using a trained model LM constructed by machine learning based on data representing a surface model provided from a 3D CAD2. Here, a strain region means a region that appears visually distorted due to an optical illusion. In this embodiment, the cross-sectional profile of the surface model is used as the lines included in the surface model. However, the lines included in the surface model are not limited to the cross-sectional profile, and may, for example, be constituent lines on the faces that constitute the surface model. These constituent lines can also be described as boundary lines or edges on the faces that constitute the surface model.

[0023] The algorithm of the trained model LM, the configuration of the strain region estimation device 1, and the flow of the strain region estimation method M1 will be described in detail later.

[0024] The input to the trained model LM consists of feature quantities derived for each of the multiple unit intervals defined in the cross-sectional profile of the surface model. These feature quantities can be appropriately selected to represent the shape characteristics of each unit interval.

[0025] The output of the trained model LM is an estimation result of whether or not each unit interval of the cross-sectional profile described above contains a strain region. For example, if it is determined that a strain region is contained in a certain unit interval, the trained model LM outputs "1", and if it is determined that a strain region is not contained, the trained model LM outputs "0".

[0026] The inventors of this application have found that there is a certain relationship between the feature quantities that are input to the trained model LM and the estimation result of whether or not a strained region is included, which is the output of the trained model LM, although it is difficult to explicitly specify this relationship as a relational expression. Therefore, by using the trained model LM that takes feature quantities as input, it is possible to accurately estimate whether or not each unit interval of the cross-sectional profile includes a strained region.

[0027] <Overview of the Machine Learning System> Machine learning device 3 is a device for executing machine learning method M3. Machine learning method M3 is a method for creating a training dataset DS using data representing a surface model provided from 3D CAD2, and for constructing a trained model LM by machine learning using the training data dataset DS. Details of the configuration of machine learning device 3 and the flow of machine learning method M3 will be described later.

[0028] The strain region estimation system S progresses through a preparation phase and a trial phase before reaching the practical application phase. A brief explanation of the preparation phase, trial phase, and practical application phase is as follows.

[0029] (1) Preparation Phase In the preparation phase, the operator determines whether or not the surface model contains strained regions. Each time the operator determines whether or not the surface model contains strained regions, the machine learning device 3 creates training data (training data) from the data representing the surface model provided by 3D CAD2 and adds the created training data to the training dataset DS. The preparation phase may end after a predetermined period of time (e.g., one week, one month, or one year) has elapsed from the start of the preparation phase, or when the number of strained region estimations in the preparation phase reaches a predetermined number (e.g., 100 times, 1000 times, or 10000 times). Once the preparation phase is complete, the machine learning device 3 constructs a trained model LM using machine learning with the training data DS. The constructed trained model LM is then transferred from the machine learning device 3 to the strained region estimation device 1.

[0030] (2) Trial Phase In the trial phase, the operator determines whether or not a strained region is included in the surface model, and the strained region estimation device 1 estimates whether or not a strained region is included in the lines included in the surface model. Each time the operator measures whether or not a strained region is included in the surface model, the strained region estimation device 1 estimates whether or not a strained region is included in the surface model using the trained model LM based on the data representing the surface model provided from 3D CAD2. The trial phase may end after a predetermined period of time (e.g., one week, one month, or one year) has elapsed from the start of the trial phase, or it may end when the number of times the surface model is estimated to include a strained region in the trial phase reaches a predetermined number (e.g., 100 times, 1000 times, or 10000 times). When the trial phase ends, the operator compares their own determination results with the estimation results of whether or not a strained region is included estimated by the strained region estimation device 1 and evaluates the estimation accuracy of the strained region estimation device 1. If the estimation accuracy is insufficient, the operator returns to the preparation phase. If the estimation accuracy is sufficient, the operator proceeds to the practical application phase. Furthermore, the accuracy of the estimation can be verified using a portion of the training dataset DS after the preparation phase is complete. In this case, the trial phase can be omitted.

[0031] (3) Practical application phase In the practical application phase, the strain region estimation device 1 estimates whether or not the surface model contains strain regions. The trained model LM used by the strain region estimation device 1 in the practical application phase has been confirmed to have sufficient estimation accuracy in the trial phase. In the practical application phase, the operator's determination of whether or not the surface model contains strain regions can be omitted. This frees the operator from the trouble of determining whether or not the surface model contains strain regions, and makes it possible to estimate whether or not the surface model contains strain regions with high accuracy.

[0032] <Configuration of the strain region estimation device> The configuration of the strain region estimation device 1 will be explained with reference to Figure 1. The upper part of Figure 1 is a block diagram showing the configuration of the strain region estimation device 1.

[0033] As shown in Figure 1, the strain region estimation device 1 comprises a memory 11, a processor 12, and a storage device 13. The memory 11, processor 12, and storage device 13 are connected to each other via a bus (not shown). An input / output interface (not shown) and a communication interface (not shown) may also be connected to this bus. This input / output interface is used, for example, to input data representing a surface model from a 3D CAD2 to the strain region estimation device 1, or to output judgment results from the strain region estimation device 1 to an external device (e.g., 3D CAD2). This communication interface is used, for example, to obtain a trained model LM from an external device (e.g., a machine learning device 3).

[0034] Memory 11 is configured to store the trained model LM constructed by machine learning. The trained model LM is an algorithm that takes the feature quantities of each unit interval of lines included in the surface model (in this embodiment, the cross-sectional profile of the surface model), which are derived by the preprocessing step described later, as input, and outputs the estimation result of whether or not each unit interval contains a strain region. For memory 11, for example, semiconductor RAM (Random Access Memory) can be used. Furthermore, gradient boosting can be suitably used as the machine learning algorithm for the trained model LM. In addition, algorithms such as Bayesian-Gaussian mixture models and support vector machines can be used as the machine learning algorithm for the trained model LM. Hereafter, the machine learning algorithm will also be simply referred to as the algorithm.

[0035] The processor 12 is configured to execute the strain region estimation method M1, described later, using the trained model LM stored in the memory 11. The processor 12 can be, for example, a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), a microprocessor, a digital signal processor, a microcontroller, a TPU (Tensor Processing Unit), or a combination thereof. The processor 12 functions as a control unit in one aspect of the present invention and is sometimes referred to as an "arithmetic unit."

[0036] Storage 13 is configured to store (non-volatile) the trained model LM. When the processor 12 estimates whether or not each unit interval contains a strained region, it loads the trained model LM stored in storage 13 onto memory 11 and uses it. For storage 13, for example, flash memory, HDD (Hard Disk Drive), SSD (Solid State Drive), or a combination thereof can be used.

[0037] In this embodiment, a configuration is adopted in which the strain region estimation method M1 is executed using a single processor (processor 12), but the present invention is not limited thereto. That is, a configuration may be adopted in which the strain region estimation method M1 is executed using multiple processors. For example, the preprocessing step S11 (described later) may be executed by the first processor, and the estimation step S12 (described later) may be executed by the second processor. In this case, the multiple processors that work together to execute the strain region estimation method M1 may be provided in a single computer and configured to communicate with each other via a bus, or they may be distributed across multiple computers and configured to communicate with each other via a network. As an example, a configuration in which a processor built into a computer constituting a cloud server and a processor built into a computer owned by a user of that cloud server work together to execute the strain region estimation method M1 can be considered.

[0038] Furthermore, in this embodiment, the trained model LM is stored in memory 11 built into the same computer as the processor (processor 12) that executes the strain region estimation method M1, but the present invention is not limited to this. That is, the trained model LM may be stored in memory built into a different computer from the one that executes the strain region estimation method M1. In this case, the computer containing the memory that stores the trained model LM is configured to communicate with the computer containing the processor that executes the strain region estimation method M1 via a network. As an example, the trained model LM may be stored in memory built into a computer that constitutes a cloud server, and the strain region estimation method M1 may be executed by a processor built into a computer owned by a user of that cloud server.

[0039] Furthermore, although this embodiment employs a configuration in which the trained model LM is stored in a single memory 11, the present invention is not limited thereto. That is, a configuration in which the trained model LM is distributed and stored in multiple memories may also be adopted. In this case, the multiple memories in which the trained model LM is stored may be provided in a single computer (which may or may not be a computer with a built-in processor for executing the strain region estimation method M1), or they may be distributed and provided in multiple computers (which may or may not include a computer with a built-in processor for executing the strain region estimation method M1). As an example, a configuration in which the trained model LM is distributed and stored in the memory built into each of the multiple computers constituting a cloud server can be considered.

[0040] <Flowchart of the strain region estimation method> The flow of the strain region estimation method M1 will be explained with reference to Figures 2 to 8. Figure 2 is a flowchart showing the flow of the strain region estimation method M1. Figure 3 is a flowchart showing the flow of the preprocessing step S11 included in the strain region estimation method M1. Figure 4 is a perspective view of the surface model SM of a part whose presence or absence of strain is estimated by the strain region estimation device 1. Figure 5 is an enlarged perspective view of a part of the surface model SM. Figure 6 is an example of the cross-sectional profile of the surface model SM. Figure 7 is an enlarged perspective view of a part of the surface model SM, illustrating points determined in the process of the feature quantity derivation step S112 included in the strain region estimation method M1. Figure 8 is a schematic diagram to explain an example of a feature quantity derived in the feature quantity derivation step S112. As shown in Figures 4, 5, and 7, in this embodiment, a Cartesian coordinate system defined by the x, y, and z axes is used in the space where the surface model SM is handled. However, the coordinate system used in the strain region estimation method M1 and the machine learning method M3 described later is not limited to a Cartesian coordinate system. The coordinate system used in the machine learning method M3 may be, for example, an oblique coordinate system, a polar coordinate system, or a general coordinate system. An example of a general coordinate system is a cylindrical coordinate system.

[0041] The strain region estimation method M1 includes a preprocessing step S11 and an estimation step S12.

[0042] (Pre-processing step) The preprocessing step S11 is a step in which the processor 12 derives feature quantities for each unit interval of the cross-sectional profile of the surface model to be input to the trained model LM. As shown in Figure 3, the preprocessing step S11 includes a cross-sectional profile generation step S111 and a feature quantity derivation step S112.

[0043] Step S111 is a step in which multiple cross-sectional profiles are generated from the surface profile of the surface model SM. The multiple cross-sectional profiles generated in the cross-sectional profile generation step S111 are preferably multiple cross-sections represented in an oblique coordinate system composed of three intersecting axes, and are cross-sectional profiles of multiple cross-sections parallel to a plane defined by two of the three axes. Furthermore, it is more preferable that the spacing between adjacent cross-sectional profiles is equal.

[0044] As described above, this embodiment uses a Cartesian coordinate system with three mutually orthogonal x, y, and z axes. Furthermore, as multiple cross-sectional profiles, (1) multiple cross-sectional sections C that are parallel to the xy plane and equally spaced. xy (2) Multiple cross-sections C that are parallel to the yz plane and are equally spaced. yz , and (3) multiple cross-sections C parallel to the zx plane and at equal intervals zx The cross-sectional profile in is used (see Figures 4 and 5). Also, the adjacent cross-section C yz interval d yz The value is set to 20 mm (see Figure 5). Adjacent section C xy The distance between them, and the adjacent cross section C zx The spacing between them, and the cross-section C yzSimilarly, it is set to 20 mm. However, these intervals can be determined as appropriate according to the size of the assumed strain region and the like.

[0045] Note that, from the shape of the component, cross-section C xy , C yz , C zx If it is known which cross-section among them is the cross-section where it is easy to estimate the strain region, for example, any one of the cross-sections C xy , C yz , C zx can be selected, and only the cross-section profile in that cross-section can be used. However, by using all the cross-section profiles of C xy , C yz , C zx , it becomes unnecessary to worry about the direction of the coordinate axes in each component, and it is possible to reduce the oversight of the strain region.

[0046] The feature quantity derivation step S112 is a step of setting a plurality of unit intervals in the line included in the surface model SM and deriving a feature quantity indicating the feature of the shape of each unit interval. In the present embodiment, as the line included in the surface model SM, the cross-section profiles in the cross-sections C xy , C yz , C zx shown in FIGS. 4 and 5 are adopted. FIG. 6 shows an example of the cross-section profile in the cross-section C yz . Taking the cross-section profile shown in FIG. 6 as an example, the feature quantity derivation step S112 sets a plurality of unit intervals Se yz (i is an integer of 3 ≤ i ≤ N, where N is a positive integer) in the cross-section profile in the cross-section C i and derives the feature quantity of each unit interval Se i . More specifically, the cross-sectional profile of the surface model SM is constructed by connecting N curves in series, each composed of a single spline curve. The feature derivation step S112 derives the feature quantities for each unit interval, using an interval consisting of a first interval consisting of the i-th spline curve (where i is an integer between 2 and N-1), a second interval consisting of the (i-1)th spline curve, and a third interval consisting of the (i+1)th spline curve as the unit interval. For example, the unit interval Se shown in Figures 6 and 7 i+1 If we consider this as the first interval, then the second interval is the unit interval Se i The third interval is the unit interval Se i+2 That is the case. In this embodiment, the first interval is considered the center, and the second and third intervals located before and after it are considered the ends, and 20 feature quantities, described later, are derived. Hereafter, the 20 feature quantities derived using the three intervals, with the first interval as the center and the second and third intervals located before and after it as the ends, will be referred to as the feature quantities of the first interval.

[0047] Next, the feature derivation step S112 involves the first interval (e.g., the unit interval Se i+1 ), the second interval (for example, the unit interval Se i ), and the third interval (for example, the unit interval Se i+2 The M pseudo-line segments obtained by dividing each of the ) into M equal parts (where M is an integer greater than or equal to 2) are designated as line segment L1j, line segment L2j, and line segment L3j (where j is an integer greater than or equal to 1 and less than or equal to M). In the cross-sectional profile shown in Figure 6, each unit interval Se i+1 Point P is the starting point and ending point of each. i and point P i+1 That is what it states. Furthermore, line segment P i P i+1 The five points obtained by dividing the area into six equal parts are point P. ij This is the case (see Figure 7). Therefore, the first interval is the unit interval Se i+1 Line segment P i P ij This corresponds to the line segment L1j. Similarly, the second interval, the unit interval Se i Line segment P i Pij This corresponds to the line segment L2j, and the third interval is the unit interval Se i+2 Line segment P i P ij This corresponds to line segment L3j. In the cross-sectional profile shown in Figure 7, M=6 is used as described above. However, M is not limited to M=6 and can be appropriately selected depending on the size and design of the surface model SM. A typical example of M is M=50, but M can also be around 10 or around 100.

[0048] In this embodiment, the cross section C is parallel to the yz plane. yz In the cross-sectional profile, the unit interval Se is generalized using i. i+1 We will explain how to derive 20 features using the case where is the first interval as an example. However, in the actual feature derivation step S112, each unit interval Se in the cross-sectional profile is i For the case where the first interval is (i is an integer between 2 and N-1), we repeatedly derive 20 features. Also, section C yz Similar to the case of the cross-sectional profile in section C, xy ,C zx For each unit interval of the cross-sectional profile in each of these, 20 features are repeatedly derived.

[0049] The feature derivation step S112 derives, as features for each unit interval, at least the maximum value of the change in angle between adjacent line segments L1j and L1j+1 in the first interval, the maximum value of the change in angle between adjacent line segments L2j and L2j+1 in the second interval, and the maximum value of the change in angle between adjacent line segments L3j and L3j+1 in the third interval. Then, in the feature derivation step S112, the maximum value of the change in angle between adjacent line segments L1j and L1j+1 in the first interval is taken as the first feature. Also in the feature derivation step S112, the maximum value of the change in angle between adjacent line segments L2j and L2j+1 in the second interval is compared with the maximum value of the change in angle between adjacent line segments L3j and L3j+1 in the third interval, and the smaller value is taken as the second feature, and the larger value is taken as the third feature.

[0050] Here, the second interval is the unit interval Se i Line segment Lij(P) i P ij ) and line segment Lij+1(line segment P i P ij+1 ) angle θ ij While acute and obtuse angles can be considered, in this embodiment, the acute angle is defined as angle θ. ij This will be adopted. In Figure 8, angle θ i1 ~angle θ i4 This is illustrated. Angle θ ij The change in the unit interval Se is what is meant by the change in Se i The length L obtained by dividing the length into 6 equal parts i angle θ ij It can be obtained by dividing by .

[0051] Furthermore, a fourth feature is defined as the relationship between the maximum value of the angle change in the first interval, the maximum value of the angle change in the second interval, and the maximum value of the angle change in the third interval, expressed as one of the values ​​0, -1, or 1. If the absolute value of the maximum value of the angle change in the first interval is small, and the absolute values ​​of the maximum values ​​of the angle change in the second and third intervals are large, then 0 is adopted. In addition, if the case described above does not fall under the category of adopting 0, and the maximum value of the angle change is positive, then -1 is adopted. Examples of this case include: (1) when the maximum values ​​of the angle change in the first to third intervals are all positive, and the maximum values ​​of the angle change decrease in the order of second, first, and third intervals; (2) when the maximum values ​​of the angle change in the second and third intervals are approximately the same, and the maximum value of the angle change in the first interval is greater than the maximum values ​​of the angle change in the second and third intervals; and (3) when the maximum value of the angle change in the second interval is greater than the maximum values ​​of the angle change in the first and third intervals, and the maximum values ​​of the angle change in the first and third intervals are approximately the same. The range within which the maximum values ​​of the angle change are considered to be approximately the same can be determined as appropriate. Furthermore, when determining whether the maximum values ​​of the angle change are approximately the same, the relative magnitude of the maximum values ​​of the angle change being compared is not considered. In addition, if the minimum value of the angle change is negative, and the above-mentioned cases of adopting 0 are not applicable, then 1 is adopted. Examples of this case include (4) when the minimum values ​​of the angle change in the first to third intervals are all negative, and the minimum values ​​of the angle change decrease in the order of second, first, and third intervals; (5) when the minimum values ​​of the angle change in the second and third intervals are approximately the same, and the minimum value of the angle change in the first interval is smaller than the minimum values ​​of the angle change in the second and third intervals; and (6) when the minimum value of the angle change in the second interval is smaller than the minimum values ​​of the angle change in the first and third intervals, and the minimum values ​​of the angle change in the first and third intervals are approximately the same. Furthermore, additional criteria may be established for each of the cases in which 0 is adopted, -1 is adopted, and 1 is adopted.

[0052] Furthermore, in the feature derivation step S112, the angle in the first interval (D1) is set as the fifth feature, and the smaller (Ds) and larger (Db) angles of the second and third intervals are set as the sixth and seventh features, respectively.

[0053] Furthermore, the difference of D1 with respect to Ds is designated as the eighth feature, and the difference of D1 with respect to Db is designated as the ninth feature.

[0054] Furthermore, the absolute value of the difference between Db and Ds is defined as the tenth feature.

[0055] Furthermore, the sum of the angles in the second interval and the sum of the angles in the third interval are used as the eleventh feature.

[0056] Furthermore, the sum of the angles in the first interval, the sum of the angles in the second interval, and the sum of the angles in the third interval is defined as the twelfth feature.

[0057] Furthermore, the 13th feature is a representation of the relative magnitudes of the maximum angle in the first interval, the maximum angle in the second interval, and the maximum angle in the third interval, expressed as one of the values ​​0, -1, or 1. If the absolute value of the maximum angle in the first interval is small, and the absolute values ​​of the maximum angles in the second and third intervals are large, then 0 is adopted. In addition, if the case does not fall under the above-mentioned cases of adopting 0, and the maximum angle is positive, then -1 is adopted. An example of this case is when the maximum angles in the first to third intervals are all positive, and the maximum angles decrease in the order of second, first, and third intervals. In addition, if the case does not fall under the above-mentioned cases of adopting 0, and the minimum angle is negative, then 1 is adopted. An example of this case is when the minimum angles in the first to third intervals are all negative, and the minimum angles decrease in the order of second, first, and third intervals. Furthermore, as with the fourth feature, additional criteria may be established for each of the following cases: adopting 0, adopting -1, and adopting 1.

[0058] Also, the length L in the first section i Let length L1 be the 14th feature.

[0059] Furthermore, the length L in the second section i (Hereinafter referred to as length L2) and length L in the third section i (Hereinafter referred to as length L3) is compared with the smaller value (Ls) and the larger value (Lb) are designated as the 15th and 16th features, respectively.

[0060] Furthermore, the ratio of Ls to L1 (ln(Ls / L1)) is defined as the 17th feature.

[0061] Furthermore, the ratio of Lb to L1 (ln(Lb / L1)) is defined as the 18th feature.

[0062] Furthermore, the ratio of Lb to Ls (ln(Lb / Ls)) is defined as the 19th feature.

[0063] Furthermore, the 20th feature is a representation of the relative sizes of L1, L2, and L3 using one of the values ​​0, -1, or 1. 0 is used when L1 is small and L2 and L3 are large, and when L1 is large and L2 and L3 are small. -1 is used when L2, L1, and L3 decrease monotonically in that order. 1 is used when L2, L1, and L3 increase monotonically in that order. Additional criteria may be established for each of the cases where 0, -1, and 1 are used.

[0064] As described above, in this embodiment, the feature derivation step S112 derives 20 features. However, the features derived by the feature derivation step S112 are not limited to the above 20. Some of the above features can be omitted, or other features can be added to the above features. The features derived by the feature derivation step S112 can be appropriately determined according to the shape of the part and the corresponding distortions that may occur.

[0065] (Estimated step) Estimation step S12 is a step in which the processor 12 uses the trained model LM to estimate whether there are any distortion regions that may be included in the lines (cross-sectional profiles in this embodiment) included in the surface model SM and that appear visually distorted due to optical illusions. In estimation step S12, the processor 12 reads data representing the feature quantities derived in preprocessing step S11 from the memory 11 and inputs the read data representing the feature quantities into the trained model LM. Then, it writes the estimation result output from the trained model LM, that is, data representing the estimation result of whether or not each unit interval contains the distortion region, to the storage 13. The form of the data representing the estimation result can be determined as appropriate, but in this embodiment, it is set to "1" when the unit interval contains a distortion region and to "0" when the unit interval does not contain a distortion region.

[0066] In the actual estimation step S12, section C xy ,C yz ,C zx In all cross-sectional profiles, each unit interval Se i For the case where the first interval is an integer between 2 and N-1 (where i is between 2 and N-1), 20 feature quantities are taken as input, and the estimation result of whether or not each unit interval contains the aforementioned distortion region is output.

[0067] The strain region estimation method M1 may further include an output step that outputs data representing the estimation results estimated in the estimation step S12. In this output step, the processor 12 may be configured to read data representing the estimation results from the storage 13 and provide the read data representing the estimation results to the 3D CAD2. This allows the 3D CAD2 to differentiate the strain regions included in each unit section constituting each cross-sectional profile of the surface model SM from other regions. In this output step, the processor 12 may also be configured to present the estimation results read from the storage 13 to the user by outputting them to a display. This allows the user to know whether or not a strain region is included in each unit section constituting each cross-sectional profile of the surface model SM.

[0068] <Machine learning equipment and machine learning methods> The configuration of the machine learning device 3 will be explained with reference to Figures 1 and 9. The lower part of Figure 1 is a block diagram showing the configuration of the machine learning device 3. Figure 9 is a flowchart showing the flow of the machine learning method M3 executed by the machine learning device 3.

[0069] The machine learning device 3 is a device that executes a machine learning method M3 for constructing a trained model LM. As shown in Figure 5, the machine learning device 3 comprises a storage 31, a processor 32, and a memory 33. The storage 31, processor 32, and memory 33 are connected to each other via a bus (not shown). An input / output interface (not shown) and a communication interface (not shown) may also be connected to this bus. This input / output interface is used, for example, to input training data from an external device (e.g., a sensor and a keyboard) to the machine learning device 3. This communication interface is used, for example, to provide the trained model LM to an external device (e.g., the strain region estimation device 1 mentioned above).

[0070] Storage 31 is configured to store the training dataset DS. The training dataset DS is a set of training data in which data representing multiple features in each unit interval that constitutes each cross-sectional profile of the surface model SM are labeled to indicate whether or not the unit interval contains a strain region.

[0071] Here, assigning a label representing the estimation result to data representing multiple features means associating the data representing multiple features with the estimation result in any way. Methods for associating the data representing multiple features with the estimation result include, for example, creating a table that establishes a one-to-one correspondence between the data representing multiple features and the estimation result, or storing the data representing multiple features in a directory corresponding to the estimation result. The storage 31 can be, for example, flash memory, HDD, SSD, or a combination thereof.

[0072] As shown in Figure 9, the machine learning method M3 includes a preprocessing step S31 and a construction step S32. The preprocessing step S31 is the same as the preprocessing step S11 shown in Figure 2. That is, the preprocessing step S31 is a step in which the processor 32 derives multiple features in each unit interval of the cross-sectional profile of the surface model SM to be input to the trained model LM. In the preprocessing step S31, the processor 32 uses the set of data representing the multiple features in each unit interval as the training dataset DS.

[0073] Each training dataset DS contains data representing multiple features for each unit interval. The data representing multiple features for each unit interval included in each training dataset DS is similar to the data representing multiple features for each unit interval input to the trained model LM, and represents multiple features derived for each of the multiple unit intervals set in the cross-sectional profile of the surface model.

[0074] In the preprocessing step S31, labels representing the estimation results described above are further assigned to the multiple features derived for each unit interval.

[0075] In the construction step S32, the processor 32 is configured to perform a construction process in which it loads the training dataset DS stored in the storage 31 onto the memory 33 and constructs a trained model LM using supervised machine learning with this training dataset DS. As described above, the trained model LM is an algorithm that takes multiple features derived for each of the multiple unit intervals set in the cross-sectional profile of the surface model as input and outputs the estimation result of whether or not each unit interval of the cross-sectional profile contains a strain region. As the processor 32, for example, an ASIC such as a CPU, GPU, microprocessor, digital signal processor, microcontroller, TPU, or a combination thereof can be used. The processor 32 functions as a control unit in one aspect of the present invention and is sometimes called an "arithmetic unit". Memory 33 is configured for storing the trained model LM obtained by the processor 32 performing the construction process. For example, semiconductor RAM can be used as memory 33. The trained model LM stored in memory 33 may be stored (non-volatile) in the storage 31 described above.

[0076] In this description, we have explained a configuration in which a single processor 32 on a single computer executes the machine learning method M3, which includes the process of constructing a trained model LM. However, this is not the only configuration. That is, it is also possible to employ a configuration in which this machine learning method M3 is executed jointly by multiple processors on a single computer or distributed across multiple computers.

[0077] Furthermore, while this description has focused on a configuration in which the training dataset DS is stored in a single storage device 31 located on a single computer, the configuration is not limited to this. That is, it is also possible to adopt a configuration in which the training dataset DS is stored in multiple storage devices located on a single computer or distributed across multiple computers. Moreover, the training dataset DS does not necessarily need to be stored in the storage device 31 built into the computer along with the processor 32 and memory 33; it may also be stored in a cloud server configured to communicate with that computer via a network.

[0078] Furthermore, while this description has focused on a configuration in which the trained model LM is stored in a single memory 33 located on a single computer, it is not limited to this configuration. In other words, it is also possible to employ a configuration in which the trained model LM is stored in multiple memories located on a single computer, or distributed across multiple computers.

[0079] The program that causes the processor 32 to execute the machine learning method M3 is recorded, for example, on a computer-readable, non-temporary, tangible recording medium. The processor 32 executes the determination method S1 by executing the instructions contained in this program. This recording medium may be storage 31, memory 33, or other recording medium. For example, tape, disk, card, semiconductor memory, and programmable logic circuits can be used as other recording media.

[0080] The machine learning device 3 can construct the trained model LM used by the strain region estimation device 1 described above. Moreover, the training data used for machine learning consists of data representing multiple features in each unit interval that constitutes each cross-sectional profile of the surface model SM, with labels indicating whether or not the unit interval contains a strain region. Conventionally, the determination of whether or not the surface model SM contains a strain region was made by displaying the surface model SM on a screen and having an operator visually inspect the surface model SM. Therefore, when creating training data, an operator can easily determine the presence or absence of a strain region. Alternatively, the results of visual inspections that have already been accumulated can be used as training data. Thus, the machine learning device 3 makes it easy to create highly accurate training data. The same can be said for the machine learning method M3 implemented using the machine learning device 3, and the program that operates the computer as the machine learning device 3.

[0081] In this embodiment, the machine learning method M3 for constructing a trained model LM and the estimation method for estimating the presence or absence of strained regions using the trained model LM are described in two separate devices. However, the present invention is not limited to this. That is, the present invention also includes an embodiment in which the machine learning method M3 for constructing a trained model LM and the estimation method for estimating the presence or absence of strained regions using the trained model LM are performed in a single device.

[0082] [Variations of the estimation step] In the estimation step S12 of the distortion region estimation method M1 shown in Figure 2, the processor 12 was configured to estimate, using the trained model LM, whether or not there are distortion regions that may be included in the lines (cross-sectional profile in this embodiment) included in the surface model SM and that appear visually distorted due to optical illusions.

[0083] However, a modified version of estimation step S12 may be configured to estimate strain regions that may be included in the surface model SM without using the trained model LM. In this case, the processor 12 performs principal component analysis (PCA) on the multiple features (20 features in this embodiment) derived by the preprocessing step S11 and converts them into two-dimensional features (hereinafter referred to as PCA features).

[0084] The upper part of Figure 10 is a scatter plot of PCA features plotted on a two-dimensional plane. The lower part of Figure 10 is a scatter plot that is an enlarged portion of the upper part of Figure 10.

[0085] In the modified estimation step S12, a region R (see lower part of Figure 10) is predetermined on the two-dimensional plane on which the PCA features are plotted. This region is estimated by the processor 12 to contain strained areas for each unit interval of the cross-sectional profile of the surface model SM. The modified estimation step S12 utilizes the tendency for features of unit intervals containing strained areas to cluster in specific regions on the two-dimensional plane when converted to PCA features. By constructing estimation step S12 using PCA features in this way, it is possible to estimate strained areas that may be included in the surface model SM without using the trained model LM.

[0086] In this modified example, PCA is performed on multiple feature quantities derived in the preprocessing step S11. However, the multiple feature quantities used in this modified example can be changed as appropriate. Also, the region R that defines the feature quantities of the unit interval including the distortion region can be changed as appropriate. [Examples]

[0087] Using the machine learning device 3 and strain region estimation device 1 included in the strain region estimation system S shown in Figure 1, the strain regions that may be included in the cross-sectional profile of the surface model SM were estimated. In this example, the Bayesian-Gaussian mixture model, the Support Vector Machine, and gradient boosting (LightGBM) were used as the algorithms for the trained model LM. Hereinafter, the Bayesian-Gaussian mixture model, the Support Vector Machine, and gradient boosting will be referred to as the first example, the second example, and the third example, respectively.

[0088] In the first to third embodiments, the same dataset was used in each phase: the learning phase, the trial phase (verification phase), and the practical phase (estimation phase). These datasets correspond to 12,214 unit intervals obtained from multiple cross-sectional profiles of the surface model SM of 13 parts. In these embodiments, in each of the first to third embodiments, 7,816 (64% of the total) of the 12,214 unit intervals were used in the learning phase, 1,955 (16% of the total) were used in the trial phase, and 2,443 (20% of the total) were used in the practical phase. In the learning phase, the machine learning device 3 was used, and the trained model LM obtained by the machine learning device 3 executing the machine learning method M3 was provided to the strain region estimation device 1. In the trial and practical phases, the strain region estimation device 1 was used.

[0089] In estimation step S12 using the trained model LM, as described above, 20 features corresponding to each unit interval were taken as input, and the estimation result of whether or not a strained region was included in that unit interval was output. The output estimation result was "1" if a strained region was included in the unit interval, and "0" if a strained region was not included in the unit interval.

[0090] Furthermore, the evaluation of the estimation results using the pre-trained model LM was performed in the following three ways. • Correct answer: The trained model LM also estimated the areas that the worker identified as distorted to be distorted. • Not detected: The trained model LM did not estimate the areas that the worker identified as having distortion as distortion. • Over-detection: The trained model LM estimated areas as distortion when the worker did not identify them as such.

[0091] For correct answers, a higher percentage (accuracy rate) results in a higher score. For undetected items, a lower percentage (undetection rate) results in a higher score. For overdetected items, it is considered acceptable if the percentage (overdetection rate) does not reach 0%.

[0092] Furthermore, determining the presence or absence of distortion in a component can be difficult even for an operator, and depends on the operator's skill level. Therefore, if the false positive rate is extremely low, it is likely that questionable cases that could be considered distortion are being overlooked, making it undesirable as a trained model (LM).

[0093] In the first embodiment, a pre-trained model LM employing a Bayesian-Gaussian mixture model achieved 32 correct answers, 11 missed detections, and 161 over-detections in 2443 trials during the practical application phase. In the second embodiment, a pre-trained model LM employing a support vector machine achieved 34 correct answers, 9 missed detections, and 10 over-detections. In the third embodiment, a pre-trained model LM employing gradient boosting achieved 38 correct answers, 5 missed detections, and 98 over-detections.

[0094] In the first embodiment, the most common results were undetected and over-detected. Additionally, there were parts with zero correct results. Furthermore, when workers examined the unit intervals that were presumed to be over-detected, many areas did not appear to exhibit distortion.

[0095] In the second embodiment, the number of correct and undetected cases was not bad, but there were few false positives. Therefore, there is a high possibility that areas that look like distortions, which may be overlooked by some workers, could not be extracted.

[0096] In the third embodiment, the number of undetected cases was small, and the number of over-detected cases was considered appropriate, being somewhere between that of the first and second embodiments. When the workers checked the unit intervals that were presumed to be over-detected, they found that the extracted unit intervals appeared distorted depending on the perspective.

[0097] From these results, it was found that gradient boosting is preferred as the algorithm for the trained model LM used in the strain region estimation system S.

[0098] 〔summary〕 A strain region estimation device according to a first aspect of the present invention includes a control unit that performs an estimation step of estimating strain regions that may be included in a three-dimensional surface model and that appear visually distorted due to optical illusions.

[0099] According to the above configuration, it is possible to detect distortions that may occur in a three-dimensional surface model, such as visual distortions caused by optical illusions.

[0100] Furthermore, in the strain region estimation device according to the second aspect of the present invention, in addition to the configuration of the strain region estimation device according to the first aspect described above, the control unit further performs a preprocessing step before the estimation step, the estimation step estimates strain regions that may be included in the lines included in the surface model and that appear visually distorted due to optical illusions, using a trained model constructed by supervised learning, the preprocessing step includes a feature quantity derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving feature quantities that indicate the shape characteristics of each unit interval, the input to the trained model is the feature quantities derived for each unit interval, and the output to the trained model is the estimation result of whether or not each unit interval includes the strain region.

[0101] According to the above configuration, a trained model, which has been input with features derived for each unit interval, can output an estimation result indicating whether or not each unit interval contains the aforementioned distortion region. Therefore, by using this distortion region estimation device, it is possible to detect distortions that may occur in a three-dimensional surface model, such as visual distortions caused by optical illusions.

[0102] Furthermore, in the strain region estimation device according to the third aspect of the present invention, in addition to the configuration of the strain region estimation device according to the second aspect described above, the lines included in the surface model are the cross-sectional profile of the surface model, and the preprocessing step is configured to set the plurality of unit intervals in the cross-sectional profile and derive the feature quantities of each unit interval.

[0103] A cross-sectional profile of a surface model is a suitable line to use for deriving feature quantities.

[0104] Furthermore, in the strain region estimation apparatus according to the fourth aspect of the present invention, in addition to the configuration of the strain region estimation apparatus according to the third aspect described above, the preprocessing step further includes a cross-sectional profile generation step that generates a plurality of cross-sectional profiles from the surface profile of the surface model.

[0105] With the above configuration, it is possible to detect areas of strain that may occur in the surface model using multiple cross-sectional profiles. Therefore, users can understand the areas where strain is occurring in the surface model in two dimensions.

[0106] Furthermore, in the strain region estimation device according to the fifth aspect of the present invention, in addition to the configuration of the strain region estimation device according to the fourth aspect described above, the plurality of cross-sectional profiles generated by the cross-sectional profile generation step are a plurality of cross-sections represented in an oblique coordinate system composed of three intersecting axes, and are cross-sectional profiles of a plurality of cross-sections parallel to a plane defined by two of the three axes.

[0107] It is preferable that multiple cross-sectional profiles are parallel to each other. More preferably, adjacent cross-sectional profiles are not only parallel but also equally spaced. Equal spacing between adjacent cross-sectional profiles reduces the probability of overlooking strain regions larger than this spacing.

[0108] Furthermore, in the strain region estimation device according to the sixth aspect of the present invention, in addition to the configuration of the strain region estimation device according to the fifth aspect described above, the three axes are each the x-axis, y-axis, and z-axis, and the cross-sectional profile is a cross-sectional profile in (1) a plurality of equally spaced cross-sections parallel to the xy-plane, (2) a plurality of equally spaced cross-sections parallel to the yz-plane, and (3) a plurality of equally spaced cross-sections parallel to the zx-plane.

[0109] With the above configuration, the preprocessing step can be performed mechanically without considering the orientation of the surface model when the cross-sectional profile generation step generates the cross-sectional profile. Therefore, the efficiency of estimating the strain region can be improved.

[0110] Furthermore, in the strain region estimation device according to the seventh aspect of the present invention, in addition to the configuration of the strain region estimation device according to any one of the third to sixth aspects described above, each of the cross-sectional profiles is configured by connecting N curves (where N is an integer of 3 or more) in series, each of the curves is configured by a single spline curve, and the feature quantity derivation step employs a configuration in which the feature quantity is derived from a unit interval consisting of a first interval consisting of the i-th spline curve (where i is an integer of 2 or more and N-1 or less), a second interval consisting of the (i-1)th spline curve, and a third interval consisting of the (i+1)th spline curve.

[0111] When a surface model is CAD data handled by 3D CAD, the surface profile consists of multiple constituent surfaces. In such cases, the cross-sectional profile is constructed by connecting multiple spline curves, and therefore the above configuration can be suitably used.

[0112] Furthermore, in the strain region estimation device according to the eighth aspect of the present invention, in addition to the configuration of the strain region estimation device according to the seventh aspect described above, the first section, the second section, and the third section are each divided into M equal parts (M is an integer of 2 or more) to obtain M pseudo-line segments, which are then designated as line segment L1j, line segment L2j, and line segment L3j (j is an integer of 1 or more and less than or equal to M), respectively. The feature quantity derivation step then derives, as feature quantities for each unit section, at least the maximum value of the change in angle between adjacent line segment L1j and line segment L1j+1 in the first section, the maximum value of the change in angle between adjacent line segment L2j and line segment L2j+1 in the second section, and the maximum value of the change in angle between adjacent line segment L3j and line segment L3j+1 in the third section.

[0113] Examples of features include those mentioned above. By using at least these features, it is possible to accurately estimate the distortion occurring in a unit interval where a first interval with relatively small curvature is sandwiched between second and third intervals with relatively large curvature.

[0114] A distortion region estimation method according to the ninth aspect of the present invention includes an estimation step of estimating distortion regions that may be included in a three-dimensional surface model and that appear visually distorted due to optical illusions. Furthermore, in the strain region estimation method according to the tenth aspect of the present invention, in addition to the configuration of the strain region estimation method according to the ninth aspect described above, the method further includes a preprocessing step performed before the estimation step, wherein the estimation step uses a trained model constructed by supervised learning to estimate strain regions that may be included in the lines included in the surface model and that appear visually distorted due to optical illusions, and the preprocessing step includes a feature quantity derivation step in which a plurality of unit intervals are set in the lines included in the surface model and feature quantities that indicate the shape characteristics of each unit interval, wherein the input to the trained model is the feature quantities derived for each unit interval and the output to the trained model is the estimation result of whether or not each unit interval includes the strain region.

[0115] Furthermore, in the strain region estimation method according to the 11th aspect of the present invention, in addition to the configuration of the strain region estimation method according to the 10th aspect described above, the lines included in the surface model are the cross-sectional profile of the surface model, and the preprocessing step is configured to set the plurality of unit intervals in the cross-sectional profile and derive the feature quantities of each unit interval.

[0116] Furthermore, in the strain region estimation method according to the 12th aspect of the present invention, in addition to the configuration of the strain region estimation method according to the 11th aspect described above, the preprocessing step further includes a cross-sectional profile generation step that generates a plurality of cross-sectional profiles from the surface profile of the surface model.

[0117] The strain region estimation methods according to the ninth to twelfth aspects of the present invention each achieve the same effects as the strain region estimation devices according to the first to fourth aspects of the present invention.

[0118] A machine learning apparatus according to a thirteenth aspect of the present invention includes a control unit that performs a preprocessing step and a construction step of constructing a trained model that estimates distortion regions that may be included in lines included in a three-dimensional surface model and that appear visually distorted due to optical illusions, by supervised learning using a training dataset, wherein the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving feature quantities that indicate the shape characteristics of each unit interval, the input to the trained model is the feature quantities derived for each unit interval, and the output to the trained model is an estimation result of whether or not each unit interval includes the distortion region.

[0119] A machine learning method according to a fourteenth aspect of the present invention includes a preprocessing step and a construction step in which a control unit constructs a trained model that estimates distortion regions that may be included in lines included in a three-dimensional surface model and that appear visually distorted due to optical illusions, by supervised learning using a training dataset, wherein the preprocessing step includes a feature derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving feature quantities that indicate the shape characteristics of each unit interval, the input to the trained model is the feature quantities derived for each unit interval, and the output to the trained model is an estimation result of whether or not each unit interval includes the distortion region.

[0120] The machine learning apparatus according to the thirteenth aspect of the present invention and the machine learning method according to the fourteenth aspect of the present invention each achieve the same effect as the strain region estimation apparatus according to the second aspect of the present invention.

[0121] [Additional Notes] The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Other embodiments obtained by appropriately combining the technical means disclosed in each of the embodiments described above are also included in the technical scope of the present invention. [Explanation of symbols]

[0122] 1. Strain region estimation device 11 memory 12 processors M1 Strain Region Estimation Method S11 Pretreatment step S12 Estimation Step 2D 3D CAD 3. Machine Learning Equipment 31 Storage 32 processors 33 memory LM pre-trained model DS training dataset

Claims

1. A control unit that performs an estimation step of estimating distortion regions that may be included in a three-dimensional surface model and that appear visually distorted due to optical illusions, The control unit further performs a preprocessing step before the estimation step, The estimation step involves using a trained model constructed through supervised learning to estimate distortion regions that may be included in the lines of the surface model and that appear visually distorted due to optical illusions. The preprocessing step includes a feature quantity derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving a feature quantity that shows the shape characteristics of each unit interval, The input to the aforementioned trained model is the feature quantities derived for each unit interval, The output of the trained model is an estimation result of whether or not each unit interval contains the strain region. A strain region estimation device characterized by the following features.

2. The lines included in the surface model are the cross-sectional profiles of the surface model, The preprocessing step involves setting the plurality of unit intervals in the cross-sectional profile and deriving the feature quantities for each unit interval. The strain region estimation device according to feature 1.

3. The preprocessing step further includes a cross-sectional profile generation step of generating a plurality of cross-sectional profiles from the surface profile of the surface model. The strain region estimation device according to feature 2.

4. The plurality of cross-sectional profiles generated by the cross-sectional profile generation step are a plurality of cross-sections represented in an oblique coordinate system composed of three intersecting axes, and are cross-sectional profiles of a plurality of cross-sections parallel to a plane defined by two of the three axes. The strain region estimation device according to feature 3.

5. Let each of the three axes be the x-axis, y-axis, and z-axis. The aforementioned cross-sectional profile is a cross-sectional profile in (1) a plurality of cross-sections parallel to the xy plane and equally spaced, (2) a plurality of cross-sections parallel to the yz plane and equally spaced, and (3) a plurality of cross-sections parallel to the zx plane and equally spaced. The strain region estimation device according to feature 4.

6. Each of the aforementioned cross-sectional profiles is constructed by connecting N curves (where N is an integer of 3 or more) in series. Each of the aforementioned curves is composed of a single spline curve, The feature derivation step involves deriving the feature quantities for each unit interval, using an interval consisting of a first interval comprising the i-th spline curve (where i is an integer between 2 and N-1), a second interval comprising the (i-1)th spline curve, and a third interval comprising the (i+1)th spline curve as the unit interval. A strain region estimation device according to any one of claims 2 to 5.

7. The M pseudo-line segments obtained by dividing each of the first, second, and third sections into M equal parts (where M is an integer greater than or equal to 2) are designated as line segment L1j, line segment L2j, and line segment L3j (where j is an integer greater than or equal to 1 and less than or equal to M), respectively. The feature derivation step derives, as feature quantities for each unit interval, at least the maximum value of the change in angle between adjacent line segments L1j and L1j+1 in the first interval, the maximum value of the change in angle between adjacent line segments L2j and L2j+1 in the second interval, and the maximum value of the change in angle between adjacent line segments L3j and L3j+1 in the third interval. The strain region estimation device according to feature 6.

8. The method includes an estimation step of estimating distortion regions that may be included in a three-dimensional surface model and that appear visually distorted due to optical illusions, The process further includes a preprocessing step performed prior to the estimation step, The estimation step involves using a trained model constructed through supervised learning to estimate distortion regions that may be included in the lines of the surface model and that appear visually distorted due to optical illusions. The preprocessing step includes a feature quantity derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving a feature quantity that shows the shape characteristics of each unit interval, The input to the aforementioned trained model is the feature quantities derived for each unit interval, The output of the trained model is an estimation result of whether or not each unit interval contains the strain region. A method for estimating a strain region, characterized by the features described above.

9. The lines included in the surface model are the cross-sectional profiles of the surface model, The preprocessing step involves setting the plurality of unit intervals in the cross-sectional profile and deriving the feature quantities for each unit interval. The strain region estimation method according to feature 8.

10. The preprocessing step further includes a cross-sectional profile generation step of generating a plurality of cross-sectional profiles from the surface profile of the surface model. The strain region estimation method according to feature 9.

11. The system includes a control unit that performs a preprocessing step and a construction step that constructs a trained model that estimates distortion regions that may be included in lines of a three-dimensional surface model and that appear visually distorted due to optical illusions, using supervised learning with a training dataset. The preprocessing step includes a feature quantity derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving a feature quantity that shows the shape characteristics of each unit interval, The input to the aforementioned trained model is the feature quantities derived for each unit interval, The output of the trained model is an estimation result of whether or not each unit interval contains the strain region. A machine learning device characterized by the following features.

12. The process includes a preprocessing step and a construction step, which involves supervised learning using a training dataset to construct a trained model that estimates distortion regions that may be included in lines within a three-dimensional surface model and that appear visually distorted due to optical illusions. The preprocessing step includes a feature quantity derivation step of setting a plurality of unit intervals in the lines included in the surface model and deriving a feature quantity that shows the shape characteristics of each unit interval, The input to the aforementioned trained model is the feature quantities derived for each unit interval, The output of the trained model is an estimation result of whether or not each unit interval contains the strain region. A machine learning method characterized by the following features.

Images