CN115600487B - Method and device for designing mold surface of automobile windshield glass based on convolutional neural network - Google Patents
Method and device for designing mold surface of automobile windshield glass based on convolutional neural network Download PDFInfo
- Publication number
- CN115600487B CN115600487B CN202211071143.2A CN202211071143A CN115600487B CN 115600487 B CN115600487 B CN 115600487B CN 202211071143 A CN202211071143 A CN 202211071143A CN 115600487 B CN115600487 B CN 115600487B
- Authority
- CN
- China
- Prior art keywords
- discrete points
- discrete
- glass
- profile
- mold
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 239000011521 glass Substances 0.000 title claims abstract description 271
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 90
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000013461 design Methods 0.000 claims abstract description 24
- 239000013598 vector Substances 0.000 claims description 71
- 238000012549 training Methods 0.000 claims description 42
- 238000012360 testing method Methods 0.000 claims description 37
- 239000011159 matrix material Substances 0.000 claims description 30
- 238000010606 normalization Methods 0.000 claims description 22
- 230000004913 activation Effects 0.000 claims description 16
- 238000007781 pre-processing Methods 0.000 claims description 14
- 238000004364 calculation method Methods 0.000 claims description 9
- 210000002569 neuron Anatomy 0.000 claims description 8
- 238000003062 neural network model Methods 0.000 claims description 7
- 238000013507 mapping Methods 0.000 claims description 6
- 238000011166 aliquoting Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000000611 regression analysis Methods 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims 1
- 238000000465 moulding Methods 0.000 abstract description 14
- 238000001514 detection method Methods 0.000 description 8
- 238000004519 manufacturing process Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 230000005484 gravity Effects 0.000 description 3
- 238000013003 hot bending Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 238000005452 bending Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000010791 quenching Methods 0.000 description 2
- 230000000171 quenching effect Effects 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000007664 blowing Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000001816 cooling Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 238000001035 drying Methods 0.000 description 1
- 238000007688 edging Methods 0.000 description 1
- 230000005489 elastic deformation Effects 0.000 description 1
- 238000010438 heat treatment Methods 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003825 pressing Methods 0.000 description 1
- 238000001179 sorption measurement Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000005496 tempering Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Computer Hardware Design (AREA)
- Computational Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Automation & Control Theory (AREA)
- Aviation & Aerospace Engineering (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
- Re-Forming, After-Treatment, Cutting And Transporting Of Glass Products (AREA)
Abstract
The invention provides a method and a device for designing a mold surface of an automobile windshield based on a convolutional neural network. And acquiring a plurality of groups of corresponding glass molded surfaces and mold molded surfaces from the historical design scheme, extracting discrete points of each glass molded surface, acquiring geometric characteristic data of each discrete point, and acquiring the normal distance from each discrete point to the corresponding mold molded surface. And establishing a convolutional neural network model according to the geometric characteristic data of all the discrete points and the normal distance from the discrete points to the corresponding mold surface. And extracting discrete points of the required glass molded surface, and acquiring geometric characteristic data of each discrete point on the required glass molded surface. According to the geometric feature data of discrete points of the required glass molding surface and the convolutional neural network model, predicting the normal distance from each discrete point on the required glass molding surface to the required mold molding surface, and generating the molding surface of the required mold by utilizing the geometric feature data of the discrete points on the required glass molding surface and the normal distance from the discrete points on the required glass molding surface to the required mold molding surface.
Description
Technical Field
The invention belongs to the technical field of glass mold design, and particularly relates to a method and a device for designing a mold surface of an automobile windshield glass based on a convolutional neural network.
Background
The production process of the automobile windshield comprises the following steps: firstly, the glass raw sheet is conveyed to a forming chamber after cutting, edging, cleaning, drying and heating. Then, the glass mold is sucked onto the suction surface of the suction mold machine in the molding chamber, and the glass mold below is held as the suction mold surface descends. When the glass raw sheet is heated to reach a softening point, the glass raw sheet is downwards bent and attached to the glass die under the influence of gravity, so that the set profile curvature of the glass die is achieved. Finally, the glass raw sheet formed by hot bending is cooled and subjected to subsequent steps to form the automobile windshield meeting production requirements.
Among them, the glass mold plays a very important role in molding of a glass raw sheet. However, since the glass sheet is subject to springback during the molding process, the geometric characteristics of the glass mold profile are not exactly the same as the design curve of the automobile windshield.
At present, the design of the mold surface of the automobile windshield adopts a trial-and-error method, and the general flow is as follows: (1) determining a desired glass profile; (2) preliminarily designing the molding surface of the glass mold according to the molding surface of the glass; (3) trial-manufacturing a glass finished product according to the molded surface of the glass mold; (4) Detecting deviation between the finished product of the trial-produced glass and the required glass profile; (5) And (3) if the deviation is not qualified, returning to the step (2) to adjust the molding surface of the glass mold. However, trial-and-error requires repeated corrections to the glass mold profile, which is not only long in design cycle, heavy in effort, but also costly, resulting in lower efficiency in the production of automotive windshields.
With the continued development of the automotive industry, automotive windshields are increasingly being patterned, and the design of glass mold profiles is facing new demands and challenges. Trial and error has been difficult to meet the development trend of the automobile windshield manufacturing industry due to its high design cost and long design cycle.
Disclosure of Invention
The embodiment of the invention provides a method and a device for designing the mold surface of an automobile windshield based on a convolutional neural network, which are used for solving the problems of higher design cost and longer design period in the prior art.
In order to solve the technical problems, the embodiment of the invention discloses the following technical scheme:
one aspect of the present invention provides a method for designing a mold surface of an automobile windshield glass mold based on a convolutional neural network, which is applied to generating a mold surface of a demand mold according to a glass surface of the demand, comprising:
obtaining multiple groups of molded surfaces from historical design schemes, wherein each group of molded surfaces comprises a glass molded surface and a corresponding mold molded surface;
discrete points of each glass molded surface are extracted respectively, and geometric characteristic data of each discrete point are obtained;
obtaining the normal distance from each discrete point to the corresponding mold surface;
establishing a convolutional neural network model according to the geometric feature data of all the discrete points and the normal distance from the discrete points to the corresponding mold surfaces, wherein the input variable of the convolutional neural network model is the geometric feature of the discrete points, and the output variable is the normal distance from the discrete points to the corresponding mold surfaces;
Discrete points of the required glass molded surface are extracted, and geometric feature data of each discrete point are obtained;
according to the geometric feature data of discrete points of the required glass molded surface and the convolutional neural network model, predicting the normal distance from each discrete point on the required glass molded surface to the required mold molded surface;
and generating the profile of the required mold by utilizing the geometric characteristic data of the discrete points of the required glass profile and the normal distance from the discrete points on the required glass profile to the profile of the required mold.
Optionally, the extracting discrete points of each glass profile and obtaining geometric feature data of each discrete point respectively includes:
for each glass profile, discrete points are extracted as follows:
determining a minimum bounding box of the glass profile;
establishing a three-dimensional coordinate system of the glass molded surface by taking the central point of the minimum bounding box as an origin, wherein the x-axis of the coordinate system is parallel to the longest side of the minimum bounding box, and the z-axis of the coordinate system is parallel to the shortest side of the minimum bounding box;
respectively taking the longest two boundary lines on the glass molded surface as an upper boundary line and a lower boundary line of the glass molded surface;
dividing the upper boundary line and the lower boundary line by p+1 equally, and connecting corresponding equally dividing points on the upper boundary line and the lower boundary line to obtain p equally dividing lines;
Obtaining p planes which are perpendicular to the XOY plane and pass through any one of p bisectors, wherein the planes are in one-to-one correspondence with the bisectors;
obtaining p intersecting lines intersecting p planes on the glass molded surface;
q+1 aliquoting is carried out on each intersecting line; sequentially connecting the equal division points of the corresponding sequences on the p intersecting lines to obtain q connecting lines; the intersection points of the p intersecting lines and the q connecting lines are discrete points of the glass profile.
Optionally, extracting discrete points of each glass profile, and obtaining geometric feature data of each discrete point, and further including:
(1) Acquiring the coordinates of each discrete point under the coordinate system of the glass molded surface;
(2) For each discrete point, a normal vector S is obtained as follows:
obtaining a tangent vector S with discrete points parallel to the x-axis x And a tangent vector S of the discrete point parallel to the y-axis y ；
The normal vector S is obtained by tangent vector cross multiplication: s=s x ×S y ；
(3) For each discrete point, the average curvature H is obtained as follows:
wherein r is 1 Is the maximum radius of curvature through the discrete points; r is (r) 2 Is the minimum curve passing through the discrete pointsRadius of rate;
(4) For each discrete point, a gaussian curvature G is obtained as follows:
G＝r 1 *r 2
Wherein r is 1 Is the maximum radius of curvature through the discrete points; r is (r) 2 Is the minimum radius of curvature through the discrete points;
(5) For each discrete point, arch height AH is obtained according to the following method;
the camber is the distance from the discrete point to the corresponding bisector:
AH＝Dist(P,l i ),P∈l′ i ,i＝1,2,…p
wherein the discrete point P is at the intersection line l' i Upper, l i Is equal to l' i Corresponding bisectors.
Optionally, the obtaining the normal distance from each discrete point to the corresponding mold surface includes:
for each discrete point, the normal distance of the discrete point is obtained in the following way:
acquiring an intersection point of a straight line passing through the discrete point and parallel to the normal vector of the discrete point and a corresponding mold surface;
and taking the distance between the discrete point and the intersection point as the normal distance from the discrete point to the corresponding mold surface.
Optionally, the building a convolutional neural network model according to the geometric feature data of all the discrete points and the normal distance from the discrete points to the corresponding mold surface includes:
preprocessing the coordinates of all the discrete points by adopting a MinMax (maximum minimum) normalization mode;
preprocessing normal vectors, average curvatures, gaussian curvatures and camber of all discrete points by adopting a Robust (Robust Normalization ) normalization mode;
Constructing a data set by utilizing the preprocessed geometric feature data of the discrete points and the normal distance from the discrete points to the corresponding mold surface, wherein the coordinates, normal vector, average curvature, gaussian curvature and camber of the discrete points are used as input variables of the data set, and the normal distance from the discrete points to the corresponding mold surface is used as output variables of the data set;
a convolutional neural network model between the input variable and the output variable is obtained using all the data in the dataset.
Optionally, the preprocessing is performed on the coordinates of all the discrete points in a MinMax normalization mode, including:
the coordinates of each discrete point are preprocessed according to the following method:
wherein f_max is the maximum value of all the discrete point coordinates; f_min is the minimum value in all the discrete point coordinates, and f is the original discrete point coordinate; f' is the preprocessed discrete point coordinates.
Optionally, the preprocessing is performed on the normal vector, the average curvature, the gaussian curvature and the camber of all the discrete points by adopting a robustnormalization mode, including:
the normal vector, average curvature, gaussian curvature and camber of each discrete point are all pre-processed as follows:
wherein f is an original geometric feature data value, and the geometric feature data value is a data value of one geometric feature of normal vector, average curvature, gaussian curvature and camber of the discrete points; f' is the preprocessed geometrical characteristic data value corresponding to f; f_mean is the median of the geometric feature data values corresponding to f of all the discrete points on the glass molded surface to which the discrete points belong; IQR is the interval length between the 1 st quartile and the 3 rd quartile in the geometric feature data values corresponding to f and all discrete points on the glass profile to which the discrete points belong.
Optionally, the obtaining a convolutional neural network model between the input variable and the output variable using all data in the dataset includes:
any one of all the glass molded surfaces is selected as a test glass molded surface, and the rest glass molded surfaces are all used as training glass molded surfaces;
constructing a test data set by utilizing the geometric characteristic data of the discrete points after the pretreatment of the test glass molded surface and the normal distance from the discrete points on the test glass molded surface to the corresponding mold molded surface, wherein the coordinates, the normal vector, the average curvature, the Gaussian curvature and the camber of the discrete points are used as input variables, and the normal distance from the discrete points to the corresponding mold molded surface is used as output variables;
for each training glass profile, a training dataset was constructed using the following method:
mapping the three-dimensional space information of the training glass molded surface onto a two-dimensional plane to obtain a two-dimensional glass molded surface, and simultaneously mapping discrete points on the three-dimensional glass molded surface onto the two-dimensional plane;
starting with the upper left corner on the two-dimensional glass profile, using a preset dimension n 0 ×n 0 The sliding window of the training glass profile is used for obtaining all discrete points mapped in the range of the sliding window and used as a sample corresponding to the training glass profile in a training data set; taking all newly acquired discrete points as one sample corresponding to the training glass molded surface in the training data set after the sliding window slides once, until the sliding window acquires all the discrete points mapped on the two-dimensional glass molded surface;
Constructing a training data set using all samples, wherein the coordinates, normal vectors, average curvature, gaussian curvature and camber of all discrete points in each sample are used as input variables, and the dimension of the input variables is 9×n 0 ×n 0 The method comprises the steps of carrying out a first treatment on the surface of the Taking the normal distance from all discrete points in each sample to the corresponding mold surface as an output variable, wherein the dimension of the output variable is n 0 ×n 0 ；
And building a convolutional neural network model according to all the training data sets and the test data sets.
Optionally, the building a convolutional neural network model according to the training data set and the test data set includes:
constructing an initial model of a convolutional neural network model according to a training data set, wherein the initial model of the convolutional neural network model is composed of a convolutional block and a full connecting block, and the initial model of the convolutional neural network model comprises the following components:
the convolutional blocks of the convolutional neural network are designed as follows:
(1) Using m 1 Convolutions (Convolitions) layer and m 1 The individual batch normalization (Batch Normalization) layer performs deep feature extraction and uses the LeakyRelu function as an activation function; the LeakyRelu function is defined as:
wherein x is the activation value of a single neuron; λ is a preset slope when x is less than 0;
(2) The dimension of the input layer is consistent with the input dimension of the sample, and is 9 multiplied by n 0 ×n 0 ；
(3) The output dimension of the ith convolution layer is c i ×n i ×n i ,i≥1；
Wherein c i The number of convolution kernels in the ith convolution layer; n is n i Length/width of output data for ith convolutional layer, n i The calculation formula of (2) is as follows:
wherein n is i-1 Length/width, k of output data for the i-1 th convolutional layer i For the size of the convolution kernel in the ith convolution layer, p i Filling length s for the ith convolution layer i Step length for moving convolution kernel in ith convolution layer;
the fully connected block of the convolutional neural network is designed as follows:
(1) Using 1 expansion layer and m 2 Regression analysis was performed on the individual full-ligation (Dense) layers, the first m 2 -activation function of 1 fully connected layerFor the Relu function, the last fully connected layer does not use an activation function; the definition of the Relu function is: f (x) =max (0, x), where x is the activation value of a single neuron;
(2) The output data of the convolution layer is expanded through the flat layer to obtain a 1-dimensional vector with the length ofThe unfolding mode is as follows:
wherein,is the dimension of the convolution layer; the output dimension of the ith convolution layer is c i ×n i ×n i Then the mth 1 The output dimension of the individual convolution layers is +.>The output dimension of the convolution layer is understood as +.>Personal (S)Is->For the 1 st row 1 st column value in the first two-dimensional matrix +. >For row 1, nth in a first two-dimensional matrix m1 Column value, +.>For the +.>Line->Values of columns; and so on, let us go of>Is->Values of row 1 and column 1 in a two-dimensional matrix,>is->Line 1 of the two-dimensional matrixColumn value, +.>Is->The first two-dimensional matrix>Line->Values of columns;
(3)m 2 the input dimension of the full connection (Dense) layer is equal to the output dimension of the upper layer; wherein the input dimension of the 1 st full connection layer is L 0 Mth, m 2 The output dimension of each full connection layer is L 1 ＝n 0 ×n 0 。
Optionally, building a convolutional neural network model according to the training data set and the test data set, and further includes:
and obtaining a finally established convolutional neural network model by adopting a plurality of loss functions according to the test data set and the initial model, wherein the loss functions comprise: average squared error loss, absolute value loss, relative entropy loss, style loss, and relevance loss functions; wherein:
the average Square Loss (MSL), the absolute value Loss (L1 Loss, L1) and the relative entropy Loss (Kullback-Leibler Divergence Loss, KLDiv) are all indexes for evaluating the estimation precision of the point, and the specific formulas are as follows:
wherein,representing the true value of the normal distance from the ith discrete point to the corresponding mould profile in the test dataset,/ >The predicted value of the normal distance from the i discrete point to the corresponding mold surface is predicted according to the geometric feature data of the convolutional neural network model and the i discrete point, and N represents the total number of the discrete points in all samples of the test data set;
the style Loss (Gram Loss, gram) and the relevance Loss (corelate Loss, corr) are indexes for evaluating the estimation precision of the two-dimensional plane;
the style loss is calculated by using the difference value of the vector inner product, and the calculation mode is as follows:
the correlation loss can represent the correlation degree of a transverse vector and a longitudinal vector in the sliding window, wherein the transverse vector is a set of each row in the sliding window, and the longitudinal vector is a set of each column in the sliding window; the correlation coefficient between vectors is calculated by using a Pearson correlation coefficient, and the calculation mode is as follows:
where m is the total number of samples in the test dataset,and->Representing the true value on the two-dimensional plane in the kth sample +.>And predictive value->The matrix is formed, K represents the matrix->Is a dimension of (2); wherein matrix Y t Sum matrix Y p The forms of (a) are as follows:
the row and column vectors of matrix Y may be expressed as
Ycol i ＝[y i1 ,y i2 ,…,y ik ],i＝1,2,…,K；
Yrow i ＝[y 1i ,y i2i ,...,y ki ],i＝1,2,…,K。
Optionally, the obtaining a final convolutional neural network model by using a plurality of loss functions according to the test data set and the initial model includes:
The average square error loss, absolute value loss, relative entropy loss, style loss and correlation loss function are weighted and averaged to obtain a total loss function;
and (3) based on the total loss function, re-determining parameters in the convolutional neural network model by adopting a gradient descent method to obtain the finally established convolutional neural network model.
Optionally, predicting the normal distance from each discrete point on the required glass profile to the required mold profile according to the geometric feature data of the discrete point on the required glass profile and the convolutional neural network model includes:
preprocessing the coordinates of all discrete points on the required glass molded surface in a MinMax normalization mode;
preprocessing normal vectors, average curvatures, gaussian curvatures and camber of all discrete points on a required glass molded surface in a Robust normalization mode;
substituting the geometric feature data of all discrete points of the required glass molded surface after pretreatment into a convolutional neural network model as input data, and calculating to obtain a predicted value of the normal distance from each discrete point on the required glass molded surface to the required mold molded surface;
determining whether there are discrete points on the desired glass profile having a plurality of predicted values,
if so, the maximum value in the plurality of predicted values is taken as the final predicted value of the corresponding discrete point.
Optionally, the generating the profile of the demand mold by using the geometric feature data of the discrete point of the profile of the demand glass and the normal distance from the discrete point of the profile of the demand glass to the profile of the demand mold includes:
calculating coordinates of each discrete point of the required mold surface by utilizing geometric feature data of the discrete points on the required glass surface and normal distance from the discrete point of the required glass surface to the required mold surface, wherein the discrete points of the required mold surface correspond to the discrete points on the required glass surface one by one and are assumed discrete points on the required mold surface;
the coordinates of each discrete point of the desired mold profile are calculated as follows:
wherein: (x' i ,y′ i ,z′ i ) Coordinates of discrete points of the required mold surface; (x) i ,y i ,z i ) Coordinates of corresponding discrete points on the required glass profile; (n) xi ,n yi ,n zi ) Normal vector of corresponding discrete point on the required glass profile; d, d i The normal distance from the corresponding discrete point on the predicted required glass molded surface to the required mold molded surface is set;
constructing a spline curve network wire frame connected with discrete points according to the coordinates of the discrete points of the required mold surface;
and (3) carrying out curved surface reconstruction on the spline curve network wire frame by using a geometric modeling engine to generate the molded surface of the required mold.
Another aspect of the present invention provides a device for designing a mold profile of an automotive windshield based on a convolutional neural network, which is applied to generating a desired mold profile according to a desired glass profile, comprising:
the glass profile and mould profile acquisition unit is used for acquiring a plurality of groups of profiles from the historical design scheme, wherein each group of profiles comprises a glass profile and a corresponding mould profile;
the glass profile discrete point data acquisition unit is used for respectively extracting discrete points of each glass profile and acquiring geometric characteristic data of each discrete point;
the normal distance acquisition unit is used for acquiring the normal distance from each discrete point to the corresponding mold surface;
the convolution neural network model building unit is used for building a convolution neural network model according to the geometric feature data of all the discrete points and the normal distance from the discrete points to the corresponding mold surfaces, wherein the input variable of the convolution neural network model is the geometric feature of the discrete points, and the output variable is the normal distance from the discrete points to the corresponding mold surfaces;
the glass profile discrete point data acquisition unit is also used for extracting discrete points requiring the glass profile and acquiring geometric characteristic data of each discrete point;
The normal distance prediction unit is used for predicting the normal distance from each discrete point on the required glass profile to the required mold profile according to the geometric feature data of the discrete points of the required glass profile and the convolutional neural network model;
the required mold profile generating unit is used for generating the profile of the required mold by utilizing the geometric characteristic data of the discrete points of the required glass profile and the normal distance from the discrete points on the required glass profile to the required mold profile.
As can be seen from the above technical solutions, in the method and the device for designing a mold surface of an automotive windshield based on a convolutional neural network provided by the embodiments of the present invention, first, multiple sets of corresponding glass mold surfaces and mold surfaces are obtained from a historical design scheme, discrete points of each glass mold surface are respectively extracted, geometric feature data of each discrete point are obtained, and a normal distance from each discrete point to the corresponding mold surface is obtained.
And secondly, establishing a convolutional neural network model according to the geometric feature data of all the discrete points and the normal distance from the discrete points to the corresponding mold surface. And extracting discrete points of the required glass molded surface, and acquiring geometric characteristic data of each discrete point on the required glass molded surface.
And finally, predicting the normal distance from each discrete point on the required glass profile to the required mold profile according to the geometric feature data of the discrete point on the required glass profile and the convolutional neural network model, and generating the profile of the required mold by utilizing the geometric feature data of the discrete point on the required glass profile and the normal distance from the discrete point on the required glass profile to the required mold profile.
Therefore, the method and the device provided by the embodiment of the invention can greatly improve the design efficiency of the mold surface of the automobile windshield and effectively reduce the iteration cost in the design process.
Drawings
Fig. 1 is a schematic flow chart of a design method of a mold surface of an automobile windshield based on a convolutional neural network according to an embodiment of the present invention;
FIG. 2 is a flowchart illustrating implementation of step S102 in FIG. 1 according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a minimum bounding box of an automobile windshield according to an embodiment of the present invention;
FIG. 4 is a schematic view of discrete points on a glass surface according to one embodiment of the present invention;
FIG. 5 is a flowchart illustrating implementation of step S104 in FIG. 1 according to an embodiment of the present invention;
FIG. 6 is a flowchart illustrating implementation of step S107 in FIG. 1 according to an embodiment of the present invention;
Fig. 7 is a schematic structural diagram of an automotive windshield mold surface design device based on a convolutional neural network according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail below with reference to the drawings and detailed description for the purpose of better understanding of the technical solution of the present invention to those skilled in the art.
Fig. 1 is a schematic flow chart of a design method of a mold surface of an automobile windshield based on a convolutional neural network, which is applied to generating a mold surface according to a required glass surface, and is characterized in that. As shown in fig. 1, the method comprises the steps of:
step S101: multiple sets of profiles are obtained from the historical design, each set of profiles including a glass profile and a corresponding mold profile.
The glass molded surfaces in each set of molded surfaces and the mold molded surfaces correspond to each other, and each glass molded surface is manufactured by the corresponding mold molded surface.
Step S102: and respectively extracting discrete points of each glass molded surface, and acquiring geometric characteristic data of each discrete point.
In one embodiment of the present disclosure, for each glass profile, as shown in FIG. 2, discrete points are extracted as follows:
Step S1021: a minimum bounding box of the glass profile is determined.
The minimum bounding box is the minimum bounding rectangle of the glass profile, which in the disclosed embodiments of the invention can be obtained using existing algorithms.
Step S1022: and establishing a three-dimensional coordinate system of the glass molded surface by taking the central point of the minimum bounding box as an origin.
As shown in fig. 3, the x-axis of the three-dimensional coordinate system is parallel to the longest side of the minimum bounding box, the z-axis of the coordinate system is parallel to the shortest side of the minimum bounding box, and the y-axis of the coordinate system is parallel to the middle length side of the minimum bounding box.
Step S1023: the longest two borderlines on the glass profile are respectively taken as an upper borderline and a lower borderline of the glass profile.
The glass profile is generally an irregular curved surface of approximately rectangular shape, and the boundary of the glass profile is formed by an outwardly curved curve. When the glass profile is upright, its boundary lines are divided into an upper boundary line, a lower boundary line, a left boundary line, and a right boundary line. Wherein, the upper and lower boundary lines of the glass profile are longest.
Step S1024: and respectively carrying out p+1 equal division on the upper boundary line and the lower boundary line, and connecting corresponding equal division points on the upper boundary line and the lower boundary line to obtain p equal division lines.
Dividing the upper boundary line and the lower boundary line by p+1 equally, thereby obtaining p equally divided points up on the upper boundary line i I=1, 2, … p, and p bisectors lp on the lower boundary line i ,i＝1,2,…p。
Connecting corresponding bisectors on the upper and lower border lines, e.g. up 1 Connection lp 1 ，up 2 Connection lp 2 . After connecting the corresponding p equal-dividing points on the upper and lower boundary lines, obtaining p equal-dividing lines l i ,i＝1,2,…p。
Step S1025: p planes perpendicular to the XOY plane and passing through any of the p bisectors are obtained.
With a bisector line l 1 For example, obtain a line of bisection l 1 At the same time, the plane is perpendicular to the XOY plane on the three-dimensional coordinate system, which plane can intersect the glass profile. In the above manner, p planes are obtained, each plane passing through a bisector l i And perpendicular to the XOY plane. Each plane passing through only one bisector line l i The planes are in one-to-one correspondence with the bisectors.
Step S1026: p intersecting lines on the glass profile intersecting p planes are obtained.
The p planes obtained all intersect the glass profile, so that p intersecting lines l 'on the glass profile can be obtained' i I=1, 2, … p, the intersection line being a curve.
Step S1027: q+1 aliquoting is carried out on each intersection line; sequentially connecting the equal division points of the corresponding sequences on the p intersecting lines to obtain q connecting lines; the intersection of the p intersecting lines and the q connecting lines is a discrete point of the glass profile.
For each intersection line l on the glass profile′ i Each of i=1, 2, … p was q+1 equally divided to obtain each intersection line l' i I=1, 2, q aliquots on … p.
Sequentially connecting corresponding sequential bisectors of p intersecting lines, e.g. connecting intersecting line l' 1 The first bisecting point on the line of intersection l' 2 The first bisecting point on the first line is connected in turn until it is connected to the intersection line l' p The first of the points of bisection, thereby obtaining a connecting line connecting the first of the points of bisection on each of the intersecting lines. After the connection is completed at the q-th bisector on each intersection line, q connection lines on the glass profile can be obtained.
As shown in fig. 4, the intersections of the p intersecting lines with the q connecting lines are taken as discrete points corresponding to the glass profile.
In the disclosed embodiments, the geometric features required to obtain discrete points are discrete point coordinates, normal vectors, average curvature, gaussian curvature, and camber.
The reason for extracting the characteristic parameters is as follows:
combining the hot bending forming and rebound processes of the glass, and lifting the glass raw sheet and attaching the glass raw sheet to the suction mold under the combined action of vacuum adsorption and blowing of the suction film after the glass raw sheet is heated at high temperature; then the suction mold and the glass descend to press the glass together with the hot ring; the glass falls on the cooling ring after being pressed and formed, and is rapidly supported and delivered to the quenching cold air grid area for quenching tempering.
The glass sheet is formed by pressing a suction mold surface and a hot ring surface, and then falls on a cold ring by means of gravity. The stress distribution conditions at each stage are as follows: when in press forming, the surface of the suction mold is firstly contacted with the middle area of the glass raw sheet, and gradually attached to the boundary, so that the stress is small and uneven. After the glass edge is extruded by the surface of the suction mold and the hot ring surface, boundary stress is increased and the distribution is more uniform. When the mould falls, the glass falls onto the cold ring at a certain speed under the action of self gravity, so that impact load is generated, and the stress at the edge of the glass suddenly increases.
Therefore, the stress distribution of the windshield in different positions during the molding process has a certain difference, which can cause the deviation between the different positions on the molded glass surface and the molded surface of the mold to be different. Based on this, positional information of discrete points on the glass surface, i.e., discrete point coordinates (x/y/z), is extracted.
In the hot bending forming process of the glass raw sheet, when the glass raw sheet is subjected to an external bending moment, the shape of the glass raw sheet is changed. When the glass is bent, the inner glass raw sheet of the deformation zone is subjected to tangential compressive stress to generate compression deformation, the outer glass raw sheet is subjected to tangential tensile stress to generate tensile deformation, and the tangential direction also influences the deformation of the glass profile, so that unit normal vectors (vector_x/vector_y/vector_z) perpendicular to the tangential direction are extracted for characterizing the tangential characteristics of the glass profile.
Meanwhile, the radius of curvature of the mold surface is a main factor determining the glass surface, the degree of curvature of the glass surface depends on the shape of the mold surface, and the degree of rebound is related to the radius of curvature of the mold surface. This is because the smaller the radius of curvature of the mold surface, the deeper the relative degree of bending, and the larger the proportion of elastic deformation in the total deformation increases, and the larger the rebound after the external force is removed. Thus, the curvature characteristics of the glass profile, including the average curvature H and gaussian curvature K, are extracted.
In addition, to characterize the degree of curvature of the glass profile, the camber AH of the discrete points was also extracted.
The data values of each geometrical feature of the discrete points are obtained respectively in the following way:
(1) The coordinates of each discrete point are obtained under the coordinate system of the glass profile.
(2) For each discrete point, a normal vector S is obtained as follows:
obtaining a tangent vector S with discrete points parallel to the x-axis x And a tangent vector S of the discrete point parallel to the y-axis y 。
The normal vector S is obtained by tangent vector cross multiplication: s=s x ×S y 。
(3) For each discrete point, the average curvature H is obtained as follows:
wherein r is 1 Is the maximum radius of curvature through the discrete points; r is (r) 2 Is the smallest radius of curvature through the discrete points.
(4) For each discrete point, a gaussian curvature G is obtained as follows:
G＝r 1 *r 2
wherein r is 1 Is the maximum radius of curvature through the discrete points; r is (r) 2 Is the smallest radius of curvature through the discrete points.
(5) For each discrete point, the arch height AH is obtained according to the following method, wherein the arch height is from the discrete point P to the corresponding bisector l i Distance of (2):
AH＝Dist(P,l i ),P∈l′ i ,i＝1,2,…p
wherein the discrete point P is at the intersection line l' i I=1, 2, …, p, l i Is equal to l' i Corresponding bisectors.
Step S103: and obtaining the normal distance from each discrete point to the corresponding mold surface.
In the disclosed embodiment of the invention, the normal distance of each discrete point to the corresponding mold profile is obtained as follows:
for each discrete point, the normal distance of the discrete point is obtained in the following manner:
(1) And acquiring the intersection point of the straight line passing through the discrete point and parallel to the normal vector of the discrete point and the corresponding mold surface.
(2) And taking the distance between the discrete point and the intersection point as the normal distance from the discrete point to the corresponding mold surface.
Step S104: and establishing a convolutional neural network model according to the geometric characteristic data of all the discrete points and the normal distance from the discrete points to the corresponding mold surface.
The input variable of the convolutional neural network model is the geometric characteristic of a discrete point, and the output variable is the normal distance from the discrete point to the corresponding mold surface.
In one embodiment of the present disclosure, as shown in FIG. 5, a convolutional neural network model is built in the following manner:
the geometric feature data of the discrete points are preprocessed before building the convolutional neural network model.
Step S1041: and preprocessing the coordinates of all the discrete points by adopting a MinMax normalization mode.
In one embodiment of the present disclosure, the coordinates of each discrete point are pre-processed as follows:
wherein f_max is the maximum value of all the discrete point coordinates; f_min is the minimum value in all the discrete point coordinates, and f is the original discrete point coordinate; f' is the preprocessed discrete point coordinates.
Step S1042: and preprocessing the normal vector, the average curvature, the Gaussian curvature and the camber of all the discrete points by adopting a Robust normalization mode.
In one embodiment of the present disclosure, the normal vector, average curvature, gaussian curvature, and camber of each discrete point are all pre-processed as follows:
wherein f is an original geometric feature data value, and the geometric feature data value is a geometric feature data value of one of normal vector, average curvature, gaussian curvature and camber of discrete points; f' is the preprocessed geometrical characteristic data value corresponding to f; f_mean is the median of the geometric feature data values corresponding to f and all the discrete points on the glass molded surface to which the discrete points belong; IQR is the interval length between the 1 st quartile and the 3 rd quartile in the geometric feature data values corresponding to f and all discrete points on the glass profile to which the discrete points belong. For example, f is the original data value of the camber of the discrete point, f' is the data value of the camber after the pretreatment of the discrete point, f_mean is the median of the camber data values of all the discrete points on the glass surface to which the discrete point belongs, and IQR is the interval length between the 1 st quartile and the 3 rd quartile of the camber data values of all the discrete points on the glass surface to which the discrete point belongs.
Step S1043: and constructing a data set by utilizing the geometric characteristic data of the preprocessed discrete points and the normal distance from the discrete points to the corresponding mold surfaces.
The coordinates, normal vector, average curvature, gaussian curvature and camber of the discrete points are used as input variables in the data set, and the normal distance from the discrete points to the corresponding mold surface is used as output variables in the data set.
Step S1044: a convolutional neural network model between the input variable and the output variable is obtained using all the data in the dataset.
In one embodiment of the present disclosure, this step may be implemented by the sub-steps of:
(1) Any one of all the glass profiles is selected as a test glass profile, and all the glass profiles except the test glass profile are used as training glass profiles.
(2) And constructing a test data set by utilizing the geometric characteristic data of the discrete points of the test glass molded surface after pretreatment and the normal distance from the discrete points on the test glass molded surface to the corresponding mold molded surface. The coordinates, normal vectors, average curvature, gaussian curvature and camber of the discrete points are used as input variables, and the normal distance from the discrete points to the corresponding mold surface is used as an output variable.
(3) For each training glass profile, a training dataset was constructed using the following method:
And mapping the three-dimensional space information of the training glass profile onto a two-dimensional plane to obtain the two-dimensional glass profile, and simultaneously mapping the discrete points on the three-dimensional glass profile onto the two-dimensional plane.
Starting with the upper left corner on the two-dimensional glass profile, using a preset dimension n 0 ×n 0 All discrete points mapped within the sliding window are obtained as a sample of the training data set corresponding to the training glass profile, in one embodiment of the present inventionIn an embodiment, the preset size of the sliding window is 32×32. And taking all newly acquired discrete points as one sample corresponding to the training glass profile in the training data set after the sliding window slides once, and sequentially acquiring the samples according to the mode until the sliding window acquires all the discrete points mapped on the two-dimensional glass profile. Wherein the number of sliding window slides, i.e. the number of samples corresponding to the training glass profile, is noted m.
Constructing a training data set using all samples, wherein the coordinates, normal vectors, average curvature, gaussian curvature and camber of all discrete points in each sample are used as input variables, and the dimension of the input variables is 9×n 0 ×n 0 Wherein 9 means the total dimension of the geometric feature, wherein the coordinates and normal vector of the discrete points have 3 dimensions respectively, and the average curvature, gaussian curvature and camber have 1 dimension respectively, and 9 dimensions in total.
Taking the normal distance from all discrete points in each sample to the corresponding mold surface as an output variable, wherein the dimension of the output variable is n 0 ×n 0 。
And acquiring a training data set corresponding to each training glass profile in the above manner.
(4) And building a convolutional neural network model according to all the training data sets and the test data sets.
In one embodiment of the present disclosure, a convolutional neural network model may be built in the following manner:
constructing an initial model of a convolutional neural network model according to all training data sets, wherein the initial model of the convolutional neural network model is composed of a convolutional block and a full connecting block, and the initial model of the convolutional neural network model is composed of the convolutional block and the full connecting block, wherein the initial model is composed of the convolutional block and the full connecting block, and the initial model is composed of the convolutional block and the full connecting block, wherein the initial model is composed of the convolutional block and the:
the convolutional blocks of the convolutional neural network are designed as follows:
(1) Using m 1 Convolutions (Convolitions) layer and m 1 The individual batch normalization (Batch Normalization) layer performs deep feature extraction and uses the LeakyRelu function as an activation function; the LeakyRelu function is defined as:
wherein x is the activation value of a single neuron; λ is the slope when x is less than 0, in one embodiment of the present disclosure, λ=0.2;
(2) The dimension of the input layer is consistent with the input dimension of the sample, and is 9 multiplied by n 0 ×n 0 ；
(3) The output dimension of the ith convolution layer is c i ×n i ×n i ,i≥1；
Wherein c i The number of convolution kernels in the ith convolution layer; n is n i Length/width of output data for ith convolutional layer, n i The calculation formula of (2) is as follows:
wherein n is i-1 Length/width, k of output data for the i-1 th convolutional layer i For the size of the convolution kernel in the ith convolution layer, p i Filling length s for the ith convolution layer i Step length for moving convolution kernel in ith convolution layer;
the fully connected block of the convolutional neural network is designed as follows:
(1) Using 1 expansion layer and m 2 Regression analysis was performed on the individual full-ligation (Dense) layers, the first m 2 The activation function of 1 fully connected layer is the Relu function, the last fully connected layer does not use the activation function; the definition of the Relu function is: f (x) =max (0, x), where x is the activation value of a single neuron;
(2) The output data of the convolution layer is expanded through the flat layer to obtain a 1-dimensional vector with the length ofThe unfolding mode is as follows:
wherein,is the dimension of the convolution layer; the output dimension of the ith convolution layer is c i ×n i ×n i The output dimension of the m1 st convolution layer is +.>The output dimension of the convolution layer is understood as +.>Personal (S)Is->For the 1 st row 1 st column value in the first two-dimensional matrix +.>For row 1, nth in a first two-dimensional matrix m1 Column value, +.>For the +.>Line->Values of columns; and so on, let us go of >Is->Values of row 1 and column 1 in a two-dimensional matrix,>is->Row 1 +.>Column value, +.>Is->The first two-dimensional matrix>Line->Values of columns;
(3)m 2 the input dimension of the full connection (Dense) layer is equal to the output dimension of the upper layer; wherein the input dimension of the 1 st full connection layer is L 0 Mth, m 2 The output dimension of each full connection layer is L 1 ＝n 0 ×n 0 。
After the convolution block and the full connection block are built, the initial model building process of the convolution neural network model is completed.
In one embodiment of the present disclosure, the convolutional portion of the convolutional neural network is formed using two convolutional block connections, each convolutional layer containing a convolutional kernel, a batch normalization layer, and a LeakyRelu activation function. Wherein the dimensions of the two convolution kernels are 4×4 and 3×3 in sequence.
The fully connected block uses three fully connected layers to downsample the unfolded layer neurons layer by layer, with each neuron using the Relu activation function except for the last fully connected layer.
In one embodiment of the present disclosure, after an initial model of the completed convolutional neural network model is built, the following steps are calculated:
and obtaining a finally established convolutional neural network model by adopting various loss functions according to the test data set and the initial model. Wherein the loss function comprises: average squared error loss, absolute value loss, relative entropy loss, style loss, and relevance loss function, wherein:
The Mean Square Loss (MSL), the absolute value Loss (L1 Loss, L1) and the relative entropy Loss (Kullback-Leibler Divergence Loss, KLDiv) are all indices that evaluate the point estimation accuracy, which is the average error between the predicted data and the true data for each discrete point in the test data set.
The specific formulas for average square loss, absolute value loss, and relative entropy loss are as follows:
wherein,representing the true value of the normal distance from the ith discrete point to the corresponding mould profile in the test dataset,/>The predicted value of the normal distance from the i discrete point to the corresponding mold surface is predicted according to the geometric feature data of the convolutional neural network model and the i discrete point, and N represents the total number of the discrete points in all samples of the test data set;
the style Loss (Gram Loss, gram) and the correlation Loss (correlay Loss, corr) are both indexes for evaluating the two-dimensional plane estimation accuracy, which is the average deviation between the predicted data and the real data of each sample in the test data set.
The style loss is calculated by using the difference value of the vector inner product, and the calculation mode is as follows:
the correlation loss can represent the correlation degree of a transverse vector and a longitudinal vector in the sliding window, wherein the transverse vector is a set of each row in the sliding window, and the longitudinal vector is a set of each column in the sliding window; the correlation coefficient between vectors is calculated by using a Pearson correlation coefficient, and the calculation mode is as follows:
Where m is the total number of samples in the test dataset,and->Representing the true value on the two-dimensional plane in the kth sample +.>And predictive value->The matrix is formed, K represents the matrix->Is a dimension of (2); wherein matrix Y t Sum matrix Y p The forms of (a) are as follows:
therefore, the row and column vectors of matrix Y can be expressed as
Ycol i ＝[y i1 ,y i2 ,…,y ik ],i＝1,2,…,K；
Yrow i ＝[y 1i ,y i2i ,...,y ki ],i＝1,2,…,K。
In one embodiment of the present disclosure, in the foregoing embodiment, a final convolutional neural network model is obtained by using a plurality of loss functions according to a test data set and an initial model, including the following steps:
(1) The average square error loss, absolute value loss, relative entropy loss, style loss, and correlation loss functions calculated in the foregoing embodiments are weighted-averaged to obtain a total loss function.
(2) And (3) based on the total loss function, re-determining parameters in the convolutional neural network model by adopting a gradient descent method, and obtaining the finally established convolutional neural network model.
After the final convolutional neural network model is built, step S105 is continued.
Step S105: and extracting discrete points requiring the glass profile, and acquiring geometric characteristic data of each discrete point.
The required glass profile is a glass profile which is required to be designed corresponding to the mold profile and is a glass profile which is designed in advance.
The three-dimensional coordinate system of the required glass profile is constructed according to the method of extracting the discrete points of the glass profile in the foregoing embodiment, and the discrete points of the required glass profile are extracted, which will not be described herein.
According to the method for acquiring the geometric feature data of each discrete point on the glass surface in the foregoing embodiment, the geometric feature data of each discrete point on the glass surface is acquired, which is not described herein.
Step S106: and predicting the normal distance from each discrete point on the required glass profile to the required mold profile according to the geometric feature data of the discrete points of the required glass profile and the convolutional neural network model.
Substituting the geometric feature data of discrete points on the required glass molded surface into a finally established convolutional neural network model, wherein the convolutional neural network model predicts the normal distance from each discrete point to the required mold molded surface.
In one embodiment of the present disclosure, this step may be accomplished by the sub-steps of:
(1) The coordinates of all discrete points on the desired glass profile were pre-processed in the manner described in the previous examples using a MinMax normalization.
(2) According to the mode in the embodiment, the normal vector, the average curvature, the Gaussian curvature and the arch height of all discrete points on the required glass molded surface are preprocessed by adopting a Robust normalization mode;
(3) Substituting the geometric characteristic data of all the discrete points of the required glass molded surface after pretreatment into a finally established convolutional neural network model as input data, and calculating to obtain a predicted value of the normal distance from each discrete point on the required glass molded surface to the required mold molded surface.
(4) Determining whether there are discrete points on the desired glass profile having a plurality of predicted values,
if so, it is stated that the one or more discrete points have more than two predicted values, and when this occurs, the maximum value of the plurality of predicted values is taken as the final predicted value of the corresponding discrete point.
If not, it is indicated that such an abnormal condition does not exist, a unique predicted value corresponding to each discrete point is normally determined.
Step S107: and generating the profile of the required mold by utilizing the geometric characteristic data of the discrete points of the required glass profile and the normal distance from the discrete points on the required glass profile to the profile of the required mold.
In the disclosed embodiment of the invention, as shown in fig. 6, step S107 may be accomplished using the following sub-steps:
step S1071: and calculating the coordinates of each discrete point of the required mold surface by utilizing the geometric characteristic data of the discrete points on the required glass surface and the normal distance from the discrete point of the required glass surface to the required mold surface.
Wherein, the discrete points of the required mold profile are in one-to-one correspondence with the discrete points on the required glass profile, and are assumed discrete points on the required mold profile.
The coordinates of each discrete point of the desired mold profile are calculated as follows:
wherein: (x' i ,y′ i ,z′ i ) Coordinates of discrete points of the required mold surface; (x) i ,y i ,z i ) Coordinates of corresponding discrete points on the required glass profile; (n) xi ,n yi ,n zi ) Normal vector of corresponding discrete point on the required glass profile; d, d i The normal distance from the corresponding discrete point on the predicted required glass profile to the required mold profile is set.
Step S1072: and constructing a spline curve network wire frame connected with the discrete points according to the coordinates of the discrete points of the required mold surface.
In one embodiment of the present disclosure, step S1072 may be implemented by:
(1) The discrete points of the required mold surface are transposed into a p multiplied by q two-dimensional matrix by using the existing geometric modeling engine.
(2) And randomly selecting a plurality of groups of preset number of discrete points from all discrete points of the required mold surface, and respectively generating a plurality of detection surfaces of the required mold by using a geometric modeling engine. Wherein each detection profile is generated from a preset number of discrete points.
(3) According to the coordinates of discrete points of the required mold surface and the detection surfaces of the multiple groups of required molds, the average error between all discrete points of the required mold surface and the detection surfaces of each required mold is calculated respectively, namely, the average error between all discrete points of the required mold surface and the detection surfaces of each required mold is calculated according to the detection surfaces of each required mold. The average error may be the average of the normal distances of all discrete points to the detection profile.
(4) And determining a detection profile with the minimum average error, and taking the corresponding discrete point as the selected discrete point.
(5) Connecting the selected discrete points to form a spline curve network wire frame.
Step S1073: and (3) carrying out curved surface reconstruction on the spline curve network wire frame by using a geometric modeling engine to generate the molded surface of the required mold.
And (3) carrying out curved surface reconstruction on the spline curve network wire frame by using the conventional geometric modeling engine, and constructing a final required mold surface.
Fig. 7 is a schematic structural diagram of an automotive windshield mold profile design device based on a convolutional neural network, which is applied to generating a required mold profile according to a required glass profile. As shown in fig. 7, the apparatus includes the following units:
a glass profile and mold profile acquiring unit 11 configured to acquire a plurality of sets of profiles each including one glass profile and a corresponding one of the mold profiles from the historical design;
a glass profile discrete point data acquisition unit 12 configured to extract discrete points of each glass profile, respectively, and acquire geometric feature data of each discrete point;
a normal distance acquisition unit 13 configured to acquire a normal distance from each discrete point to a corresponding mold surface;
The convolutional neural network model building unit 14 is configured to build a convolutional neural network model according to the geometric feature data of all the discrete points and the normal distance from the discrete points to the corresponding mold surfaces, wherein the input variable of the convolutional neural network model is the geometric feature of the discrete points, and the output variable is the normal distance from the discrete points to the corresponding mold surfaces;
a glass profile discrete point data acquisition unit 12 further configured to extract discrete points requiring a glass profile and acquire geometric feature data of each discrete point;
a normal distance predicting unit 15 configured to predict a normal distance from each discrete point on the required glass profile to the required mold profile according to the geometric feature data of the discrete points of the required glass profile and the convolutional neural network model;
the required mold profile generating unit 16 is configured to generate a profile of the required mold using geometric feature data of discrete points of the required glass profile and a normal distance from the discrete points on the required glass profile to the required mold profile.
It is to be understood that the above embodiments are merely illustrative of the application of the principles of the present invention, but not in limitation thereof. Various modifications and improvements may be made by those skilled in the art without departing from the spirit and substance of the invention, and are also considered to be within the scope of the invention.
Claims (14)
1. The utility model provides a car windshield mould profile design method based on convolutional neural network, is applied to and generates demand mould profile according to demand glass profile, characterized by comprising:
obtaining multiple groups of molded surfaces from historical design schemes, wherein each group of molded surfaces comprises a glass molded surface and a corresponding mold molded surface;
discrete points of each glass molded surface are extracted respectively, and geometric characteristic data of each discrete point are obtained;
obtaining the normal distance from each discrete point to the corresponding mold surface;
establishing a convolutional neural network model according to the geometric feature data of all the discrete points and the normal distance from the discrete points to the corresponding mold surfaces, wherein the input variable of the convolutional neural network model is the geometric feature of the discrete points, and the output variable is the normal distance from the discrete points to the corresponding mold surfaces;
discrete points of the required glass molded surface are extracted, and geometric feature data of each discrete point are obtained;
according to the geometric feature data of discrete points of the required glass molded surface and the convolutional neural network model, predicting the normal distance from each discrete point on the required glass molded surface to the required mold molded surface;
and generating the profile of the required mold by utilizing the geometric characteristic data of the discrete points of the required glass profile and the normal distance from the discrete points on the required glass profile to the profile of the required mold.
2. The method of claim 1, wherein extracting discrete points of each glass profile and obtaining geometric feature data for each of the discrete points, respectively, comprises:
for each glass profile, discrete points are extracted as follows:
determining a minimum bounding box of the glass profile;
establishing a three-dimensional coordinate system of the glass molded surface by taking the central point of the minimum bounding box as an origin, wherein the x-axis of the coordinate system is parallel to the longest side of the minimum bounding box, and the z-axis of the coordinate system is parallel to the shortest side of the minimum bounding box;
respectively taking the longest two boundary lines on the glass molded surface as an upper boundary line and a lower boundary line of the glass molded surface;
dividing the upper boundary line and the lower boundary line by p+1 equally, and connecting corresponding equally dividing points on the upper boundary line and the lower boundary line to obtain p equally dividing lines;
obtaining p planes which are perpendicular to the XOY plane and pass through any one of p bisectors, wherein the planes are in one-to-one correspondence with the bisectors;
obtaining p intersecting lines intersecting p planes on the glass molded surface;
q+1 aliquoting is carried out on each intersecting line; sequentially connecting the equal division points of the corresponding sequences on the p intersecting lines to obtain q connecting lines; the intersection points of the p intersecting lines and the q connecting lines are discrete points of the glass profile.
3. The method of claim 2, wherein the separately extracting discrete points of each glass profile and obtaining geometric feature data for each of the discrete points further comprises:
(1) Acquiring the coordinates of each discrete point under a coordinate system corresponding to the glass molded surface;
(2) For each discrete point, a normal vector S is obtained as follows:
obtaining a tangent vector S with discrete points parallel to the x-axis x And a tangent vector S of the discrete point parallel to the y-axis y ；
Normal vector S is obtained by tangent vector cross multiplication s=s x S y ；
(3) For each discrete point, the average curvature H is obtained as follows:wherein r is 1 Is the maximum radius of curvature through the discrete points; r is (r) 2 Is the minimum radius of curvature through the discrete points;
(4) For each discrete point, a gaussian curvature G is obtained as follows:wherein r is 1 Is the maximum radius of curvature through the discrete points; r is (r) 2 Is the minimum radius of curvature through the discrete points;
(5) For each discrete point, arch height AH is obtained according to the following method;
the camber is the distance from the discrete point to the corresponding bisector:
wherein the discrete point P is at the intersection lineUpper partIs associated withCorresponding bisectors.
4. The method of claim 1, wherein the obtaining a normal distance of each discrete point to a corresponding mold profile comprises:
for each discrete point, the normal distance of the discrete point is obtained in the following way:
acquiring an intersection point of a straight line passing through the discrete point and parallel to the normal vector of the discrete point and a corresponding mold surface;
and taking the distance between the discrete point and the intersection point as the normal distance from the discrete point to the corresponding mold surface.
5. A method according to claim 3, wherein said modeling a convolutional neural network based on geometric feature data of all discrete points and normal distances of the discrete points to corresponding mold surfaces comprises:
preprocessing the coordinates of all the discrete points by adopting a MinMax normalization mode;
preprocessing normal vectors, average curvatures, gaussian curvatures and arch heights of all discrete points by adopting a Robust normalization mode;
constructing a data set by utilizing the preprocessed geometric feature data of the discrete points and the normal distance from the discrete points to the corresponding mold surface, wherein the coordinates, normal vector, average curvature, gaussian curvature and camber of the discrete points are used as input variables of the data set, and the normal distance from the discrete points to the corresponding mold surface is used as output variables of the data set;
A convolutional neural network model between the input variable and the output variable is obtained using all the data in the dataset.
6. The method of claim 5, wherein the preprocessing of the coordinates of all discrete points using MinMax normalization comprises:
the coordinates of each discrete point are preprocessed according to the following method:
wherein (1)>Maximum value in all discrete point coordinates; />For the minimum value in all discrete point coordinates, +.>Is the original discrete point coordinates; />Is the discrete point coordinates after pretreatment.
7. The method of claim 5, wherein the preprocessing of the normal vector, the average curvature, the gaussian curvature and the camber of all discrete points by the Robust normalization method comprises:
the normal vector, average curvature, gaussian curvature and camber of each discrete point are all pre-processed as follows:
wherein (1)>The geometric characteristic data value is an original geometric characteristic data value, wherein the geometric characteristic data value is a geometric characteristic data value of one of normal vector, average curvature, gaussian curvature and camber of the discrete points; />Is->Corresponding preprocessed geometrical characteristic data values; / >For all discrete points and +.>The median of the corresponding geometric feature data values; IQR is the sum of all discrete points and +.>The interval length between the 1 st quartile and the 3 rd quartile in the corresponding geometric feature data value.
8. The method of claim 5, wherein the obtaining a convolutional neural network model between the input variable and the output variable using all data in the dataset comprises:
any one of all the glass molded surfaces is selected as a test glass molded surface, and the rest glass molded surfaces are all used as training glass molded surfaces;
constructing a test data set by utilizing the geometric characteristic data of the discrete points after the pretreatment of the test glass molded surface and the normal distance from the discrete points on the test glass molded surface to the corresponding mold molded surface, wherein the coordinates, the normal vector, the average curvature, the Gaussian curvature and the camber of the discrete points are used as input variables, and the normal distance from the discrete points to the corresponding mold molded surface is used as output variables;
for each training glass profile, a training dataset was constructed using the following method:
mapping the three-dimensional space information of the training glass molded surface onto a two-dimensional plane to obtain a two-dimensional glass molded surface, and simultaneously mapping discrete points on the three-dimensional glass molded surface onto the two-dimensional plane;
Starting with the upper left corner on the two-dimensional glass profile, using preset dimensionsThe sliding window of the training glass profile is used for obtaining all discrete points mapped in the range of the sliding window and used as a sample corresponding to the training glass profile in a training data set; taking all newly acquired discrete points as one sample corresponding to the training glass molded surface in the training data set after the sliding window slides once, until the sliding window acquires all the discrete points mapped on the two-dimensional glass molded surface;
constructing a training data set by using all samples, wherein coordinates, normal vectors, average curvature, gaussian curvature and camber of all discrete points in each sample are used as input variables, and the dimension of the input variables isThe method comprises the steps of carrying out a first treatment on the surface of the Taking the normal distance from all discrete points in each sample to the corresponding mold surface as an output variable, wherein the dimension of the output variable is；
And building a convolutional neural network model according to all the training data sets and the test data sets.
9. The method of claim 8, wherein the modeling the convolutional neural network from all training data sets and test data sets comprises:
constructing an initial model of a convolutional neural network model according to a training data set, wherein the initial model of the convolutional neural network model is composed of a convolutional block and a full connecting block, and the initial model of the convolutional neural network model comprises the following components:
The convolutional blocks of the convolutional neural network are designed as follows:
(1) Using m 1 Convolutional layers m 1 Extracting deep features by the batch normalization layer, and using a LeakyRelu function as an activation function; the LeakyRelu function is defined as:
wherein x is the activation value of a single neuron; λ is a preset slope when x is less than 0;
(2) The dimension of the input layer is consistent with the input dimension of the sample, and is；
(3) The output dimension of the ith convolution layer is；
Wherein,the number of convolution kernels in the ith convolution layer;the length/width of the data is output for the ith convolutional layer,the calculation formula of (2) is as follows:
wherein,is the firsti-1 length/width of convolutional layer output data,for the size of the convolution kernel in the ith convolution layer,for the fill length of the ith convolutional layer,step length for moving convolution kernel in ith convolution layer;
the fully connected block of the convolutional neural network is designed as follows:
(1) Using 1 spreading layer and m 2 Regression analysis was performed on all the connection layers, the first m 2 The activation function of 1 fully connected layer is the Relu function, the last fully connected layer does not use the activation function; the definition of the Relu function is:wherein x is the activation value of a single neuron;
(2) The output data of the convolution layer is unfolded through an unfolding layer to obtain a 1-dimensional vector with the length of The unfolding mode is as follows:
wherein,is the dimension of the convolution layer; the output dimension of the ith convolution layer isThen the mth 1 The output dimension of each convolution layer isThe method comprises the steps of carrying out a first treatment on the surface of the Roll upThe output dimension of the laminate is understood to bePersonal (S)Then (2) a two-dimensional matrix ofFor the 1 st row and 1 st column values in the first two-dimensional matrix,for row 1 of the first two-dimensional matrixThe value of the column is used to determine,for the first two-dimensional matrixLine 1Values of columns; and so on,is the firstThe values of row 1 and column 1 in a two-dimensional matrix,is the firstLine 1 of the two-dimensional matrixThe value of the column is used to determine,is the firstThe first two-dimensional matrixLine 1Values of columns;
（3）m 2 the input dimension of each full-connection layer is equal to the output dimension of the previous layer; wherein the input dimension of the 1 st full connection layer isMth, m 2 The output dimension of each full connection layer is。
10. The method of claim 9, wherein building a convolutional neural network model from the training dataset and the test dataset, further comprises:
and obtaining a finally established convolutional neural network model by adopting a plurality of loss functions according to the test data set and the initial model, wherein the loss functions comprise: average squared error loss, absolute value loss, relative entropy loss, style loss, and relevance loss functions; wherein:
The average square error loss, the absolute value loss and the relative entropy loss are all indexes of the estimation precision of the evaluation point, and the specific formulas are as follows:
where MSL represents the average squared error loss, L1 represents the absolute value loss, KLDiv represents the relative entropy loss,representing the actual value of the normal distance from the ith discrete point to the corresponding mold surface in the test dataset,the predicted value of the normal distance from the i discrete point to the corresponding mold surface is predicted according to the geometric feature data of the convolutional neural network model and the i discrete point, and N represents the total number of the discrete points in all samples of the test data set;
the style loss and the correlation loss are indexes for evaluating the estimation precision of the two-dimensional plane;
the style loss is calculated by using the difference value of the vector inner product, and the calculation mode is as follows:
wherein Gram represents style loss;
the correlation loss represents the correlation degree of a transverse vector and a longitudinal vector in the sliding window, wherein the transverse vector is a set of each row in the sliding window, and the longitudinal vector is a set of each column in the sliding window; the correlation coefficient between vectors is calculated by using a Pearson correlation coefficient, and the calculation mode is as follows:
Where Corr represents the correlation loss, m is the total number of samples in the test dataset,andrepresenting the true values on the two-dimensional plane in the kth sample, respectivelyAnd predicted valueThe matrix is formed, K represents the matrixIs a dimension of (2); wherein the matrixSum matrixThe forms of (a) are as follows:
the row and column vectors of matrix Y are denoted +.>；。
11. The method of claim 10, wherein said obtaining a final established convolutional neural network model using a plurality of loss functions based on the test dataset and the initial model comprises:
the average square error loss, absolute value loss, relative entropy loss, style loss and correlation loss function are weighted and averaged to obtain a total loss function;
and (3) based on the total loss function, re-determining parameters in the convolutional neural network model by adopting a gradient descent method to obtain the finally established convolutional neural network model.
12. The method of claim 11, wherein predicting the normal distance of each discrete point on the desired glass profile to the desired mold profile based on the geometric feature data of the discrete points on the desired glass profile and the convolutional neural network model comprises:
preprocessing the coordinates of all discrete points on the required glass molded surface in a MinMax normalization mode;
Preprocessing normal vectors, average curvatures, gaussian curvatures and camber of all discrete points on a required glass molded surface in a Robust normalization mode;
substituting the geometric feature data of all discrete points of the required glass molded surface after pretreatment into a convolutional neural network model as input data, and calculating to obtain a predicted value of the normal distance from each discrete point on the required glass molded surface to the required mold molded surface;
determining whether there are discrete points on the desired glass profile having a plurality of predicted values,
if so, the maximum value in the plurality of predicted values is taken as the final predicted value of the corresponding discrete point.
13. The method of any one of claims 1 to 12, wherein generating the profile of the desired mold using geometric feature data of discrete points on the desired glass profile and a normal distance from the discrete points on the desired glass profile to the desired mold profile comprises:
calculating coordinates of each discrete point of the required mold surface by utilizing geometric feature data of the discrete points on the required glass surface and normal distance from the discrete point of the required glass surface to the required mold surface, wherein the discrete points of the required mold surface correspond to the discrete points on the required glass surface one by one and are assumed discrete points on the required mold surface;
The coordinates of each discrete point of the desired mold profile are calculated as follows:
wherein:coordinates of discrete points of the required mold surface;to demand the seating of corresponding discrete points on the glass profileMarking;normal vector of corresponding discrete point on the required glass profile;the normal distance from the corresponding discrete point on the predicted required glass molded surface to the required mold molded surface is set;
constructing a spline curve network wire frame connected with discrete points according to the coordinates of the discrete points of the required mold surface;
and (3) carrying out curved surface reconstruction on the spline curve network wire frame by using a geometric modeling engine to generate the molded surface of the required mold.
14. An automobile windshield mold profile design device based on a convolutional neural network, which is applied to generating a required mold profile according to a required glass profile, and is characterized by comprising:
the glass profile and mould profile acquisition unit is used for acquiring a plurality of groups of profiles from the historical design scheme, wherein each group of profiles comprises a glass profile and a corresponding mould profile;
the glass profile discrete point data acquisition unit is used for respectively extracting discrete points of each glass profile and acquiring geometric characteristic data of each discrete point;
the normal distance acquisition unit is used for acquiring the normal distance from each discrete point to the corresponding mold surface;
The convolution neural network model building unit is used for building a convolution neural network model according to the geometric feature data of all the discrete points and the normal distance from the discrete points to the corresponding mold surfaces, wherein the input variable of the convolution neural network model is the geometric feature of the discrete points, and the output variable is the normal distance from the discrete points to the corresponding mold surfaces;
the glass profile discrete point data acquisition unit is also used for extracting discrete points requiring the glass profile and acquiring geometric characteristic data of each discrete point;
the normal distance prediction unit is used for predicting the normal distance from each discrete point on the required glass profile to the required mold profile according to the geometric feature data of the discrete points of the required glass profile and the convolutional neural network model;
the required mold profile generating unit is used for generating the profile of the required mold by utilizing the geometric characteristic data of the discrete points of the required glass profile and the normal distance from the discrete points on the required glass profile to the required mold profile.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211071143.2A CN115600487B (en) | 2022-09-02 | 2022-09-02 | Method and device for designing mold surface of automobile windshield glass based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211071143.2A CN115600487B (en) | 2022-09-02 | 2022-09-02 | Method and device for designing mold surface of automobile windshield glass based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115600487A CN115600487A (en) | 2023-01-13 |
| CN115600487B true CN115600487B (en) | 2024-02-02 |
Family
ID=84843712
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211071143.2A Active CN115600487B (en) | 2022-09-02 | 2022-09-02 | Method and device for designing mold surface of automobile windshield glass based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115600487B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004145674A (en) * | 2002-10-25 | 2004-05-20 | Nippon Sheet Glass Co Ltd | Method for designing mold surface of press bending form block |
| CN102169521A (en) * | 2011-04-25 | 2011-08-31 | 天津职业技术师范大学 | Design method for molded surface of automobile cover part die based on molded surface deformation compensation |
| CN103617302A (en) * | 2013-10-17 | 2014-03-05 | 中南大学 | Method for building accurate molded surface of automobile covering part drawing mold |
| CN104200054A (en) * | 2014-09-26 | 2014-12-10 | 山东山大华天软件有限公司 | Gap design method used for automobile covering part die |
| CN104573281A (en) * | 2015-01-29 | 2015-04-29 | 中南大学 | Complex space curved surface thin plate forming die face designing method taking springback compensation |
| CN109822295A (en) * | 2019-03-29 | 2019-05-31 | 鹤壁职业技术学院 | The high accurate numerical-control processing method of magnesium alloy automobile panel outside plate die face |
| CN112396690A (en) * | 2020-11-11 | 2021-02-23 | 大连理工大学 | Curved surface high-precision reconstruction method based on improved centripetal parameterization method |
| CN114972387A (en) * | 2022-07-28 | 2022-08-30 | 南京航空航天大学 | Three-dimensional actual measurement-based mould deformation repairing method and system in composite material forming process |
-
2022
- 2022-09-02 CN CN202211071143.2A patent/CN115600487B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004145674A (en) * | 2002-10-25 | 2004-05-20 | Nippon Sheet Glass Co Ltd | Method for designing mold surface of press bending form block |
| CN102169521A (en) * | 2011-04-25 | 2011-08-31 | 天津职业技术师范大学 | Design method for molded surface of automobile cover part die based on molded surface deformation compensation |
| CN103617302A (en) * | 2013-10-17 | 2014-03-05 | 中南大学 | Method for building accurate molded surface of automobile covering part drawing mold |
| CN104200054A (en) * | 2014-09-26 | 2014-12-10 | 山东山大华天软件有限公司 | Gap design method used for automobile covering part die |
| CN104573281A (en) * | 2015-01-29 | 2015-04-29 | 中南大学 | Complex space curved surface thin plate forming die face designing method taking springback compensation |
| CN109822295A (en) * | 2019-03-29 | 2019-05-31 | 鹤壁职业技术学院 | The high accurate numerical-control processing method of magnesium alloy automobile panel outside plate die face |
| CN112396690A (en) * | 2020-11-11 | 2021-02-23 | 大连理工大学 | Curved surface high-precision reconstruction method based on improved centripetal parameterization method |
| CN114972387A (en) * | 2022-07-28 | 2022-08-30 | 南京航空航天大学 | Three-dimensional actual measurement-based mould deformation repairing method and system in composite material forming process |
Non-Patent Citations (1)
| Title |
|---|
| 沙智华 ; 陈翠翠 ; 张生芳 ; .汽车玻璃曲面展开系统研究与开发.大连交通大学学报.(第05期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115600487A (en) | 2023-01-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wu et al. | Application of the Box–Behnken design to the optimization of process parameters in foam cup molding | |
| Rayegani et al. | Fused deposition modelling (FDM) process parameter prediction and optimization using group method for data handling (GMDH) and differential evolution (DE) | |
| CN107704697B (en) | A kind of profile three-dimensional bending formability prediction and evaluation optimization method | |
| CN104794256A (en) | Springback compensation in formed sheet metal parts | |
| Lee et al. | Shape optimization of the workpiece in the forging process using equivalent static loads | |
| CN104268349A (en) | Method for accurately controlling trimming line of turned edge under complex curved surface | |
| Lu et al. | Shape optimisation of preform design for precision close-die forging | |
| Alimirzaloo et al. | A novel method for preform die design in forging process of an airfoil blade based on Lagrange interpolation and meta-heuristic algorithm | |
| CN108108582A (en) | A kind of method for numerical simulation of curved-surface piece flexible rolling forming process | |
| CN115600487B (en) | Method and device for designing mold surface of automobile windshield glass based on convolutional neural network | |
| CN115600488B (en) | Method and device for designing mold surface of automobile windshield glass mold based on encoder-decoder model | |
| CN111014415A (en) | Method for manufacturing draw bead based on CAE technology | |
| CN115600366B (en) | Automobile windshield mold profile design method and device based on regression model | |
| CN109918785A (en) | A kind of large-scale complex thin-wall titanium alloy member hot forming corrugation prediction and control method | |
| CN111931117B (en) | Rapid prediction method for removal rate of spiral curved surface grinding material | |
| CN107153727B (en) | Tolerance distribution method and device for flexible thin-wall structure based on deformation base | |
| CN115600309B (en) | Method and device for designing mold surface of automobile windshield based on curved surface reconstruction | |
| CN112784456A (en) | Design method and system for automobile glass forming die surface | |
| CN110765693A (en) | Design method of noble metal tantalum capacitor shell optimization die based on CAE simulation technology | |
| CN114163112A (en) | Digital design method of hot-press forming die for automobile rear windshield glass | |
| CN112364549B (en) | Method for establishing rolling variable-thickness plate forming limit field | |
| CN1595404A (en) | Digitalized design method of rubber product extrusion die | |
| JP7276584B1 (en) | Springback amount evaluation method, apparatus and program for press-formed product, and method for manufacturing press-formed product | |
| JP7184671B2 (en) | How to set the shape of the press mold | |
| CN116933442B (en) | Stamping process simulation model modeling method and related device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |