KR20220120414A - Apparatus for determining the shape of a preform using a artificial neural network and method therefor - Google Patents

Abstract

A method for predicting a shape of a preform using an artificial neural network comprises: a step of converting, by a pre-processing part, a forged shape, for which is a 3D image, into a forged shape matrix, for which is a 3D matrix expression; a step of deriving, by a shape prediction part, a plurality of predicted shape matrices, for which are matrix expressions of the shape of the preform for the forged shape, through a plurality of prediction models learned to predict the shape of the preform; a step of converting, by a post-processing part, the plurality of predicted shape matrices into a plurality of predicted shapes for which are 3D images; and a step of selecting, by a shape determination part, one of the plurality of predicted shapes that best meets a preset design condition as the preform.

Description

Apparatus for determining the shape of a preform using a artificial neural network and method therefor

The present invention relates to a technique for predicting the shape of a preform, and more particularly, to an apparatus for predicting the shape of a preform using an artificial neural network and a method therefor.

Forging refers to a processing method that creates a desired shape of a material through linear motion of a mold. Since the shape that can be obtained according to the flow of the material is limited, it is generally forged through several steps, and the forging shape before the final product is called a preform. The design of the preform prevents forging defects such as non-filling and overlapping during forging into the final shape, and reduces the forging load, so that the mold life can be improved. However, the design of the preform was mainly carried out by the trial error of the experienced person, and it is highly dependent on the experience of the engineer.

Korea Patent Publication No. 2014-0126306 (published on October 30, 2014)

An object of the present invention is to provide an apparatus for predicting the shape of a preform using an artificial neural network and a method therefor.

A method for learning a model for predicting the shape of a preform according to a preferred embodiment of the present invention for achieving the object as described above includes the steps of: a forging shape generating unit generating a forged shape; generating a plurality of corresponding candidate preforms; selecting, by the label setting unit, any one of the plurality of candidate preforms corresponding to the forging shape, as a label preform corresponding to the forging shape; and a learning unit and training a predictive model for predicting the shape of the preform by using the forged shape and the label preform corresponding to the forged shape as learning data.

The step of selecting the preform includes the label setting unit performing finite element analysis on the plurality of preform candidates to derive the mold filling rate, the presence of overlapping defects, and forging load, and the label setting unit performing finite element analysis of the finite element analysis. and selecting one of the plurality of preform candidates as a label preform corresponding to the forging shape according to the result.

In the step of selecting one of the plurality of preform candidates as the label preform corresponding to the forging shape according to the result of the finite element analysis, the label setting unit has an unfilled defect according to the mold filling rate among the plurality of preform candidates. specifying a formed body candidate and excluding the specified preform candidate, the label setting unit excluding a preform candidate having an overlap defect from among the plurality of preform candidates, the label setting unit including the unfilled defect and and selecting a preform candidate having the lowest forging load among a plurality of preform candidates without the overlapping defect as a label preform corresponding to the forging shape.

The step of generating the forging shape includes the steps of: setting a bus bar at a position spaced apart by a preset radius along a bottom line indicating a bottom surface perpendicular to the rotation axis from a rotation axis perpendicular to the ground by the forging shape generating unit; allocating a control point along a bottom line between the rotation axis and the bus bar by a generating unit; and combining a plurality of preset height levels by the forging shape generating unit in a vertical direction from the bottom line of the rotation axis, the bus bar, and the control point. setting, the forging shape generating unit generating a connecting line by connecting the rotation shaft, the bus bar, and the control point at which the height is set; and generating a forged shape by rotating it along the rotation axis.

The generating of the plurality of candidate preforms includes: setting the busbars at positions spaced apart by the same radius as the corresponding forging shape along a reference line perpendicular to the rotation axis and the vertical axis of the label setting unit perpendicular to the ground; Allocating a plurality of control points along a reference line between the rotation axis and the bus bar, and setting the height in a vertical direction from the reference line of the rotation axis, the bus bar and the control point by combining a plurality of height levels preset by the label setting unit And, the label setting unit connecting the rotation shaft, the bus bar, and the control point to which the height is set to generate a connecting line; setting, and generating a candidate preform by rotating the label setting unit along the axis of rotation, the connecting line, the bus bar, and the bottom line.

In the step of setting the height between the bottom line and the reference line, the volume of the candidate preform generated by the label setting unit rotating the rotation shaft, the connecting line, the bus bar and the bottom line, and the volume of the forging shape corresponding to the candidate preform It is characterized in that the height between the bottom line and the reference line is set to match.

The step of learning the predictive model comprises the steps of: a learning unit providing a plurality of learning data including a plurality of forged shapes and a plurality of label preforms corresponding to each of the plurality of forged shapes; It includes the steps of classifying into a plurality of training data groups, and generating a plurality of predictive models by performing learning by using the training data belonging to the corresponding training data group for each training data group.

A method for learning a model for predicting the shape of a preform according to a preferred embodiment of the present invention for achieving the object as described above includes the steps of: a forging shape generating unit generating a forged shape; Generating a corresponding label preform, the learning unit using the forging shape and the label preform corresponding to the forging shape as learning data, it comprises the steps of learning a predictive model for predicting the shape of the preform.

The step of learning the predictive model comprises the steps of: a learning unit providing a plurality of learning data including a plurality of forged shapes and a plurality of label preforms corresponding to each of the plurality of forged shapes; It includes the steps of classifying into a plurality of training data groups, and generating a plurality of predictive models by performing learning by using the training data belonging to the corresponding training data group for each training data group.

Classifying the plurality of learning data groups into the plurality of learning data groups is characterized in that the plurality of monotonic shapes are classified into a plurality of learning data groups based on a combination of height differences between a plurality of feature points of the monotone shapes.

In the step of classifying the plurality of learning data groups, the learning unit represents a combination of a first condition indicating a combination of height differences between a predetermined number of feature points in an order close to the rotation axis, and a combination of a height difference between a predetermined number of feature points in an order close to the mother bar providing a second condition; and classifying the plurality of monotone shapes into the plurality of learning data groups according to whether the learning unit satisfies the first condition or the second condition.

The step of generating the plurality of predictive models includes the steps of, by the learning unit, voxelizing the monotonic shape and label preform of the learning data belonging to any one learning data group into a monotonic shape matrix and a label preform matrix in matrix form; , a predictive model deriving a predicted shape matrix that is a matrix expression predicting the shape of the preform through a plurality of operations to which a weight for which the learning between a plurality of layers is not completed with respect to the monotonic shape matrix is applied, and the learning unit Comprising the steps of calculating a loss representing the difference between the predicted shape matrix and the label preform matrix through a loss function, and performing optimization of modifying the weight of the prediction model so that the loss is minimized.

를 통해 상기 손실을 산출하고, 상기

는 예측형상행렬과 레이블 예비성형체행렬과의 차이를 나타내는 손실이고, 상기

는 단조형상행렬에 대응하는 레이블인 레이블 예비성형체행렬이고, 상기

는 입력된 단조형상행렬로부터 도출된 예측형상행렬인 것을 특징으로 한다. The step of calculating the loss is a loss function of the learning unit

Calculate the loss through

is the loss representing the difference between the predicted shape matrix and the label preform matrix,

is a label preform matrix that is a label corresponding to the forging shape matrix, and

is a predicted shape matrix derived from the input monotonic shape matrix.

In the step of converting the monotonic shape matrix and the label preform matrix, the learning unit maps the longest side of the forged shape and the label preform, which are mesh files, to 1 voxel, and the mesh file corresponds to the volume occupied by any one voxel. It is characterized in that the voxel is converted into a monotonic shape matrix and a label preform matrix by assigning a value of 0 or 1.

An apparatus for training a model for predicting the shape of a preform according to a preferred embodiment of the present invention for achieving the above object includes a forging shape generating unit for generating a forged shape, and a plurality of forging shapes corresponding to the forged shape. a label setting unit generating a candidate preform and selecting any one of the plurality of candidate preforms corresponding to the forging shape as a label preform corresponding to the forging shape; the forging shape and corresponding to the forging shape and a learning unit for learning a predictive model for predicting the shape of the preform using the label preform as learning data.

The label setting unit performs finite element analysis on the plurality of preform candidates to derive the mold filling rate, the presence of overlapping defects, and forging load, and according to the result of the finite element analysis, one of the plurality of preform candidates is applied to the forging shape. Characterized in selecting the corresponding label preform.

The label setting unit specifies a preform candidate having an unfilled defect according to a mold filling rate among the plurality of preform candidates, excludes the specified preform candidate, and a preform candidate having an overlap defect among the plurality of preform candidates. , and a preform candidate having the lowest forging load among a plurality of preform candidates without the unfilled defect and the overlapping defect is selected as a label preform corresponding to the forging shape.

The forging shape generating unit sets a bus bar at a position spaced apart by a preset radius along a bottom line indicating a bottom surface perpendicular to the rotation axis on a rotation axis perpendicular to the ground, and a control point along the bottom line between the rotation axis and the bus bar. Allocate, combine a plurality of preset height levels to set the vertical height from the bottom line of the rotation shaft, the bus bar, and the control point, and connect the rotation shaft, the bus bar, and the control point to which the height is set to create a connection line and rotating the rotation shaft, the connecting line, the bus bar, and the bottom line along the rotation axis to generate a forged shape.

The label setting unit sets the bus bar at a position spaced apart by the same radius as the corresponding forging shape along the reference line perpendicular to the rotation axis perpendicular to the ground, and allocates a plurality of control points along the reference line between the rotation axis and the bus bar, By combining a plurality of set height levels, a height in the vertical direction is set from a reference line of the rotation shaft, the bus bar, and the control point, and a connection line is generated by connecting the rotation shaft, the bus bar and the control point at which the height is set, and a plurality of preset Select any one of the height levels of , set the height between the bottom line and the reference line, and rotate the rotation shaft, the connecting line, the bus bar, and the bottom line along the rotation axis to generate a candidate preform.

The label setting unit sets the height between the bottom line and the reference line so that the volume of the candidate preform generated by the rotation of the rotation shaft, the connecting line, the bus bar, and the bottom line matches the volume of the forging shape corresponding to the candidate preform characterized in that

The learning unit provides a plurality of learning data including a plurality of forging shapes and a plurality of label preforms corresponding to each of the plurality of forging shapes, classifying the plurality of training data into a plurality of training data groups, and the learning It is characterized in that for each data group, a plurality of predictive models are generated by performing learning using the training data belonging to the corresponding training data group.

An apparatus for training a model for predicting the shape of a preform according to a preferred embodiment of the present invention for achieving the above object includes a forging shape generating unit generating a forged shape and a label preform corresponding to the forging shape. It includes a label setting unit for generating the forged shape, and a learning unit for learning a predictive model for predicting the shape of the preform using the forged shape and the label preform corresponding to the forged shape as learning data.

The learning unit provides a plurality of learning data including a plurality of forged shapes and a plurality of label preforms corresponding to each of the plurality of forging shapes, and the learning unit classifies the plurality of learning data into a plurality of training data groups. and generating a plurality of predictive models by performing learning by using the training data belonging to the corresponding training data group for each training data group.

Classifying the plurality of learning data groups into the plurality of learning data groups is characterized in that the plurality of monotonic shapes are classified into a plurality of learning data groups based on a combination of height differences between a plurality of feature points of the monotone shapes.

The learning unit provides a first condition indicating a combination of a height difference between a predetermined number of feature points in an order close to the rotation axis, and a second condition indicating a combination of a height difference between a predetermined number of feature points in an order close to the mother bar, and the first It is characterized in that the plurality of monotonic shapes are classified into the plurality of learning data groups according to whether the condition or the second condition is satisfied.

The learning unit voxelizes the monotonic shape and label preform of the learning data belonging to any one training data group and converts it into a monotonic shape matrix and a label preform matrix in matrix form, and a predictive model includes a plurality of monotonic shape matrices for the monotonic shape matrix. If the predicted shape matrix, which is a matrix expression that predicted the shape of the preform, is derived through a plurality of operations to which the inter-layer learning is not completed, the weight of the preform is applied. It is characterized in that the optimization of calculating the loss and correcting the weight of the prediction model so that the loss is minimized is performed.

를 통해 상기 손실을 산출하고, 상기

는 예측형상행렬과 레이블 예비성형체행렬과의 차이를 나타내는 손실이고, 상기

는 단조형상행렬에 대응하는 레이블인 레이블 예비성형체행렬이고, 상기

는 입력된 단조형상행렬로부터 도출된 예측형상행렬인 것을 특징으로 한다. The learning unit is a loss function

Calculate the loss through

is the loss representing the difference between the predicted shape matrix and the label preform matrix,

is a label preform matrix that is a label corresponding to the forging shape matrix, and

is a predicted shape matrix derived from the input monotonic shape matrix.

The learning unit maps the longest side to 1 voxel in the monotonic shape and label preform that are the mesh files, and assigns a value of 0 or 1 to the voxel according to the volume occupied by the voxel to any one of the mesh files, thereby forming a monotonic shape matrix and It is characterized in that it is converted into a label preform matrix.

In the method for predicting the shape of a preform using an artificial neural network according to a preferred embodiment of the present invention for achieving the above object, the pre-processing unit converts a monotone shape, which is a three-dimensional image, into a monotone shape matrix, which is a three-dimensional matrix expression. and deriving a plurality of predicted shape matrices that are matrix representations of the shape of the preform for the forged shape through a plurality of predictive models trained to predict the shape of the preform by the shape prediction unit, and the post-processing unit includes a plurality of It includes the steps of converting the predicted shape matrix of the three-dimensional image into a plurality of predicted shapes, and the shape determining unit selecting any one of the plurality of predicted shapes that best meets preset design conditions as a preform.

The step of selecting the preform includes the steps of excluding or correcting a predicted shape having a singular shape among the plurality of predicted shapes by the shape determining unit, and performing a finite element analysis on the plurality of predicted shapes by the shape determining unit to form a mold. Deriving the filling factor, the existence of overlapping defects, and the forging load, and the shape determining unit selecting one of the plurality of predicted shapes as a preform corresponding to the forging shape according to the result of the finite element analysis.

In the step of excluding or correcting the predicted shape having the singular shape, the shape determining unit checks the section at a predetermined interval with respect to the predicted shape, and if the difference between the areas in the neighboring sections is greater than or equal to a preset threshold, the corresponding part is set to the singular shape. characterized in that it is specified as

The step of excluding or correcting the predicted shape having the singular shape is specified as the singular shape, but when the cross-section is blocked in all directions, the shape determining unit fills the corresponding part with an image representing the predicted shape to correct it.

In the step of selecting one of the plurality of predicted shapes as a preform corresponding to the forging shape, the shape determining unit specifies a predicted shape with unfilled defects from among the plurality of predicted shapes according to the mold filling rate, and selects the specified predicted shape. The step of excluding, the step of excluding the predicted shape having an overlap defect from among the plurality of predicted shapes by the shape determining unit, and the shape determining unit having the lowest forging load among the plurality of predicted shapes without the unfilled defect and the overlapping defect and selecting the predicted shape recorded with the preform corresponding to the forged shape.

In the step of deriving the plurality of predictive shape matrices, the plurality of predictive models derive a plurality of predictive shape matrices by performing a plurality of operations to which the weights learned between a plurality of layers are applied to the monotonic shape matrix. do.

In the step of converting the monotonic shape into a monotonic shape matrix, which is a three-dimensional matrix expression, the preprocessor maps the longest side of the monotonic shape, which is a mesh file, to 1 voxel, and the mesh file corresponds to the volume occupied by any one voxel. It is characterized in that the voxel is assigned a value of 0 or 1.

In the step of converting the plurality of predicted shape matrices into a plurality of predicted shapes that are three-dimensional images, the post-processing unit converts a plurality of predicted shape matrices that are voxel files through a marching cube algorithm according to the scale to a plurality of predicted shapes that are mesh files. and performing a smoothing process on the surfaces of a plurality of predicted shapes that are mesh files using a smoothing algorithm.

The method includes a plurality of learning data including a plurality of forging shapes and a plurality of label preforms corresponding to each of the plurality of forging shapes, before the step of converting the monotonic shape into a monotonic shape matrix which is a three-dimensional matrix expression. preparing, classifying the plurality of learning data into a plurality of learning data groups by the learning unit; It further comprises the step of generating a predictive model of.

In a method for predicting the shape of a preform using an artificial neural network according to a preferred embodiment of the present invention for achieving the object as described above, the learning unit uses a plurality of learning data groups each including a plurality of learning data. Generating a plurality of predictive models that are learned to predict the preform corresponding to the forged shape, and the formed body derivation unit predicting the preform corresponding to the forged shape using the plurality of predictive models.

The generating of the plurality of predictive models includes the steps of classifying, by the learning unit, a plurality of learning data including a plurality of forged shapes and a plurality of label preforms corresponding to each of the plurality of forged shapes into a plurality of training data groups; , generating, by the learning unit, a plurality of predictive models corresponding to each of the plurality of training data groups.

The classifying into the plurality of learning data groups is characterized in that the learning unit classifies the plurality of monotone shapes into a plurality of learning data groups based on a combination of height differences between a plurality of feature points of the monotone shapes.

In the step of classifying the plurality of learning data groups, the learning unit is a combination of a first condition indicating a combination of differences between the heights of a predetermined number of feature points in an order close to the rotation axis, and a difference between the heights of a predetermined number of feature points in an order close to the mother bar and preparing a second condition indicating that , and classifying the plurality of monotone shapes into the plurality of learning data groups according to whether the learning unit satisfies the first condition or the second condition.

In the step of generating a plurality of predictive models corresponding to each of the plurality of training data groups, the learning unit voxelizes monotone shapes and label preforms of training data belonging to any one training data group to form a monotonic shape matrix in matrix form and A predictive shape matrix that is a matrix expression in which the shape of the preform is predicted through a plurality of operations in which a weight is applied for which the learning between a plurality of layers has not been completed for the step of converting to a label preform matrix and the prediction model for the monotonic shape matrix The step of deriving , the learning unit calculating a loss representing the difference between the predicted shape matrix and the label preform matrix through the loss function, and optimization of correcting the weight of the predictive model so that the loss is minimized including the steps of

The predicting of the preform corresponding to the forging shape using the plurality of prediction models includes the steps of: the preprocessing unit transforming the forged shape, which is a three-dimensional image, into a monotonic shape matrix, which is a three-dimensional matrix expression; and the shape prediction unit of the shape of the preform Deriving a plurality of predictive shape matrices that are matrix representations of the shape of the preform for the forged shape through a plurality of predictive models trained to predict It includes the steps of converting to a predicted shape, and selecting, by a shape determining unit, any one of a plurality of predicted shapes that best meets preset design conditions as a preform.

The step of selecting the preform includes the steps of excluding or correcting a predicted shape having a singular shape among the plurality of predicted shapes by the shape determining unit, and performing a finite element analysis on the plurality of predicted shapes by the shape determining unit to form a mold. Deriving the filling factor, the existence of overlapping defects, and the forging load, and the shape determining unit selecting one of the plurality of predicted shapes as a preform corresponding to the forging shape according to the result of the finite element analysis.

The apparatus for predicting the shape of a preform using an artificial neural network according to a preferred embodiment of the present invention for achieving the above object is a pre-processing of converting a monotonic shape, which is a three-dimensional image, into a monotonic shape matrix, which is a three-dimensional matrix expression. and a shape prediction unit for deriving a plurality of predicted shape matrices that are matrix representations of the shape of the preform for the forged shape through a plurality of predictive models trained to predict the shape of the preform, and a plurality of predicted shape matrices It includes a post-processing unit that converts a plurality of predicted shapes that are three-dimensional images, and a shape determining unit that selects one of the plurality of predicted shapes that best meets a preset design condition as a preform.

The shape determining unit excludes or corrects a predicted shape having a singular shape among the plurality of predicted shapes, and performs a finite element analysis on the plurality of predicted shapes to derive a mold filling rate, whether there is an overlap defect, and a forging load, It is characterized in that one of the plurality of predicted shapes is selected as a preform corresponding to the forged shape according to the result of element analysis.

The shape determining unit checks the cross-section at a predetermined interval with respect to the predicted shape, and when the difference between the areas in the neighboring cross-sections is equal to or greater than a preset threshold, the corresponding part is specified as a singular shape.

Although the shape determining part is specified as a specific shape, when the cross-section is blocked in all directions, the shape is corrected by filling the corresponding part with an image representing the predicted shape.

The shape determining unit specifies a predicted shape having an unfilled defect according to a mold filling rate among the plurality of predicted shapes, excludes the specified predicted shape, and excludes a predicted shape having an overlap defect from among the plurality of predicted shapes, It is characterized in that the shape determining unit selects the predicted shape having the lowest forging load among the plurality of predicted shapes without the unfilled defect and the overlapping defect as a preform corresponding to the forging shape.

The plurality of prediction models is characterized in that the plurality of prediction shape matrices are derived by performing a plurality of operations to which the weights learned between a plurality of layers are applied to the monotonic shape matrix.

The preprocessor may map the longest side of a monotonic shape that is a mesh file to 1 voxel, and assign a value of 0 or 1 to a corresponding voxel according to a volume occupied by a voxel in any one of the mesh files.

The pre-processing unit converts a plurality of predicted shape matrices that are voxel files into a mesh file through a marching cube algorithm according to the scale, and smoothes the surfaces of a plurality of predicted shapes that are mesh files using a smoothing algorithm. characterized by performing.

The apparatus provides a plurality of training data including a plurality of forging shapes and a plurality of label preforms corresponding to each of the plurality of forging shapes, classifying the plurality of training data into a plurality of training data groups, and the training data Each group further includes a learning unit for generating a plurality of predictive models by performing learning using a plurality of learning data belonging to the corresponding learning data group.

An apparatus for predicting the shape of a preform using an artificial neural network according to a preferred embodiment of the present invention for achieving the above object is a monotonic shape using a plurality of learning data groups each including a plurality of learning data. and a learning unit for generating a plurality of predictive models that are learned to predict a preform corresponding to , and a formed body derivation unit for predicting a preform corresponding to a forged shape using the plurality of predictive models.

The learning unit classifies a plurality of training data including a plurality of forging shapes and a plurality of label preforms corresponding to each of the plurality of forging shapes into a plurality of training data groups, and a plurality of training data groups corresponding to each of the plurality of training data groups. It is characterized in that it generates a predictive model of

The learning unit classifies the plurality of monotonous shapes into a plurality of learning data groups based on a combination of height differences between a plurality of feature points of the forged shapes.

The learning unit provides a first condition indicating a combination of differences between the heights of a predetermined number of feature points in an order close to the rotation axis, and a second condition indicating a combination of differences between the heights of a predetermined number of feature points in an order close to the mother bar, and It is characterized in that the plurality of monotonic shapes are classified into the plurality of learning data groups according to whether the first condition or the second condition is satisfied.

The learning unit voxelizes the monotonic shape and label preform of the learning data belonging to any one training data group and converts it into a monotonic shape matrix and a label preform matrix in matrix form, and a predictive model includes a plurality of monotonic shape matrices for the monotonic shape matrix. If the predicted shape matrix, which is a matrix expression that predicted the shape of the preform, is derived through a plurality of operations to which the inter-layer learning is not completed, the weight of the preform is applied. It is characterized in that the optimization of calculating the loss and correcting the weight of the prediction model so that the loss is minimized is performed.

The shape of the preform for the forging shape through a pre-processing unit that converts the forged shape, which is a three-dimensional image, into a forged shape matrix, which is a three-dimensional matrix expression, and a plurality of predictive models learned to predict the shape of the preform. A shape prediction unit for deriving a plurality of predicted shape matrices, which is a matrix expression of and a shape-determining unit that selects any one of them as a preform.

The shape determining unit excludes or corrects a predicted shape having a singular shape among the plurality of predicted shapes, and performs finite element analysis on the plurality of predicted shapes to derive a mold filling rate, whether there is an overlap defect, and a forging load, and It is characterized in that the shape determining unit selects one of the plurality of predicted shapes as a preform corresponding to the forging shape according to the result of the finite element analysis.

In the method for determining the shape of the preform corresponding to the non-axially symmetric forged shape using an artificial neural network according to a preferred embodiment of the present invention for achieving the object as described above, when the shape processing unit inputs the non-axially symmetric forged shape, the input Generating a plurality of different axisymmetric forging shapes from the non-axially symmetric forging shape; A step of deriving a non-axially symmetric preform, the preform derivation unit deriving a plurality of axisymmetric preforms from the plurality of axisymmetric forging shapes using the plurality of predictive models; Comparing the forging load of the non-axially symmetric preform derived through the finite element analysis by the shape selection unit and the forging load of the simulated preform and the forging of the non-axially symmetric preform and selecting a non-axially symmetric preform as a final preform when the difference between the load and the forging load of the simulated preform is less than a preset reference value.

The generating of the plurality of different axisymmetric forged shapes includes: recognizing a rotation axis that the shape processing unit vertically passes through the center of gravity of the non-axially symmetric forging shape; and the shape processing unit rotates at different angles based on the rotation axis Generating a plurality of longitudinal cross-sections that are cross-sections cut by a plane including the rotation axis along each, and the shape processing unit rotating the plurality of longitudinal cross-sections based on the rotation axis to generate the plurality of different forging shapes include

The plurality of axisymmetric preforms are characterized in that they are divided according to the rotation angle at which the longitudinal section is generated.

The step of generating the simulated preform is to extract a part of the axisymmetric preform according to the rotation angle applied when the shape selection unit generates the longitudinal section, and combine the extracted parts to produce the simulated preform. do.

The step of deriving the non-axially symmetric preform includes the steps of: a pre-processing unit converting a non-axially symmetric forging shape into a non-axially symmetric forging shape matrix through voxelization; A step of deriving a plurality of predicted shape matrices that are matrix expressions of the predicted shape of the symmetric preform, and a post-processing unit transforming the plurality of predicted shape matrices into a plurality of predicted shapes through post-processing of the plurality of predicted shape matrices The step of excluding or correcting the predicted shape having a singular shape among the plurality of predicted shapes by the shape determining unit, and the step of the shape determining unit excluding the predicted shape having unfilled defects and overlapping defects among the plurality of predicted shapes and selecting a predicted shape having the lowest forging load among predicted shapes from which the shape determining part is not excluded as the non-axially symmetric preform corresponding to the non-axially symmetric forged shape.

The step of deriving the plurality of axisymmetric preforms includes, for each of the plurality of axisymmetric forging shapes, the preprocessing unit converting the axisymmetric forging shape into an axisymmetric forging shape matrix through voxelization; Deriving a plurality of predicted shape matrices that are matrix expressions of the predicted shape of the axisymmetric preform from the axisymmetric monotonic shape matrix through a predictive model, and a post-processing unit through post-processing of the plurality of predicted shape matrices converting a plurality of predicted shape matrices into a plurality of predicted shapes; excluding or correcting, by a shape determining unit, a predicted shape having a singular shape among the plurality of predicted shapes; excluding the predicted shape having unfilled defects and overlapping defects; including selecting.

In the method, before the step of generating a plurality of different axisymmetric forging shapes from the non-axially symmetric forging shape, the learning unit uses a plurality of learning data including an axisymmetric forging shape and a label preform corresponding to the forging shape. The method further includes generating a plurality of predictive models.

A method for determining the shape of a preform corresponding to a non-axially symmetric forged shape using an artificial neural network according to a preferred embodiment of the present invention for achieving the object as described above is a method for determining the shape of the preform corresponding to the non-axially symmetric forged shape and the forging shape in which the learning unit is axisymmetric. Generating a plurality of predictive models trained to predict the preform from the forging shape using a plurality of training data including the corresponding label preform; Generating a plurality of different axisymmetric forging shapes from a forging shape, and deriving a non-axially symmetrical preform from the non-axially symmetric forged shape by a preform derivation unit using the plurality of predictive models; deriving a plurality of axisymmetric preforms from the plurality of axisymmetric forging shapes using the plurality of predictive models; and a shape selection unit combining the plurality of axisymmetric preforms to generate a simulated preform, and the copying and selecting the non-axially symmetric preform as the final preform through validation with the preform.

The generating of the plurality of predictive models includes the steps of, by the learning unit, voxelizing the monotonic shape and label preform of the training data to convert the monotonic shape matrix and the label preform matrix in matrix form, and the predictive model is the monotonic shape matrix Deriving a predicted shape matrix, which is a matrix expression that predicted the shape of the preform, through a plurality of operations to which a weight for which learning between a plurality of layers has not been completed for Comprising the steps of calculating a loss representing a difference from the label preform matrix, and performing optimization of modifying the weight of the prediction model so that the loss is minimized.

를 통해 상기 손실을 산출하고, 상기

는 예측형상행렬과 레이블 예비성형체행렬과의 차이를 나타내는 손실이고, 상기

는 단조형상행렬에 대응하는 레이블인 레이블 예비성형체행렬이고, 상기

는 입력된 단조형상행렬로부터 도출된 예측형상행렬인 것을 특징으로 한다. The step of calculating the loss is a loss function of the learning unit

Calculate the loss through

is the loss representing the difference between the predicted shape matrix and the label preform matrix,

is a label preform matrix that is a label corresponding to the forging shape matrix, and

is a predicted shape matrix derived from the input monotonic shape matrix.

In the step of converting the monotonic shape matrix and the label preform matrix, the learning unit maps the longest side of the forged shape and label preform, which are mesh files, to a scale of 1 voxel, and the volume occupied by the voxel in any one of the mesh files. Accordingly, it is characterized in that the voxel is converted into a monotonic shape matrix and a label preform matrix by assigning a value of 0 or 1.

The step of selecting the final preform includes the steps of: a shape selection unit combining the plurality of axisymmetric preforms to produce a simulated preform; Comparing the forging load of the imitation preform, and if the difference between the forging load of the non-axial preform and the forging load of the imitation preform is less than a preset reference value, selecting the non-axial preform as the final preform.

The apparatus for determining the shape of the preform corresponding to the non-axial forging shape using an artificial neural network according to a preferred embodiment of the present invention for achieving the object as described above A non-axially symmetrical preform is derived from the non-axially symmetric forged shape by using a shape processing unit that generates a plurality of different axisymmetric forged shapes from the forging shape, and a plurality of predictive models learned to predict the preform from the forged shape, and the A preform extracting unit for deriving a plurality of axisymmetric preforms from the plurality of axisymmetric forging shapes using a plurality of predictive models, and combining the plurality of axisymmetric preforms to generate a simulated preform; Comparing the forging load of the non-axially symmetric preform derived through element analysis and the forging load of the simulated preform, and if the difference between the forging load of the non-axially symmetric preform and the forging load of the simulated preform is less than the preset standard, non-axially symmetrical preliminary and a shape selector for selecting the molded body as the final preform.

The shape processing unit recognizes a rotational axis vertically penetrating the center of gravity of the non-axially symmetrical forged shape, and generates a plurality of longitudinal cross-sections that are cross-sections cut in a plane including the rotational axis along different rotational angles based on the rotational axis, The shape processing unit rotates the plurality of longitudinal cross-sections based on the rotation axis to generate the plurality of different forged shapes.

The plurality of axisymmetric preforms are characterized in that they are divided according to the rotation angle at which the longitudinal section is generated.

The shape selection unit extracts a part of the axisymmetric preform according to the rotation angle applied when generating the longitudinal section, and combines the extracted parts to generate the simulated preform.

The preform derivation unit includes a preprocessing unit that converts the non-axially symmetric forging shape into a non-axially symmetric forging shape matrix through voxelization, and a shape predicting the non-axially symmetric preform from the non-axially symmetric forging shape matrix through the plurality of predictive models. A shape prediction unit for deriving a plurality of predicted shape matrices that are matrix expressions of Excludes or corrects the predicted shape having a singular shape among the shapes, excludes the predicted shape having unfilled defects and overlapping defects among the plurality of predicted shapes, and the predicted shape having the lowest monotonic load among the predicted shapes that are not excluded and a shape determining unit for selecting the non-axially symmetric preform corresponding to the non-axially symmetric forging shape.

The preform derivation unit includes a pre-processing unit that converts an axisymmetric forging shape into an axisymmetric monotonic shape matrix through voxelization, and a matrix of shapes predicted from the axisymmetric monotonic shape matrix through a plurality of predictive models A shape prediction unit for deriving a plurality of predicted shape matrices as expressions; Excluding or correcting the predicted shape having a singular shape, excluding the predicted shape having unfilled defects and overlapping defects among the plurality of predicted shapes, and the predicted shape having the lowest monotonic load among the predicted shapes that are not excluded and a shape determining unit for selecting the axisymmetric preform corresponding to the axisymmetric forging shape.

The apparatus further includes a learning unit for generating a plurality of predictive models using a plurality of training data including an axisymmetric forging shape and a label preform corresponding to the forging shape.

An apparatus for determining the shape of a preform corresponding to a non-axially symmetric forged shape using an artificial neural network according to a preferred embodiment of the present invention for achieving the above object is an axisymmetric forging shape and corresponding to the forging shape. A learning unit generating a plurality of predictive models that are learned to predict a preform from a forging shape using a plurality of learning data including a label preform, and when a non-axially symmetric forging shape is input, from the input non-axially symmetric forging shape to each other a shape processing unit for generating a plurality of different axisymmetric forging shapes; and deriving a non-axially symmetric preform from the non-axially symmetric forging shape using the plurality of predictive models, and using the plurality of predictive models for the plurality of axisymmetric forgings A preform extracting unit for deriving a plurality of axisymmetric preforms from a shape, and the plurality of axisymmetric preforms are combined to produce a simulated preform, and the non-axially symmetric preform is finally preformed through verification using the simulated preform It includes a shape selection unit for selecting the molded body.

The learning unit voxelizes the monotonic shape and the label preform of the training data to convert it into a monotonic shape matrix and a label preform matrix in matrix form, and the predictive model weights the monotonic shape matrix for which learning between a plurality of layers is not completed When a predicted shape matrix, which is a matrix expression that predicts the shape of the preform, is derived through a plurality of operations to which is applied, a loss representing the difference between the predicted shape matrix and the label preform matrix is calculated through a loss function, and the loss is It is characterized in that the optimization of correcting the weight of the prediction model so as to be minimized.

를 통해 상기 손실을 산출하고, 상기

는 예측형상행렬과 레이블 예비성형체행렬과의 차이를 나타내는 손실이고, 상기

는 단조형상행렬에 대응하는 레이블인 레이블 예비성형체행렬이고, 상기

는 입력된 단조형상행렬로부터 도출된 예측형상행렬인 것을 특징으로 한다. The learning unit is a loss function

Calculate the loss through

is the loss representing the difference between the predicted shape matrix and the label preform matrix,

is a label preform matrix that is a label corresponding to the forging shape matrix, and

is a predicted shape matrix derived from the input monotonic shape matrix.

The learning unit maps the longest side of the forged shape and the label preform, which are mesh files, to a scale of 1 voxel, and assigns a value of 0 or 1 to the voxel according to the volume occupied by the voxel to any one of the mesh files. Characterized in converting to matrix and label preform matrix.

The shape selection unit generates a simulated preform by combining the plurality of axisymmetric preforms, compares the forging load of the non-axially symmetric preform derived through finite element analysis and the forging load of the simulated preform, and a non-axially symmetric preform When the difference between the forging load of and the forging load of the simulated preform is less than a preset reference value, the non-axially symmetric preform is selected as the final preform.

According to the present invention, a preform having excellent performance can be derived by deriving a plurality of predicted shapes and selecting one with the best characteristics from these predicted shapes.

1 is a view for explaining the overall configuration of an apparatus for deriving the shape of a preform according to an embodiment of the present invention.
2 is a view for explaining the detailed configuration of the model generating unit of the apparatus for deriving the shape of the preform according to the embodiment of the present invention.
3 is a view for explaining the detailed configuration of a molded body extraction unit and a non-axial symmetry operation unit of the apparatus for deriving the shape of the preform according to an embodiment of the present invention.
4 is a view for explaining the configuration of a prediction model for predicting the shape of the preform according to an embodiment of the present invention.
5 is a view for explaining a filter of a prediction model for predicting the shape of a preform according to an embodiment of the present invention.
6 is a view for explaining an example of the operation of the prediction model for predicting the shape of the preform according to the embodiment of the present invention.
7 is a flowchart illustrating a method of generating learning data according to an embodiment of the present invention.
8 is a view for explaining a method of generating a monotone shape of learning data according to an embodiment of the present invention.
9 is a view for explaining a method of generating a candidate preform corresponding to the forged shape of the learning data according to an embodiment of the present invention.
10 to 12 are views for explaining a method of selecting a label preform corresponding to the forge shape of the learning data through finite element analysis according to an embodiment of the present invention.
13 is a flowchart illustrating a method of generating a plurality of predictive models (EM) according to an embodiment of the present invention.
14 is a view for explaining a method of generating a plurality of predictive models (EM) according to an embodiment of the present invention.
15 is a flowchart illustrating a method of generating a predictive model (EM) according to an embodiment of the present invention.
16 is a diagram for explaining a method of generating a predictive model (EM) according to an embodiment of the present invention.
17 is a flowchart for explaining a method for predicting a shape of a preform according to an embodiment of the present invention.
18 is a diagram for explaining a method of converting a voxel file into a mesh file according to an embodiment of the present invention.
19 is a diagram for explaining a method of detecting a predicted shape having a singular shape among predicted shapes according to an embodiment of the present invention.
20 is a diagram for explaining a method of correcting a singular shape of a predicted shape according to an embodiment of the present invention.
21 is a flowchart for explaining a method for predicting the shape of a non-axially symmetric preform according to an embodiment of the present invention.
22 is a view for explaining an example of a non-axially symmetric forging shape according to an embodiment of the present invention.
23 is a view for explaining a method of deriving an axisymmetric forged shape from a non-axially symmetric forged shape according to an embodiment of the present invention.
24 is a view for explaining a method of deriving a simulated preform from a plurality of preforms according to an embodiment of the present invention.
25 is a diagram illustrating a computing device according to an embodiment of the present invention.

Prior to the detailed description of the present invention, the configurations shown in the embodiments and drawings described in the present specification are only the most preferred embodiments of the present invention, and do not represent all the technical spirit of the present invention, so at the time of application It should be understood that there may be various equivalents and variations that may be substituted for them.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In this case, it should be noted that the same components in the accompanying drawings are denoted by the same reference numerals as much as possible. In addition, detailed descriptions of well-known functions and configurations that may obscure the gist of the present invention will be omitted. For the same reason, some components are exaggerated, omitted, or schematically illustrated in the accompanying drawings, and the size of each component does not fully reflect the actual size.

In particular, the terms or words used in the present specification and claims described below should not be construed as being limited to conventional or dictionary meanings, and the inventor of the term in order to best describe his invention. Based on the principle that it can be appropriately defined as In particular, in an embodiment of the present invention, prediction means calculating a value by calculation of a weight learned of a prediction model based on an artificial neural network.

First, an apparatus for deriving a shape of a preform according to an embodiment of the present invention will be described. 1 is a view for explaining the overall configuration of an apparatus for deriving the shape of a preform according to an embodiment of the present invention. 2 is a view for explaining the detailed configuration of the model generating unit of the apparatus for deriving the shape of the preform according to the embodiment of the present invention. 3 is a view for explaining the detailed configuration of a molded body extraction unit and a non-axial symmetry operation unit of the apparatus for deriving the shape of the preform according to an embodiment of the present invention.

Referring to FIG. 1 , an apparatus for deriving the shape of a preform according to an embodiment of the present invention (10, hereinafter abbreviated as 'shape deriving device') is a model generating unit 100 and a molded body extracting unit 200. and a non-axisymmetric operation unit 300 .

The model generator 100 is for generating training data and generating a plurality of estimation models (EMs) according to an embodiment of the present invention through deep learning using the generated training data. Here, the prediction model (EM) is an artificial neural network, and may be, for example, a convolution neural network (CNN).

Referring to FIG. 2 , the model generating unit 100 includes a monotonic shape generating unit 110 , a label setting unit 120 , and a learning unit 130 .

The forge shape generating unit 100 is for generating a forge shape from the learning data according to an embodiment of the present invention. This forged shape is an axisymmetric rotating body, and is generated as a three-dimensional image.

The label setting unit 120 generates a plurality of candidate preforms corresponding to the forging shape generated by the forging shape generating unit 100, and selects any one of the plurality of candidate preforms as a label preform corresponding to the forging shape. do. These label preforms are also axisymmetric rotating bodies and are created as three-dimensional images.

The learning unit 130 uses the forging shape generated by the forging shape generating unit 100 and the label preform generated by the label setting unit 120 corresponding to the forging shape as training data to obtain a predictive model (EM). create At this time, when the forging shape is input, the learning unit 130 learns the prediction model EM to predict the shape of the preform corresponding to the input forging shape. The learning unit 130 may generate a plurality of predictive models EM using a training data group obtained by grouping a plurality of different training data.

Again, referring to FIG. 1 , the molded body extracting unit 200 receives a plurality of predictive models EM from the model generator 100 , and predicts a preform corresponding to the forged shape using the plurality of predictive models. it is to do

Referring to FIG. 3 , the molded body extracting unit 200 includes a pre-processing unit 210 , a shape predicting unit 220 , a post-processing unit 230 , and a shape determining unit 240 .

The pre-processing unit 210 is for converting a monotone shape, which is a three-dimensional image, into a monotone shape matrix, which is a three-dimensional matrix expression. To this end, the preprocessor 210 voxelizes a monotone shape that is a 3D image. In this case, the preprocessor 210 may map the longest side of the monotonic shape that is the mesh file to 1 voxel, and assign a value of 0 or 1 to the corresponding voxel according to the volume occupied by the voxel to any one of the mesh files.

The shape prediction unit 220 is for deriving a plurality of predicted shape matrices that are matrix representations of the shape of the preform for the forged shape. To this end, the shape prediction unit 220 uses a plurality of prediction models (EM) learned to predict the shape of the preform.

The post-processing unit 230 transforms the plurality of predicted shape matrices into a plurality of predicted shapes that are 3D images. At this time, the post-processing unit 230 converts the plurality of predicted shape matrices that are voxel files to the plurality of predicted shapes that are mesh files through the marching cube algorithm, and smoothes the surfaces of the plurality of predicted shapes that are mesh files using a smoothing algorithm. can be performed.

The shape determining unit 240 selects one of the plurality of predicted shapes that most closely meets the preset design conditions as the preform. To this end, the shape determining unit 240 excludes or corrects the predicted shape having a singular shape among the plurality of predicted shapes, and performs finite element analysis on the plurality of predicted shapes to determine the mold filling rate, the presence of overlapping defects, and the forging load. derive Then, the shape determining unit 240 specifies a predicted shape having an unfilled defect according to the mold filling rate among the plurality of predicted shapes, and excludes the specified predicted shape. In addition, the shape determining unit 240 excludes the predicted shape having an overlap defect from among the plurality of predicted shapes. Then, the shape determining unit 240 selects a predicted shape having the lowest forging load among a plurality of predicted shapes without unfilled defects and overlapping defects as a preform corresponding to the forging shape.

Referring back to FIG. 1 , the non-axisymmetric operation unit 300 is for predicting a preform having a non-axisymmetric forging shape. That is, the predictive model EM is generated by using an axisymmetric monotonic shape as training data. However, the non-axisymmetric calculation unit 300 uses the preform derivation unit 200 including the prediction model (EM) to derive the preform most suitable for the non-axial forging shape even for the non-axial forging shape. Can be derived.

Referring back to FIG. 3 , the non-axial symmetry operation unit 300 includes a shape processing unit 310 and a shape selection unit 320 .

When a non-axially symmetric forged shape is input, the shape processing unit 310 generates a plurality of different axisymmetric forged shapes from the input non-axially symmetric forged shape. In addition, the shape processing unit 310 provides a non-axially symmetric forged shape and a plurality of different axisymmetric forged shapes to the preform guide unit 200 . Then, the preform derivation unit 200 uses a plurality of predictive models (EM) learned to predict the preform from the axisymmetric forging shape to derive the non-axially symmetric preform from the non-axially symmetric forging shape, and a plurality of predictive models (EM) is used to derive a plurality of axisymmetric preforms from a plurality of axisymmetric forging shapes.

The shape selection unit 320 generates a simulated preform by combining a plurality of axisymmetric preforms, and selects the non-axially symmetric preform as a final preform through verification using the simulated preform. Accordingly, the preform corresponding to the non-axially symmetric forging shape can be predicted.

Next, the configuration of the predictive model (EM) according to an embodiment of the present invention will be described. 4 is a view for explaining the configuration of a prediction model for predicting the shape of the preform according to an embodiment of the present invention. 5 is a view for explaining a filter of a prediction model for predicting the shape of a preform according to an embodiment of the present invention. 6 is a view for explaining an example of the operation of the prediction model for predicting the shape of the preform according to the embodiment of the present invention.

The prediction model (EM) according to an embodiment of the present invention includes a plurality of layers, and the plurality of layers includes a plurality of convolution layers, a plurality of pooling layers, a plurality of upconvolution layers ( up-convolution layer) and a plurality of concatenate layers. These plurality of layers generate a feature map through operation.

The convolution layer performs a convolution operation and an operation by an activation function. The activation function may be exemplified by Sigmoid, Hyperbolic tangent (tanh), Exponential Linear Unit (ELU), Rectified Linear Unit (ReLU), Leakly ReLU, Maxout, Minout, Softmax, and the like. The pooling layer PL is for down-sampling and performs a max pooling operation. An up-convolution layer is for up-sampling and performs an up-convolution operation. The concatenate layer performs an operation that combines the feature map of the first half with the feature map of the second half.

As described above, the predictive model EM includes a plurality of layers, and the plurality of layers includes a plurality of operations. In addition, a plurality of layers are connected by a weight (w: weight). The calculation result of one layer is weighted and input to the next layer. That is, one layer of the predictive model (EM) receives a weighted value from the previous layer, performs an operation on it, and transmits the operation result to the input of the next layer.

For example, referring to FIG. 5 , in the convolution operation, a weight is applied through a filter matrix (or kernel, W). The filter matrix W is a three-dimensional matrix, and each element of the three-dimensional matrix becomes a weight. The filter matrix W moves in a sliding manner within an image matrix (or feature map, img_M) that is a matrix representation of a three-dimensional image having a predetermined height (img_H), width (img_W), and depth (img_D). A weight may be applied to the image matrix img_M by performing an operation between the elements of the image matrix img_M and the elements of the filter matrix W, ie, a convolution operation.

In particular, when performing a convolution operation according to an embodiment of the present invention, the convolution operation may be performed after padding is added to the image matrix img_M. For example, referring to FIG. 6 , the element value is 0 in all of the height (img_H), width (img_W), and depth (img_D) of one image matrix (img_M) having a size of 128 * 128 * 128 (zero padding) is added to generate one image matrix (img_M) with a size of 130*130*130. Then, a convolution operation is performed on the zero-padded image matrix (img_M) through 32 filter matrices (W) of size 6*6*6, and an activation function (f) is performed on the result of the convolution operation. Thus, it is possible to create a convolutional layer having 32 image matrices (or feature maps) of size 128*128*128.

That is, the above-described operation is performed according to the following Equation (1).

Here, X is a zero-padded image matrix (img_M), and W is a filter matrix. b is a threshold, and is used to remove an operation result that is less than the threshold. f stands for the activation function. Since the activation function (f) differentiates the error function with weights in the optimization process except for the final output, the ReLU function, which always has a differential value in the positive region, is used. For the last layer, the sigmoid function is used because the voxel value must be between 0 and 1 in order to predict the candidate shape (preform) in the last layer for the final output.

Next, a method for generating learning data according to an embodiment of the present invention will be described. 7 is a flowchart illustrating a method of generating learning data according to an embodiment of the present invention. 8 is a view for explaining a method of generating a monotone shape of learning data according to an embodiment of the present invention. 9 is a view for explaining a method of generating a candidate preform corresponding to the forged shape of the learning data according to an embodiment of the present invention. 10 to 12 are views for explaining a method of selecting a label preform corresponding to the forge shape of the learning data through finite element analysis according to an embodiment of the present invention.

Referring to FIG. 7 , the monotonic shape generating unit 110 prepares a plurality of monotonic shapes for learning the predictive model EM in step S110 . An example of a forged shape is shown in FIG. 8 . As shown, the forged shape has a shape of axis symmetry with respect to the rotation axis (X) perpendicular to the ground. As shown in (A) of FIG. 8 , the forging shape generating unit 110 first has a preset radius D/ 2) Set the busbar (G) at the spaced position. Then, the forging shape generating unit 110 randomly allocates control points C1, C2, C3, and C4 along the bottom line B between the rotation axis X and the bus bar G. Next, the forging shape generating unit 110 combines a plurality of preset height levels (eg, 0.24D, 0.36D, 0.48D, 0.6D, ..., where D is a diameter) to combine the rotation axis X ), the bus bar G, and the vertical heights H0, H1, H2, H3, H4, H5 from the bottom line B of the control points C1, C2, C3, C4. Then, the forging shape generating unit 110 sets the heights H0, H1, H2, H3, H4, H5 of the rotation axis X, the bus bar G, and the control points C1, C2, C3, C4 to B- Connect according to the spline technique to create a connecting line (L). Then, the forging shape generating unit 110 rotates the rotation shaft (X), the connecting line (L), the bus bar (G), and the bottom line (B) along the rotation axis (X), as shown in (B) of FIG. Similarly, forged shapes can be created. In particular, the forging shape generating unit 110 may generate a plurality of forging shapes by applying a combination of a plurality of height levels to the aforementioned heights H0, H1, H2, H3, H4, and H5.

Next, the label setting unit 120 generates a plurality of candidate preforms corresponding to each of the plurality of forged shapes previously generated in step S120 . At this time, the label setting unit 120 generates a candidate preform as follows in order to eliminate the user's bias. Referring to (A) of Figure 9, the label setting unit 120 first, along the reference line (S) perpendicular to the axis of rotation (X) in the vertical axis of rotation (X) and the same radius as the forging shape (D/2) ), set the busbar (G) at the spaced position. Then, the label setting unit 120 divides the distance between the rotation axis (X) and the bus bar (G) into a plurality along the reference line (S), and assigns the control points (C1, C2, C3) to the divided positions. Then, the label setting unit 120 combines a plurality of preset height levels (eg, 0.3D, 0.4D, 0.5D, where D is a diameter) to combine the axis of rotation (X), the busbar (G), and the control point (C1, C1, Heights H1, H2, H3, H4, and H5 in the vertical direction from the reference line S of C2 and C3 are set. Then, the label setting unit 120 connects the rotation axis (X), the bus bar (G) and the control points (C1, C2, C3) with the heights (H1, H2, H3, H4, H5) set according to the B-spline technique. to create a connecting line (L). Next, the label setting unit 120 selects any one of a plurality of preset height levels according to a combination and applies the height H6 between the bottom line B and the reference line S. Then, the label setting unit 120 may generate a candidate preform by rotating the axis of rotation (X), the connecting line (L), the bus bar (G), and the bottom line (B) along the axis of rotation (X). In particular, as shown in (B) of FIG. 6 , the label setting unit 120 applies a combination of a plurality of height levels to the above-described heights H1, H2, H3, H4, H5, H6, whereby a plurality of candidates Preforms can be produced. According to another embodiment, the label setting unit 120 does not select any one of a plurality of preset height levels, and the rotation axis (X), the connection line (L), the bus bar (G), and the bottom line (B) to the rotation of the The height H6 between the bottom line B and the reference line S may be set so that the volume of the candidate preform generated by the method coincides with the volume of the forging shape corresponding to the candidate preform.

Next, the label setting unit 120 performs finite element analysis on the plurality of preform candidates generated in step S130 to derive the mold filling rate, the presence of overlapping defects, and the forging load.

At this time, the label setting unit 120 models the preform candidate pf through the finite element analysis model as shown in FIG. 10 , and the preform modeled between the upper mold TD and the lower mold BD. The candidate (pf) is placed and forging is performed to perform finite element analysis to analyze the mold filling rate, the presence of overlapping parts, and the results of the forging load of the modeled preform candidate (pf). According to this finite element analysis, the label setting unit 120 may derive the mold filling rate, the presence of overlapping defects, and the forging load of the plurality of preform candidates.

Then, the label setting unit 120 selects one of a plurality of preform candidates as a preform which is a label corresponding to the forging shape according to the result of the finite element analysis in step S140.

First, the label setting unit 120 excludes a preform candidate having an unfilled defect according to the calculated mold filling rate among a plurality of preform candidates. For example, as shown in FIG. 11, in the case of preform candidate No. 6 (#6), as can be seen from ① of the graph, even when the pressure is increased to the limit value, the mold is not completely filled, so it is excluded.

In addition, the label setting unit 120 excludes a preform candidate having an overlap defect from among the plurality of preform candidates. For example, as shown in FIG. 12(A), the label setting unit 120 models the nodes of the outer part of the candidate preform, and performs numerical analysis as shown in FIG. 12(B). When the overlapping portion OL occurs, it is determined that there is an overlap defect, and the candidate for the preform may be excluded.

In addition, the label setting unit 120 selects a preform candidate having the lowest forging load among a plurality of preform candidates without the aforementioned unfilled defects and overlapping defects as a label preform corresponding to the forging shape. For example, in the case of preform candidate No. 6 (#6) among preform candidates No. 6 (#6), No. 135 (#135), and No. 126 (#126) of FIG. 11, it is excluded due to unfilled defects, and the preliminary As can be seen by comparing ② and ③ in the graph among the green body candidates No. 135 (#135) and No. 126 (#126), the preform candidate No. 126 (#126) with a lower forging load was selected as the label preform. do.

As described above, when the forged shape and the label preform corresponding thereto are selected, the label setting unit 120 maps the forged shape and the label preform corresponding thereto in step S150 and stores it as learning data.

Then, a method of generating a plurality of predictive models EM through deep learning using the above-described learning data will be described. 13 is a flowchart illustrating a method of generating a plurality of predictive models (EM) according to an embodiment of the present invention. 14 is a view for explaining a method of generating a plurality of predictive models (EM) according to an embodiment of the present invention.

The learning unit 130 classifies the plurality of monotonic shapes generated earlier ( S110 ) in step S210 into a plurality of learning data groups based on a combination of height differences between a plurality of feature points of the monotone shape.

According to an embodiment, the learning unit 130 represents a combination of the difference in heights H1, H2, H3 between the feature points C1, C2, and C3 of a predetermined number (eg, 3) in an order close to the rotation axis X. A first condition and a second condition representing a combination of a difference in heights H2, H3, H4 between a predetermined number (for example, 3) of feature points C4, C3, C2 in an order close to the bus G are provided. And the learning unit 130 classifies the plurality of monotone shapes into a plurality of learning data groups according to whether the first condition or the second condition is satisfied. A forging shape according to a combination of various height differences is shown in FIG. 14 . Since each of these monotone shapes belongs to a corresponding learning data group when any one of the first and second conditions is satisfied, any one monotone shape may belong to a plurality of learning data groups. For example, the learning data group may be classified according to the conditions shown in Table 1 below.

training data group first condition second condition 1st group H1 < H3 < H2 H2 < H4 < H3 2nd group H3 < H1 < H2 H4 < H2 < H3 3rd group H2 < H1 < H3 H3 < H2 < H4 4th group H2 < H3 < H1 H3 < H4 < H2 5th group H3 < H2 < H1 H4 < H3 < H2 6th group H1 < H2 < H3 H2 < H3 < H4

Next, the learning unit 130 generates a plurality of predictive models (EM) by performing learning using the training data belonging to the corresponding training data group for each training data group in step S220. That is, one predictive model (EM) is trained using a plurality of learning data belonging to any one training data group. Then, a method of generating any one predictive model (EM) through learning will be described in more detail. 15 is a flowchart illustrating a method of generating a predictive model (EM) according to an embodiment of the present invention. 16 is a diagram for explaining a method of generating a predictive model (EM) according to an embodiment of the present invention.

Referring to FIG. 15 , the learning unit 130 initializes the prediction model EM in step S310 . This initialization means initializing a parameter of the prediction model EM, that is, the weight w. That is, the learner 130 may initialize all elements of the filter matrix W.

When the initialization is completed, the learning unit 130 performs pre-processing on the corresponding learning data in step S320 to form a monotone shape in a three-dimensional image format and a label preform corresponding to the monotone shape in a matrix format monotone shape matrix and label Convert to preform matrix. A more detailed description of this is as follows.

The forged shape generated above (S110, S120) and the label preform corresponding to the forged shape are three-dimensional images, which are mesh files (eg, stl format files). First, the learning unit 130 converts the three-dimensional image of the mesh file format into a monotonic shape matrix and a label preform matrix by voxelization. In this case, the learning unit 130 performs voxelization according to a preset space dimension of all voxel files. In this case, the learner 130 maps the longest side of the mesh file to 1 voxel. For example, as shown in FIG. 16 , 0.0164D (here, D is a diameter) is mapped to 1 voxel in the mesh file. At this time, the learning unit 130 sets the voxel value of the mesh file to 1 if the volume occupied by the voxel in any one of the mesh files is 0.5 or more, and sets the voxel value to 0 if it is less than 0.5. After voxelization is completed, the learner 130 adds a predetermined number of voxels having a value of 0 to all surfaces of the 3D voxel file as a buffer. In this way, the voxelized voxel file becomes a matrix (the monotonic shape matrix and the label preform matrix), one voxel of the voxel file becomes each element of the matrix, and the voxel value of each voxel becomes the element value of the matrix.

After the preprocessing, the learning unit 130 inputs the monotonic shape matrix to the prediction model EM in step S330.

Then, the predictive model EM derives a predicted shape matrix, which is a matrix expression that predicts the shape of the preform through a plurality of operations to which the weights for which the learning between the plurality of layers are not completed in step S330 are applied.

Next, the learning unit 130 calculates a loss representing the difference between the predicted shape matrix and the label preform matrix through the loss function in step S340 . According to an embodiment, the learning unit 130 may obtain a loss through a binary cross-entropy function, which is a loss function as in Equation 2 below.

는 예측형상행렬과 레이블 예비성형체행렬과의 차이를 나타내는 손실을 나타낸다. 또한,

는 입력된 단조형상행렬에 대응하는 레이블인 레이블 예비성형체행렬이고,

는 입력된 단조형상행렬로부터 도출된 예측형상행렬을 의미한다. here,

is the loss representing the difference between the predicted shape matrix and the label preform matrix. In addition,

is a label preform matrix, which is a label corresponding to the input forging shape matrix,

denotes a predicted shape matrix derived from the input monotonic shape matrix.

Next, the learning unit 130 performs optimization of correcting the weight of the prediction model EM so that the loss derived through the loss function is minimized in step S360. That is, the learning unit 130 optimizes the weight of the prediction model EM so that the loss derived through Equation 2, that is, the binary cross-entropy loss, is minimized.

Steps S330 to S360 described above are repeatedly performed using a plurality of different training data, that is, a plurality of monotonic shape matrices and a label preform matrix belonging to one training data group, and according to this iteration, the predictive model (EM) The weight of is updated iteratively.

And this repetition is made until the loss is less than or equal to a preset target value when the loss is calculated using the monotonic shape matrix and the label preform matrix, which are training data for verification belonging to the same training data group. Therefore, the learning unit 130 determines whether the loss calculated by using the learning data prepared for verification in step S370 is less than or equal to a preset target value, and if the loss is less than or equal to the preset target value, learning is completed in step S380.

Each of the plurality of predictive models EM is trained by the method as described above. In this way, when the learning of the plurality of predictive models EM is completed, the learning unit 130 provides the learned plurality of predictive models EM to the preform extracting unit 200 .

Then, the preform derivation unit 200 derives the preform most suitable for the forging shape using a plurality of predictive models EM. This method will be described in more detail. 17 is a flowchart for explaining a method for predicting a shape of a preform according to an embodiment of the present invention. 18 is a diagram for explaining a method of converting a voxel file into a mesh file according to an embodiment of the present invention. 19 is a diagram for explaining a method of detecting a predicted shape having a singular shape among predicted shapes according to an embodiment of the present invention. 20 is a diagram for explaining a method of correcting a singular shape of a predicted shape according to an embodiment of the present invention.

Referring to FIG. 17 , the forging shape for predicting the shape of the preform is input to the preform extracting unit 200 in step S410 .

Then, the pre-processing unit 210 converts the monotone shape, which is a three-dimensional image, into a monotone shape matrix through pre-processing in step S420 . To this end, the preprocessor 210 first voxelizes the monotone shape of the mesh file format to generate the monotone shape of the voxel file format. In this case, the preprocessor 210 performs voxelization according to a preset dimension of the entire voxel file. For example, the size may be determined to be 122*122*122. In this case, the preprocessor 210 maps the longest side of the mesh file to a scale of 1 voxel. For example, as shown in FIG. 16 , 0.0164D (here, D is a diameter) may be mapped to 1 voxel in the mesh file (0.0164D=1voxel). In particular, the preprocessor 210 sets the voxel value of the mesh file to 1 when the volume occupied by a voxel in any one of the mesh files is 0.5 or more, and sets the voxel value to 0 when it is less than 0.5. When the voxel value is allocated, the preprocessor 210 adds a predetermined number of voxels having a value of 0 to all surfaces of the 3D voxel file as a buffer. For example, a voxel file having a size of 128*128*128 is generated by adding three voxels each having a value of 0 to six sides of a basic dimension of 122*122*122. In this way, the voxelized voxel file becomes the monotonic shape matrix, one voxel of the voxel file becomes each element of the monotonic shape matrix, and the voxel value (0 or 1) of each voxel becomes the element value of the monotonic shape matrix. .

The shape prediction unit 220 receives the monotone shape matrix from the preprocessor 210 and inputs the monotone shape matrix input in step S430 to each of the plurality of prediction models EM.

Then, the plurality of predictive models EM derive a plurality of predictive shape matrices that are matrix representations of the predicted shape of the preform by performing a plurality of operations to which a plurality of inter-layer weights are applied in step S440 . For example, when it is assumed that six prediction models EM exist, each of the six prediction models EM may derive a prediction shape matrix to derive a total of 6 prediction shape matrices.

Next, the post-processing unit 230 converts the plurality of predicted shape matrices into a plurality of predicted shapes that are 3D images in a mesh file format through post-processing of the plurality of predicted shape matrices in step S450 . This is done in reverse of step S420. To this end, the post-processing unit 230 removes the buffer from the predictive shape matrix of the voxel file format. For example, from a voxel file having a size of 128*128*128, three buffer voxels are removed from each of the six upper, lower, left, and right front and rear sides, and converted into a 122*122*122 voxel file. Then, the post-processing unit 230 converts the voxel file into a mesh file through a marching cube algorithm. At this time, as shown in FIG. 18, the scale is matched according to the scale (0.0164D=1voxel) used in the preprocessing. Then, the post-processing unit 230 may perform a smoothing process on the surface of the mesh file using a smoothing algorithm, for example, Humphrey's smoothing algorithm.

Next, the shape determining unit 240 selects any one predicted shape that most meets the set condition among the plurality of predicted shapes as the preform in step S460 .

To this end, the shape determining unit 240 first excludes a predicted shape having a singular shape, such as a hole, among a plurality of predicted shapes.

According to an embodiment, as shown in FIG. 19 , the shape determining unit 240 checks the cross-section at a predetermined interval for the predicted shape, and if the difference (dif) between the areas in the neighboring cross-sections is greater than or equal to a preset threshold, , it is judged to be a singular form such as a hole. For example, in the case of (A) of FIG. 19, the cross section of part A-A is shown in the predicted shape, and in the case of FIG. 19(B), the cross section of part B-B is shown in the same predicted shape. The shape determining unit 240 measures the difference (dif) in the area of the cross-section, and if the difference (diff) is greater than or equal to a preset threshold value, the shape determination unit 240 specifies that the portion is a singular shape (S) such as a hole (hole).

However, as a result of cross-sectional confirmation of the predicted shape, the shape determining unit 240 is specified as a singular shape (S) such as a hole as shown in FIG. As shown in (B) of (B), the corresponding part (N) can be corrected by filling the image representing the predicted shape.

As described above, after correcting or excluding the predicted shape having the singular shape (S), the shape determining unit 240 performs finite element analysis on the plurality of remaining predicted shapes to determine the mold filling rate, the presence of overlapping defects, and Derive the forging load. Then, the shape determining unit 240 selects one of the plurality of predicted shapes as a preform corresponding to the forging shape according to the result of the finite element analysis. This is done in the same way as the method of selecting one label preform from a plurality of preform candidates. Again, referring to FIGS. 10 to 12 , the shape determining unit 240 models the predicted shape through the finite element analysis model, and forging the modeled predicted shape between the upper mold TD and the lower mold BD. , perform finite element analysis to analyze the mold filling rate of the modeled predicted shape, the presence of overlapping parts, and the results of forging load. According to the finite element analysis, the shape determining unit 240 may derive the mold filling rate, the presence of overlapping defects, and the forging load of the plurality of predicted shapes.

First, the shape determining unit 240 specifies a predicted shape having an unfilled defect according to the calculated mold filling rate among a plurality of predicted shapes, and excludes the specified predicted shape. For example, raising the pressure to a limit is excluded because the mold is not completely filled.

Also, the shape determining unit 240 excludes the predicted shape having an overlap defect from among the plurality of predicted shapes. For example, the shape determining unit 240 models the nodes of the outer part of the predicted shape, and when an overlapping part OL occurs by performing numerical analysis, it is determined that there is an overlap defect, and the predicted shape may be excluded. .

In addition, the shape determining unit 240 selects a predicted shape having the lowest forging load among a plurality of predicted shapes without the aforementioned unfilled defects and overlapping defects as a preform corresponding to the forging shape.

The plurality of prediction models (EM) described above are generated by using an axisymmetric monotonic shape as training data. According to the embodiment of the present invention, the preform extracting unit 200 uses a plurality of predictive models (EM) learned by using such an axisymmetric forging shape as learning data, even for a non-axially symmetric forging shape. It is possible to derive the most suitable preform for This method will be described in more detail. 21 is a flowchart illustrating a method for predicting a shape of a non-axially symmetric preform according to an embodiment of the present invention. 22 is a view for explaining an example of a non-axially symmetric forging shape according to an embodiment of the present invention. 23 is a view for explaining a method of deriving an axisymmetric forged shape from a non-axially symmetric forged shape according to an embodiment of the present invention. 24 is a view for explaining a method of deriving a simulated preform from a plurality of preforms according to an embodiment of the present invention.

Referring to FIG. 21 , the shape processing unit 310 of the non-axisymmetric calculation unit 300 prepares a plurality of axisymmetric forged shapes from the non-axial forged shape in step S520 when the non-axially symmetric forged shape is input in step S510. A method of preparing a plurality of axisymmetric forging shapes in step S520 will be described in more detail with reference to FIGS. 22 and 23 . 22 shows an example of a non-axially symmetric forged shape. Fig. 22 (A) is a perspective view of a non-axially symmetric forging shape, and Fig. 22 (B) is a plan view of a non-axially symmetric forging shape. As shown, the forged shape is non-axially symmetric, but has a rotation axis (X) penetrating the center of gravity vertically.

Referring to FIGS. 22 and 23 , the shape processing unit 310 first recognizes the rotational axis X vertically penetrating the center of gravity of the non-axially symmetric forged shape. Then, the shape processing unit 310, as shown in (B), (C) and (D) of FIG. 23 (B), (C) and (D) from the non-axial forging shape shown in FIG. A plurality of longitudinal cross-sections, which are cross-sections cut in a plane including the axis of rotation along different rotation angles, for example, 0 degrees, 45 degrees, and 90 degrees respectively, are generated as a reference. Then, the shape processing unit 310 provides a plurality of different forging shapes by rotating a plurality of longitudinal cross-sections as shown in FIGS. 23 (B), (C) and (D) based on the rotation axis (X).

The shape processing unit 310 sequentially provides the non-axially symmetric forged shape and the plurality of axisymmetric forged shapes generated from the non-axially symmetric forged shape to the preform guide 200 .

Accordingly, the preform derivation unit 200 derives a non-axially symmetric preform corresponding to the non-axially symmetric forged shape from the non-axially symmetric forged shape by using a plurality of predictive models (EM) in step S530 . This step S530 is performed in the same manner as the method described with reference to FIGS. 17, 18, 19, and 20 above. The step S530 will be briefly described as follows. First, the preprocessor 210 converts a non-axial monotonic shape, which is a three-dimensional image in a mesh file format, into a non-axial monotonic shape matrix through pre-processing, for example, voxelization. The shape prediction unit 220 inputs the non-axially symmetric monotonic shape matrix to each of the plurality of prediction models EM. Then, the plurality of predictive models EM derive a plurality of predictive shape matrices that are matrix representations of the predicted shapes of the non-axially symmetric preform by performing a plurality of calculations to which the weights between the plurality of layers are applied. For example, when it is assumed that six prediction models EM exist, each of the six prediction models EM may derive a prediction shape matrix to derive a total of 6 prediction shape matrices. Next, the post-processing unit 230 converts the plurality of predicted shape matrices into a plurality of predicted shapes that are 3D images in a mesh file format through post-processing of the plurality of predicted shape matrices. Then, the shape determining unit 240 first excludes or corrects a predicted shape having a singular shape such as a hole among a plurality of predicted shapes. Then, the shape determining unit 240 performs finite element analysis on the plurality of predicted shapes that are not excluded to detect the presence of unfilled defects and overlapping defects, and derives a forging load. Then, the shape determining unit 240 excludes the predicted shape having the unfilled defect according to the calculated mold filling rate among the plurality of predicted shapes. In addition, the shape determining unit 240 excludes the predicted shape having an overlap defect from among the plurality of predicted shapes. Next, the shape determining unit 240 selects a predicted shape having the lowest forging load among a plurality of predicted shapes without the aforementioned unfilled defects and overlapping defects as a non-axially symmetric preform corresponding to the non-axially symmetric forging shape.

Next, the preform derivation unit 200 uses a plurality of predictive models (EM) in step S540 to derive a plurality of axisymmetric preforms corresponding to each of the plurality of axisymmetric forging shapes from the plurality of axisymmetric forging shapes. .

Step S540 is also performed in the same manner as the method described with reference to FIGS. 17, 18, 19, and 20 above. Step S540 is sequentially performed for each of the plurality of axisymmetric forging shapes, and is performed as follows for any one axisymmetric forging shape among the plurality of axisymmetric forging shapes.

First, the preprocessor 210 converts an axisymmetric monotonic shape, which is a three-dimensional image in a mesh file format, into an axisymmetric monotonic shape matrix through preprocessing, for example, voxelization. The shape prediction unit 220 inputs the axisymmetric monotonic shape matrix to each of the plurality of prediction models EM. Then, the plurality of predictive models EM derive a plurality of predictive shape matrices that are matrix representations of the predicted shape of the preform by performing a plurality of operations to which the weights between the plurality of layers are applied. For example, when it is assumed that six prediction models EM exist, each of the six prediction models EM may derive a prediction shape matrix to derive a total of 6 prediction shape matrices. Next, the post-processing unit 230 converts the plurality of predicted shape matrices into a plurality of predicted shapes that are 3D images in a mesh file format through post-processing of the plurality of predicted shape matrices. Then, the shape determining unit 240 first excludes or corrects a predicted shape having a singular shape such as a hole among a plurality of predicted shapes. Then, the shape determining unit 240 performs finite element analysis on the plurality of remaining predicted shapes to derive the mold filling rate, the existence of overlapping defects, and the forging load. Then, the predicted shape having unfilled defects is excluded according to the calculated mold filling rate among the plurality of predicted shapes. In addition, the shape determining unit 240 excludes the predicted shape having an overlap defect from among the plurality of predicted shapes. Next, the shape determining unit 240 selects the predicted shape having the lowest forging load among the plurality of predicted shapes without the aforementioned unfilled defects and overlapping defects as an axisymmetric preform corresponding to the axisymmetric forging shape.

An example of the plurality of axisymmetric preforms derived in step S540 is shown in (A) of FIG. 24 . As shown, different axisymmetric preforms were predicted depending on the rotation angle at which the longitudinal section was created.

The preform extracting unit 200 outputs the non-axially symmetric preform and the plurality of axisymmetric preforms to the shape selection unit 320 of the non-axially symmetrical operation unit 300 .

The shape selection unit 320 generates a simulated preform by combining a plurality of axisymmetric preforms in step S550 . At this time, the shape selection unit 320 selects a part of the axisymmetric preform according to the rotation angle applied when generating the longitudinal section according to the rotation angle in order to derive the axisymmetric forged shape of the corresponding axisymmetric preform from the non-axially symmetric forged shape. Extraction and combining the extracted parts to produce a simulated preform.

For example, the first axisymmetric preform P1, the second axisymmetric preform P2, and the third axisymmetric preform P3 of FIG. It is assumed that it is derived from an axisymmetric forging shape created from a longitudinal section, which is a section cut in a plane including the axis of rotation along 0 degrees, 45 degrees, and 90 degrees, respectively, based on . Accordingly, the shape selection unit 320 includes a 0 degree to 45 degree part and a 180 degree to 225 degree part based on the rotation axis X of the first axisymmetric preform P1, and a second axisymmetric preform ( 45 degree to 90 degree part and 225 degree to 270 degree part with respect to the axis of rotation X of P2), and a 90 degree to 180 degree part with respect to the axis of rotation X of the third axisymmetric preform P3 and 270 to 360 degree portions can be combined to create a simulated preform. An example of the resulting simulated preform is shown in FIG. 24(C). 24(C) shows a 1/4 longitudinal section of the simulated preform.

Next, the shape selection unit 320 calculates the forging load of the simulated preform through finite element analysis in step S560. This is done in the same way as the post-processing unit 230 performs.

Next, the shape selection unit 320 selects the final preform by comparing the forging load of the non-axially symmetric preform and the simulated preform in step S570. Here, the forging load of the non-axially symmetric preform is the forging load calculated when deriving the non-axially symmetric preform corresponding to the non-axially symmetric forged shape from the non-axially symmetric forged shape in step S530 described above.

At this time, the shape selection unit 320 compares the forging load of the non-axially symmetric preform and the simulated preform, and the difference between the forging load of the non-axially symmetric preform and the simulated preform is less than a preset reference value. Select the preform.

The shape selection unit 320 compares the forging loads of the non-axially symmetric preform and the simulated preform, and if the difference between the forging loads of the non-axially symmetric preform and the simulated preform is greater than or equal to a preset reference value, an error is returned.

25 is a diagram illustrating a computing device according to an embodiment of the present invention. The computing device TN100 of FIG. 25 is the device described herein, for example, the shape extraction device 10, or the model generation unit 100, the molded body extraction unit 200, and the non-axial symmetry operation unit 300, respectively. can be

In the embodiment of FIG. 25 , the computing device TN100 may include at least one processor TN110 , a transceiver device TN120 , and a memory TN130 . Also, the computing device TN100 may further include a storage device TN140 , an input interface device TN150 , an output interface device TN160 , and the like. Components included in the computing device TN100 may be connected by a bus TN170 to communicate with each other.

The processor TN110 may execute a program command stored in at least one of the memory TN130 and the storage device TN140. The processor TN110 may mean a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor on which methods according to an embodiment of the present invention are performed. The processor TN110 may be configured to implement procedures, functions, and methods described in connection with an embodiment of the present invention. The processor TN110 may control each component of the computing device TN100 .

Each of the memory TN130 and the storage device TN140 may store various information related to the operation of the processor TN110 . Each of the memory TN130 and the storage device TN140 may be configured as at least one of a volatile storage medium and a nonvolatile storage medium. For example, the memory TN130 may include at least one of a read only memory (ROM) and a random access memory (RAM).

The transceiver TN120 may transmit or receive a wired signal or a wireless signal. The transceiver TN120 may be connected to a network to perform communication.

Meanwhile, the above-described method according to an embodiment of the present invention may be implemented in the form of a program readable by various computer means and recorded in a computer readable recording medium. Here, the recording medium may include a program command, a data file, a data structure, etc. alone or in combination. The program instructions recorded on the recording medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software. For example, the recording medium includes magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floppy disks ( magneto-optical media) and hardware devices specially configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions may include not only machine language wires such as those generated by a compiler, but also high-level language wires that can be executed by a computer using an interpreter or the like. Such hardware devices may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

Although the present invention has been described above using several preferred embodiments, these examples are illustrative and not restrictive. As such, those of ordinary skill in the art to which the present invention pertains will understand that various changes and modifications can be made in accordance with the doctrine of equivalents without departing from the spirit of the present invention and the scope of rights set forth in the appended claims.

10: shape extraction device 100: model generation unit
110: forging shape generating unit 120: label setting unit
130: learning unit 200: molded body extracting unit
210: pre-processing unit 220: shape prediction unit
230: post-processing unit 240: shape determining unit
300: non-axial symmetry operation unit 310: shape processing unit
320: shape selection unit

Claims (15)

A method for predicting the shape of a preform, the method comprising:
converting, by a pre-processing unit, a monotone shape, which is a three-dimensional image, into a monotone shape matrix, which is a three-dimensional matrix representation;
deriving a plurality of predicted shape matrices that are matrix representations of the shape of the preform with respect to the forged shape through the plurality of prediction models trained to predict the shape of the preform by the shape prediction unit;
converting, by a post-processing unit, a plurality of predicted shape matrices into a plurality of predicted shapes that are three-dimensional images;
selecting, by a shape determining unit, any one of a plurality of predicted shapes that most closely matches a preset design condition as a preform;
characterized by comprising
A method for predicting the shape of a preform. According to claim 1,
The step of selecting the preform is
excluding or correcting, by the shape determining unit, a predicted shape having a singular shape among the plurality of predicted shapes;
performing finite element analysis on the plurality of predicted shapes by the shape determining unit to derive a mold filling rate, an overlapping defect, and a forging load; and
selecting, by the shape determining unit, one of the plurality of predicted shapes as a preform corresponding to a forged shape according to a result of the finite element analysis;
characterized in that it comprises
A method for predicting the shape of a preform. According to claim 1,
The step of excluding or correcting the predicted shape having the singular shape is
The shape determining unit checks the cross section at a predetermined interval with respect to the predicted shape, and if the difference between the areas in the adjacent cross sections is equal to or greater than a preset threshold, specifying the corresponding part as a singular shape
A method for predicting the shape of a preform. 4. The method of claim 3,
The step of excluding or correcting the predicted shape having the singular shape is
If it is specified as a specific shape, but the cross section is blocked in all directions, the shape determining unit fills in the corresponding part with an image representing the predicted shape and corrects it
A method for predicting the shape of a preform. 3. The method of claim 2,
The step of selecting one of the plurality of predicted shapes as a preform corresponding to the forged shape comprises:
specifying, by the shape determining unit, a predicted shape having an unfilled defect according to a mold filling rate among the plurality of predicted shapes, and excluding the specified predicted shape;
excluding, by the shape determining unit, a predicted shape having an overlap defect from among the plurality of predicted shapes; and
selecting, by the shape determining unit, a predicted shape having the lowest forging load among a plurality of predicted shapes without the unfilled defect and the overlapping defect as a preform corresponding to the forging shape;
characterized by comprising
A method for predicting the shape of a preform. According to claim 1,
The step of deriving the plurality of predictive shape matrices is
characterized in that the plurality of prediction models derive a plurality of predictive shape matrices by performing a plurality of operations to which the weights learned between a plurality of layers are applied to the monotonic shape matrix
A method for predicting the shape of a preform. According to claim 1,
The step of converting the monotonic shape into a monotonic shape matrix, which is a three-dimensional matrix expression,
wherein the pre-processing unit maps the longest side to 1 voxel in a monotonic shape that is a mesh file, and assigns a value of 0 or 1 to a corresponding voxel according to a volume occupied by a voxel to any one of the mesh files.
A method for predicting the shape of a preform. 8. The method of claim 7,
The step of transforming the plurality of predicted shape matrices into a plurality of predicted shapes that are three-dimensional images includes:
The post-processing unit converts a plurality of predicted shape matrices that are voxel files through a marching cube algorithm according to the scale to a plurality of predicted shapes that are mesh files, and smoothes the surfaces of a plurality of predicted shapes that are mesh files using a smoothing algorithm. characterized by performing
A method for predicting the shape of a preform. The method of claim 1,
Before the step of converting the monotonic shape into a monotonic shape matrix, which is a three-dimensional matrix expression,
providing, by a learning unit, a plurality of learning data including a plurality of forged shapes and a plurality of label preforms corresponding to each of the plurality of forged shapes;
classifying, by the learning unit, a plurality of learning data into a plurality of learning data groups; and
generating, by the learning unit, a plurality of predictive models by performing learning using a plurality of training data belonging to the corresponding training data group for each training data group;
characterized in that it further comprises
A method for predicting the shape of a preform. A method for predicting the shape of a preform, the method comprising:
generating a plurality of predictive models that are learned to predict a preform corresponding to a forged shape by using a plurality of learning data groups each including a plurality of learning data by the learning unit; and
predicting a preform corresponding to a forged shape by a molded body derivation unit using the plurality of prediction models;
characterized in that it comprises
A method for predicting the shape of a preform. 11. The method of claim 10,
The step of generating the plurality of predictive models is
classifying, by the learning unit, a plurality of learning data including a plurality of forging shapes and a plurality of label preforms corresponding to each of the plurality of forging shapes into a plurality of training data groups; and
generating, by the learning unit, a plurality of predictive models corresponding to each of the plurality of training data groups;
characterized in that it comprises
A method for predicting the shape of a preform. 12. The method of claim 11,
The step of classifying the plurality of learning data groups is
providing, by the learning unit, a first condition indicating a combination of differences between heights of a predetermined number of feature points in an order close to the rotation axis, and a second condition indicating a combination of differences between heights of a predetermined number of feature points in an order close to the mother line; and
classifying the plurality of monotone shapes into the plurality of learning data groups according to whether the learning unit satisfies the first condition or the second condition;
characterized by comprising
A method for predicting the shape of a preform. 12. The method of claim 11,
The step of generating a plurality of predictive models corresponding to each of the plurality of training data groups comprises:
converting, by the learning unit, a monotonic shape matrix and a label preform matrix in a matrix format by voxelizing the monotonic shape and label preform of the learning data belonging to any one learning data group;
deriving a predicted shape matrix, which is a matrix expression in which the predictive model predicts the shape of the preform through a plurality of operations to which the weights for which the learning between the plurality of layers are not completed with respect to the monotonic shape matrix;
calculating, by the learning unit, a loss representing a difference between a predicted shape matrix and a label preform matrix through a loss function; and
performing optimization of modifying the weight of the prediction model so that the loss is minimized;
characterized in that it comprises
A method for predicting the shape of a preform. An apparatus for predicting the shape of a preform, comprising:
a pre-processing unit for converting a monotone shape, which is a three-dimensional image, into a monotone shape matrix, which is a three-dimensional matrix representation;
a shape prediction unit for deriving a plurality of predicted shape matrices that are matrix representations of the shape of the preform with respect to the forged shape through a plurality of prediction models learned to predict the shape of the preform;
a post-processing unit that transforms a plurality of predicted shape matrices into a plurality of predicted shapes that are three-dimensional images;
a shape determining unit that selects one of a plurality of predicted shapes as a preform that most meets a preset design condition;
characterized in that it comprises
A device for predicting the shape of a preform. 15. The method of claim 14,
The shape determining unit
Excluding or correcting a predicted shape having a singular shape among the plurality of predicted shapes,
The finite element analysis is performed on the plurality of predicted shapes to derive the mold filling rate, the presence of overlapping defects, and the forging load,
characterized in that one of the plurality of predicted shapes is selected as a preform corresponding to the forging shape according to the result of the finite element analysis
A device for predicting the shape of a preform.

Images