WO2022007409A1 - 一种基于图像特征学习的曲线加筋结构布局智能设计方法 - Google Patents

一种基于图像特征学习的曲线加筋结构布局智能设计方法 Download PDF

Info

Publication number
WO2022007409A1
WO2022007409A1 PCT/CN2021/077160 CN2021077160W WO2022007409A1 WO 2022007409 A1 WO2022007409 A1 WO 2022007409A1 CN 2021077160 W CN2021077160 W CN 2021077160W WO 2022007409 A1 WO2022007409 A1 WO 2022007409A1
Authority
WO
WIPO (PCT)
Prior art keywords
network model
image
curve
neural network
training
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.)
Ceased
Application number
PCT/CN2021/077160
Other languages
English (en)
French (fr)
Inventor
郝鹏
张坤鹏
刘大川
王博
李刚
段于辉
石云峰
王宇同
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian University of Technology
Original Assignee
Dalian University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Dalian University of Technology filed Critical Dalian University of Technology
Priority to US17/424,785 priority Critical patent/US20220138582A1/en
Publication of WO2022007409A1 publication Critical patent/WO2022007409A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/17—Mechanical parametric or variational design
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/12—Computing arrangements based on biological models using genetic models
    • G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06V10/7753—Incorporation of unlabelled data, e.g. multiple instance learning [MIL]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/06—Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA]

Definitions

  • the invention belongs to the field of engineering thin-walled reinforced structure design, in particular to an intelligent design method of curved reinforced structure layout based on image feature learning.
  • the layout design of curved reinforcement will make the stiffness distribution and loading path of the reinforced structure more flexible, and improve the bearing efficiency of the structure. Therefore, it has become a research hotspot in the fields of launch vehicles, aircraft, and ships.
  • the path characterization function of the curved reinforced structure is more complex, which leads to the explosive growth of design variables, which seriously restricts the optimal design of the layout of the curved reinforced structure, especially for curves with dynamic changes in the number of design variables.
  • the structural optimization design method based on the traditional surrogate model is more difficult to carry out.
  • the present invention proposes an intelligent design method of the layout of the curved reinforced structure based on image feature learning.
  • image feature learning By building a deep learning network, the structural features of the curved path layout image are extracted, and the curve is further optimized.
  • the layout design of reinforced structure solves the difficulties faced by traditional optimization methods and provides an effective and feasible method for related fields.
  • An intelligent design method for curved reinforced structure layout based on image feature learning comprising the following steps:
  • Step 100 Select the curve reinforcement path function to generate an image set, input the self-encoding network to perform unsupervised learning training, and complete the extraction of the structural features of the curved tendon image, including the following sub-steps:
  • Step 101 Select the path function B(t), and determine the design variables of the path function for the reinforced thin-walled structure, as shown in formula (1.1);
  • B(t) is the path function
  • t is the control variable of the path function
  • P s (x s , y s ) is the coordinate of the starting point of the path
  • P m (x m , y m ) is the coordinate of a point in the path
  • P e (x e , y e ) are the coordinates of the end point of the path
  • Step 102 define the path function of the curved stiffened structure, specifically: determine the path function type according to the combination of different boundary types of the structure, and constrain the design domain space of the curved stiffened structure path function;
  • Step 103 generating an image set of the curvilinear reinforcement structure layout, specifically: determining the size m*n of each curve reinforcement structure image, and generating a training image set N 0 used for unsupervised learning training;
  • Step 104 Build a decoding network model E and an encoding network model D for the curvilinear reinforcement structure layout image
  • Step 105 Combine the image decoding network model E and the encoding network model D to form a self-encoding network model;
  • Step 106 Input the curve reinforcement layout image set N 0 into the self-encoding network model
  • Step 107 Complete the training process of the self-encoding network model on the curve reinforcement image set N 0;
  • Step 108 extract the decoding network model E trained by the self-encoding network model
  • Step 200 establish an analysis model of the mechanical response of the curved stiffened structure, form a data set for supervised learning and training, and further input the convolutional neural network model constructed by the decoding network model and the fully connected layer in step 108 to complete the curved stiffened structure
  • the learning of mechanical response includes the following sub-steps:
  • Step 201 establish a curve stiffening structure model according to the curve path function B(t);
  • Step 202 setting structural displacement load boundary conditions, and performing structural mechanical response analysis
  • Step 203 Generate training set and test set images for training and testing of the learning model, specifically: determine the size m*n of each curve-reinforced structural image, and generate a supervised learning model according to the corresponding structural mechanical response of the image
  • the training set N 1 and the test set N 2 for training and testing additionally set the evaluation criteria for the quality of the model, as shown in formula (1.2), select the root mean square (%RMSE) as the error evaluation of the model;
  • n is the number of samples
  • y i is the structural response value, is the predicted value of the model
  • Step 204 build a convolutional neural network model F by the decoding network model E and two fully connected layers in step 108;
  • Step 205 the mechanical response of the training set containing N 1 of the input label convolutional neural network model in the training F;
  • Step 206 The test set of N 2 F convolutional neural network model accuracy of determination, the completion of the training process of the neural network convolution response curve of the mechanical reinforcement structure;
  • Step 300 Based on the convolutional neural network model F for predicting the mechanical response of the curved stiffened structure in step 206, use an evolutionary algorithm to complete the optimal design of the curved stiffened structure layout, including the following sub-steps:
  • Step 301 Build an evolutionary algorithm optimization framework, and first generate an initial curve reinforcement image set N g at the beginning of the optimization iteration;
  • Step 302 Input the image set N g into the convolutional neural network model F extracted in step 206;
  • Step 303 use an evolutionary algorithm to search for optimization on the established convolutional neural network model F to obtain a new sample point K;
  • Step 304 establish a curve stiffening structure model from the obtained sample point K, and mark it through mechanical response analysis;
  • Step 305 supplement the new sample point K to the training image set N g to form an image set N g+k , and further input the convolutional neural network model F in step 206 for retraining;
  • Step 306 Retrain the convolutional neural network model Instead of the convolutional neural network model F in step 302, continue to carry out the optimization process of the evolutionary algorithm;
  • Step 307 Determine whether the current optimization process has reached the algorithm convergence condition, if it converges, output the optimal design variables, otherwise, return to step 301, wherein the convergence condition is reaching the maximum iteration number of the optimization algorithm.
  • the selected path function needs to constrain that the curvature of the function cannot be too large and the intermediate path of the function cannot exceed the plate design area, including but not limited to spline functions.
  • the pixel size of the structure image in the image set is not fixed, and can be adjusted by itself according to the complexity of the specific research structure.
  • the network structure and hyperparameter settings used to build the self-encoding network model can be adjusted according to specific research problems.
  • the mechanical response of the structure includes response characteristics such as static force, dynamic force or structural buckling
  • the analysis method used may be finite element analysis method, boundary element analysis method, isogeometric analysis method and meshless analysis method. method and similar analysis methods.
  • the setting of the number of samples in the generated training set and test set can be adjusted according to the research problem, and the model error evaluation used needs to be global, including but not limited to %RSME.
  • the error of the convolutional neural network model will gradually converge as the number of training steps increases, and the setting of the number of training steps can be adjusted according to the overall optimization efficiency of the overall optimization problem complexity and the model convergence speed.
  • step 301 includes: genetic algorithm, simulated annealing algorithm, artificial neural network algorithm, particle swarm algorithm, ant colony algorithm and other similar optimization methods.
  • step 301 to step 307 it is necessary to optimize the fixed number of ribs and the variable number of ribs respectively.
  • the convolutional neural network has completed the learning process of the structural feature mechanical response of the curved tendon image, and there is no need to generate additional training sets of variable tendons. It is only necessary to redo the optimization process from step 301 to step 307 based on the program code of variable tendons.
  • the optimal design of the curve reinforcement layout with dynamically variable number of ribs can be realized.
  • the beneficial effects of the present invention are as follows: an intelligent design method for curved reinforced structure layout based on image feature learning is proposed, a convolutional neural network model is constructed from the curve path representation image to the structural mechanical response prediction, and the model is further applied to the curve In the optimization design of reinforced structure layout.
  • the deep learning network model based on the curve-stiffened image has better structural response prediction effect, and the curve layout optimization carried out in this way obtains a feasible optimal solution.
  • the present invention is expected to be one of the most potential methods involving optimal design of component layouts in engineering structures.
  • Fig. 1 is the realization flow chart of the intelligent design method of curve reinforcement layout based on image feature learning provided by the embodiment of the present invention
  • Fig. 2 is the structure diagram of the image feature learning network
  • Fig. 2 (a) is the self-encoding network model structure that encodes and decodes the image
  • the resulting convolutional neural network model, the numerical change in the figure represents the image size change of the input image after layer-by-layer processing
  • Fig. 3 is the convergence diagram of the self-encoding network model training process
  • Fig. 4 is a picture showing the effect of self-encoding 20,000 steps of self-learning and training on the inputted curve-stiffened image;
  • Fig. 4(a) is the input image of the curve-stiffened structure, and
  • Fig. 4(b) is the self-encoding network model learning and training The output image after;
  • Figure 5 is a schematic diagram of the boundary conditions of the calculation example; the numerical values in the figure represent the magnitude of the load;
  • Fig. 6 is the weight reduction optimization process of the image sample set with the fixed number of ribs
  • Fig. 7 shows the quality lightweight optimization process of the image sample set with variable number of ribs.
  • FIG. 1 is a flowchart for realizing an image feature learning intelligent design method for a curvilinear reinforced structure layout provided by an example of the present invention.
  • the images involved in the present invention and the deep learning network are all generated based on the TensorFlow environment of the Python language.
  • the image feature learning process provided by the embodiment of the present invention to intelligently design the layout of the curved reinforced structure includes:
  • Step 100 Select the quadratic Bessel spline function as the reinforced path function, constrain the design variables of the path function according to the design domain space of the reinforced thin-walled structure, generate an image set for unsupervised learning training, and further input the
  • the self-encoding network built by multiple convolutional layers and pooling layers is trained to obtain the self-encoding network model for image structure feature extraction, including the following sub-steps:
  • Step 101 Determine the control parameters of the stiffening path function based on the quadratic Bessel spline function, as shown in formula (1.1), where B(t) is the path function, t is the path function control variable, and P s (x s , y s ) the coordinates of the starting point of the path, P m (x m , y m ) are the coordinates of a point in the path, and P e (x e , y e ) are the coordinates of the end point of the path;
  • Step 102 Determine six types of reinforcement paths according to different boundary combinations of the starting point and the end point of the curvilinear reinforcement path, select four types for combination to obtain a curvilinear reinforcement structure controlled by 20 variables, and then according to the reinforcement thin-wall structure The design space of , restricts the stiffening variables;
  • Step 103 Generate 10,000 image sets N 0 representing the path layout based on the curve path function type, and set the size of each image to 64*64;
  • Step 104 Build an image decoding network model E with three convolution layers and three pooling layers, build an image encoding network model D with three convolution layers, and combine to form a self-encoding network model for image self-learning.
  • the specific network model structure is as follows: As shown in Figure 2(a);
  • Step 105 Adjust the hyperparameters in the self-encoding network model, such as: learning rate 0.001, convolution kernel size 3*3, data input batch 100, number of training steps 20000, etc., and determine the type of Loss function, as shown in formula (1.3) shown, where N is the data training input batch, o (n) is the input image of the self-encoding network, and y (n) is the output image of the self-encoding network;
  • Step 106 Input the 10,000 curve-stiffened image sets N 0 in step 103 into batches for training the self-encoding network model in step 105;
  • Step 107 Complete the image training process.
  • the training process is shown in Figure 3. After 20,000 steps of self-encoding network model training, the training effect of image self-learning is shown in Figure 4;
  • Step 108 extract the decoding network model E trained by the self-encoding network model
  • Step 200 Create a finite element model according to the curve path type function, perform structural linear buckling analysis to obtain a data set for supervised learning training, and further input the convolutional neural network model constructed by the decoding network model and two fully connected layers in step 108 , after training, the learning process of the response by the structural quality and buckling eigenvalues is completed, including the following sub-steps:
  • Step 201 According to the curve reinforcement path function determined in Step 101, a finite element numerical model with variable number of ribs and fixed number of ribs is established by ABAQUS commercial software.
  • the reinforced thin-walled structure in this example is a flat plate with a size of 629.6*731.2mm, the thickness of the skin is 1.5mm, the height and width of the rib are 18.0mm and 2.4mm, respectively, the structural material is aluminum 2139, and its elastic modulus The amount was 72.50GPa, the Poisson's ratio was 0.3, and the density was 2.8e-6Kg/mm 3 .
  • Step 202 As shown in Figure 5, set the four-side simply supported displacement boundary condition, set the axial-shear combined load boundary condition, apply unit 1 shear force on the four sides, apply unit 1 unit axial force on the upper and lower sides, and apply uneven axial force on the left and right sides, as shown in the formula (1.4) and Equation (1.5), where P left is the left axial force, P right is the right axial force, and l is the height of the curvilinear stiffened plate, further completing the finite element linear buckling analysis of the curvilinear stiffened structure;
  • Step 203 Use the Latin hypercube method to independently sample 5 times in the design domain space, generate 5 groups of 250 curvilinear reinforced structure images with labels (mass, buckling eigenvalues), and select a set of images as the training set N 1 , the other four groups are used as the test set for cross-checking N 2 , and the root mean square error (%RMSE) is selected as the error evaluation of the model, as shown in formula (1.2), where n is the number of samples, y i is the structural response value, is the predicted value of the model;
  • Step 204 build a convolutional neural network model from the decoding network model E extracted in step 108 and two fully connected layers;
  • Step 205 Feed the training set containing labels into the convolutional neural network model for training, set the learning rate to 0.005, the data input batch to 100, and the number of training steps to 1000.
  • the training process only adjusts the parameters in the last two fully connected layers;
  • Step 206 complete the training process of multiple sets of images, and extract the trained convolutional neural network model
  • Step 300 Based on the convolutional network model for predicting the quality and buckling eigenvalues of the curvilinear reinforced structure in step 206, using the genetic optimization algorithm to select the buckling eigenvalue of the curvilinear reinforced structure not greater than 8.40 as the constraint condition, and carry out the structural weight reduction curve calculation.
  • Rib layout optimization design including the following sub-steps:
  • Step 301 build a genetic algorithm optimization framework, the optimization iteration starts by setting the initial population number 150, the genetic algebra 15 generations, the maximum optimization times 50 times, and generating the initial population curve reinforcement image set N g , wherein the size of each image is 64 *64;
  • Step 302 Input the image set N g generated by the initial population into the convolutional neural network extracted in step 206;
  • Step 303 Use the convolutional neural network model to carry out the optimal design of the layout of the curved stiffening structure based on the genetic algorithm, and obtain a new sample point K by optimization;
  • Step 304 Establish a finite element model from the obtained sample point K, and perform structural linear buckling analysis and inspection to complete the marking of the image of the sample point;
  • Step 305 generate a 64*64 size image from the obtained sample point K, and add it to the training image set, and further input the expanded image set N g+k into the convolutional neural network in step 206 for retraining;
  • Step 306 Retrain the convolutional neural network model Substitute the convolutional neural network model F in step 302 to continue genetic algorithm optimization;
  • Step 307 determine whether the genetic algorithm optimization process has reached the convergence of the maximum number of iterations, and if so, output the optimal design variables and structural buckling eigenvalues; otherwise, return to step 305 .
  • the present invention designs a characteristic learning method of curved reinforced path representation image, fully excavates the structural information in the structural image, and constructs a convolutional neural network model for structural quality and buckling eigenvalues.
  • the predicted root mean square error of the response is all about 5%, which greatly guarantees the model accuracy in the layout design problem of the curvilinear reinforced structure.
  • the optimal result of the weight reduction based on the convolutional neural network is 0.100, and the weight reduction ratio reaches 24.8% .
  • the optimal weight reduction result of the variable rib is 0.0954, and the weight loss ratio is 28.3%.
  • the present invention is a deep learning method based on structural image feature extraction. Compared with the traditional surrogate model optimization method, the invention significantly improves the model accuracy in the multivariate complex structural optimization problem, and obtains a curved reinforcement layout with higher structural mechanics bearing efficiency. design.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Geometry (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Databases & Information Systems (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Multimedia (AREA)
  • Computational Mathematics (AREA)
  • Physiology (AREA)
  • Genetics & Genomics (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

一种基于图像特征学习的曲线加筋结构布局智能设计方法，属于工程薄壁加筋结构优化设计领域。首先基于路径函数确定曲线加筋结构的设计变量，通过搭建自编码网络完成对图像结构特征的学习，进一步进行模型的迁移学习，并搭建卷积神经网络完成对带有力学响应标签的图像集进行学习，最后基于该模型实现演化类算法对曲线加筋结构布局的优化设计。上述方法解决了传统优化方法难以处理设计变量众多且可变的优化设计问题，有望成为工程领域中涉及部件布局设计问题的最具潜力的技术手段之一。

Description

一种基于图像特征学习的曲线加筋结构布局智能设计方法 技术领域
本发明属于工程薄壁加筋结构设计领域，尤其涉及一种基于图像特征学习的曲线加筋结构布局智能设计方法。
背景技术
曲线加筋布局设计由于具有更大的结构设计空间，会使得加筋结构的刚度分布、加载路径更加灵活，提高结构的承载效率，因此成为运载火箭、飞机、船舶等工程领域的研究热点。然而相对于传统的直线加筋结构，曲线加筋结构的路径表征函数更加复杂，导致设计变量爆炸式增长，进而严重地制约了曲线加筋结构布局优化设计，尤其对于设计变量数目动态变化的曲线加筋结构，基于传统代理模型的结构优化设计方法更加难以开展。
发明内容
针对曲线加筋结构布局优化设计中出现的诸多难点，本发明提出一种基于图像特征学习的曲线加筋结构布局智能设计方法，通过搭建深度学习网络提取曲线路径布局图像的结构特征，进一步优化曲线加筋结构布局设计，解决了传统优化方法所面临的困难，并为相关领域提供一种有效可行的方法。
为了达到上述目的，本发明采用的技术方案如下：
一种基于图像特征学习的曲线加筋结构布局智能设计方法，包括以下步骤：
步骤100：选取曲线加筋路径函数生成图像集，输入自编码网络进行无监督学习训练，完成曲筋图像结构特征的提取，包括以下子步骤：
步骤101：选取路径函数B(t)，并确定加筋薄壁结构路径函数设计变量，如式(1.1)所示；
B(t)＝(1-t) 2P s(x s,y s)+2t(1-t)P m(x m,y m)+t 2P e(x e,y e),t∈[0,1]      (1.1)
其中，B(t)为路径函数，t为路径函数控制变量，P s(x s，y s)为路径起点坐标，P m(x m，y m)为路径内一点坐标，P e(x e，y e)为路径终点坐标；
步骤102：对曲线加筋结构路径函数进行限定，具体的：根据结构不同边界类型组合确定路径函数类型，并对曲线加筋结构路径函数的设计域空间进行约束；
步骤103：生成曲线加筋结构布局的图像集，具体的：确定每张曲线加筋结构图像的大小m*n，生成无监督学习训练所用的训练图像集N 0；
步骤104：搭建对曲线加筋结构布局图像的解码网络模型E和编码网络模型D；
步骤105：将图像解码网络模型E和编码网络模型D进行组合，形成自编码网络模型；
步骤106：将曲线加筋布局图像集N 0输入自编码网络模型；
步骤107：完成自编码网络模型对曲线加筋图像集N 0的训练过程；
步骤108：提取自编码网络模型训练后的解码网络模型E；
步骤200：建立曲线加筋结构力学响应的分析模型，形成用于有监督学习训练的数据集，进一步输入由步骤108解码网络模型和全连接层搭建的卷积神经网络模型，完成曲线加筋结构力学响应的学习，包括以下子步骤：
步骤201：根据曲线路径函数B(t)建立曲线加筋结构模型；
步骤202：设置结构位移载荷边界条件，进行结构力学响应分析；
步骤203：生成用于学习模型训练及检验的训练集、测试集图像，具体的：确定每张曲线加筋结构图像的大小m*n，根据图像对应结构力学响应，生成用于有监督学习模型训练及检验的训练集N 1、检验集N 2；另外设置用于模型优劣的评价标准，如式(1.2)所示，选取均方根(％RMSE)作为模型的误差评估；
Figure PCTCN2021077160-appb-000001
其中，n为样本数量，y i为结构响应值，
Figure PCTCN2021077160-appb-000002
为模型预测值；
步骤204；由步骤108的解码网络模型E和两个全连接层搭建卷积神经网络模型F；
步骤205：将含力学响应标签的训练集N 1输入卷积神经网络模型F中进行训练；
步骤206：根据检验集N 2对卷积神经网络模型F的准确性进行判定，完成卷积神经网络对曲线加筋结构力学响应的训练过程；
步骤300：基于步骤206对曲线加筋结构力学响应预测的卷积神经网络模型F，利用演化类算法完成曲线加筋结构布局的优化设计，包括以下子步骤：
步骤301：搭建演化类算法优化框架，优化迭代起始首先生成初始曲线加筋图像集N g；
步骤302：将图像集N g输入由步骤206提取的卷积神经网络模型F中；
步骤303：利用演化类算法在所建立的卷积神经网络模型F上寻优获得新样本点K；
步骤304：由获得的样本点K建立曲线加筋结构模型，并通过力学响应分析进行标记；
步骤305：将新样本点K补充到训练图像集N g形成图像集N g+k，进一步输入步骤206中的卷积神经网络模型F进行重新训练；
步骤306：将重新训练的卷积神经网络模型
Figure PCTCN2021077160-appb-000003
代替步骤302中的卷积神经网络模型F，继续开展演化类算法的优化过程；
步骤307：判断当前优化过程是否达到算法收敛条件，如果收敛，输出最优设计变量，否则，返回执行步骤301，其中所述的收敛条件为达到优化算法的最大迭代次数。
进一步，所述的步骤101中，选取的路径函数需要约束函数曲率不能过大且函数中间路径不能超出平板设计区域，包括但不限于与样条函数。
进一步，所述的步骤103中，所述图像集中结构图像像素大小不固定，可根据具体研究结构的复杂性自行调整。
进一步，所述的步骤104和步骤105中，搭建自编码网络模型所用的网络结构及超参数均设置可根据具体的研究问题自行调整。
进一步，所述的步骤202中，结构的力学响应包括静力、动力或结构屈曲 等响应特征，所用到的分析方法可以是有限元分析法、边界元分析法、等几何分析法和无网格法等类似分析方法。
进一步，所述的步骤203中，生成的训练集、测试集中的样本数量的设定可根据研究问题自行调整，采用的模型误差评估需要具有全局性，包括但不局限于％RSME。
进一步，所述的步骤205中，卷积神经网络模型随着训练步数的增加误差会逐渐地收敛，训练步数的设置可根据整体优化问题复杂性及模型收敛速度的综合优化效率自行调整。
进一步，所述的步骤301中的演化类算法包括：遗传算法，模拟退火算法，人工神经网络算法，粒子群算法以及蚁群算法等类似优化方法。
进一步，所述的步骤301至步骤307过程，需要对固定数目筋条和可变数目筋条分别进行优化，在可变数目筋条结构开展优化布局设计过程中，由于步骤100和步骤200形成的卷积神经网络已对曲筋图像完成结构特征力学响应的学习过程，无需额外生成可变筋条的训练集，只需基于可变筋条的程序代码重新开展步骤301至步骤307的优化过程，即可实现筋条数目动态可变的曲线加筋布局优化设计。
本发明的有益效果是：提出一种基于图像特征学习的曲线加筋结构布局智能设计方法，搭建了由曲线路径表征图像到结构力学响应预测的卷积神经网络模型，进一步将该模型应用在曲线加筋结构布局的优化设计中。相较于传统的代理模型优化方法，基于曲线加筋图像的深度学习网络模型具有更好的结构响应预测效果，且以此开展的曲线布局优化得到可行的最优解。本发明有望成为工程结构中涉及部件布局优化设计问题的最具潜力的方法之一。
附图说明
图1为本发明实施例提供的基于图像特征学习的曲线加筋布局智能设计方法的实现流程图；
图2为图像特征学习网络的结构图；图2(a)为对图像进行编码及解码的 自编码网络模型结构；图2(b)为将自编码中的编码部分网络与全连接网络连接而成的卷积神经网络模型，图中的数值变化表示输入图像经过逐层处理的图像大小变化；
图3为自编码网络模型训练过程的收敛图；
图4为对输入的曲线加筋图片经过自编码20000步自学习训练后的效果展示图；图4(a)为曲线加筋结构输入图像，图4(b)为经过自编码网络模型学习训练后的输出图像；
图5为算例边界条件示意图；图中数值代表载荷大小；
图6为固定筋条数目图像样本集的质量轻量化优化过程；
图7为可变筋条数目图像样本集的质量轻量化优化过程。
具体实施方式
为使本发明解决的技术问题、采用的技术方案和达到的技术效果更为详尽，下面结合附图和实施例对本发明作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅用于解释本发明，而非对本发明的限定。另外还需要说明的是，为了便于描述，附图中仅示出了本发明相关的部分而非全部内容。
图1为本发明实例提供的针对曲线加筋结构布局的图像特征学习智能设计方法实现流程图。本发明中涉及的图像以及深度学习网络均是基于Python语言的TensorFlow环境生成的，本发明实施例提供曲线加筋结构布局智能设计的图像特征学习过程包括：
步骤100：选取二次贝塞尔样条函数作为加筋路径函数，根据加筋薄壁结构设计域空间对路径函数的设计变量进行约束，生成用于无监督学习训练的图像集，进一步输入由多个卷积层和池化层搭建的自编码网络，经过训练得到对图像结构特征提取的自编码网络模型，包括以下子步骤：
步骤101：基于二次贝塞尔样条函数，确定加筋路径函数控制参数，具体如式(1.1)所示，其中B(t)为路径函数，t为路径函数控制变量，P s(x s，y s)路径起点坐标，P m(x m，y m)为路径内一点坐标，P e(x e，y e)为路径终点坐标；
B(t)＝(1-t) 2P s(x s,y s)+2t(1-t)P m(x m,y m)+t 2P e(x e,y e),t∈[0,1]      (1.1)
步骤102：根据曲线加筋路径起点和终点所在的不同边界组合，确定六种加筋路径类型，选取四种类型进行组合得到由20个变量控制的曲线加筋结构，再根据加筋薄壁结构的设计空间对加筋变量进行约束限定；
步骤103：基于曲线路径函数类型的生成10000张表征路径布局的图像集N 0，并且设置每张图像大小为64*64；
步骤104：由三个卷积层和三个池化层搭建图像解码网络模型E，由三个卷积搭建图像编码网络模型D，组合形成图像自学习的自编码网络模型，具体网络模型结构如图2(a)所示；
步骤105：调整自编码网络模型中的超参数，例如：学习率0.001、卷积核大小3*3、数据输入批量100、训练步数20000等，并确定Loss函数类型，如式(1.3)所示，其中N为数据训练输入批量，o (n)为自编码网络输入图像，y (n)为自编码网络输出图像；
Figure PCTCN2021077160-appb-000004
步骤106：将步骤103中的10000张曲线加筋图像集N 0进行分批次输入步骤105中的自编码网络模型训练；
步骤107：完成图像训练过程，训练过程如图3所示，经过20000步的自编码网络模型训练，图像自学习的训练效果如图4所示；
步骤108：提取自编码网络模型训练后的解码网络模型E；
步骤200：根据曲线路径类型函数创建有限元模型，进行结构线性屈曲分析得到用于有监督学习训练的数据集，进一步输入由步骤108解码网络模型和两 个全连接层搭建的卷积神经网络模型，经过训练完成由结构质量、屈曲特征值响应的学习过程，包括以下子步骤：
步骤201：根据步骤101确定的曲线加筋路径函数，通过ABAQUS商用软件建立变筋条数目、固定筋条数目的有限元数值模型。本例中的加筋薄壁结构的尺寸为629.6*731.2mm的平板，蒙皮厚度为1.5mm，加筋肋的高度和宽度分别为18.0mm和2.4mm，结构材料为铝2139，其弹性模量为72.50GPa，泊松比为0.3，密度为2.8e-6Kg/mm 3。
步骤202：如图5所示，设置四边简支位移边界条件，设置轴剪组合载荷边界条件，四边施加单位1剪力，上下边施加单位1轴力，左右边施加不均匀轴力，如式(1.4)和式(1.5)所示，其中P left为左边轴力，P right为右边轴力，l为曲线加筋板的高度，进一步完成曲线加筋结构的有限元线性屈曲分析；
Figure PCTCN2021077160-appb-000005
Figure PCTCN2021077160-appb-000006
步骤203：在设计域空间中采用拉丁超立方方法独立采样5次，生成5组250张含标签(质量、屈曲特征值)的曲线加筋结构图像，选一组图像集作为训练集N 1，其他四组作为测试集进行交叉检验N 2，选取均方根误差(％RMSE)作为模型的误差评估，如式(1.2)所示,其中n为样本数量，y i为结构响应值，
Figure PCTCN2021077160-appb-000007
为模型预测值；
Figure PCTCN2021077160-appb-000008
步骤204；由步骤108提取的解码网络模型E和两层全连接层搭建卷积神经网络模型；
步骤205：将含有标签的训练集喂入卷积神经网络模型训练，设置学习率0.005、数据输入批量100、训练步数1000，训练过程仅对最后两个全连接层中的参数进行训练调整；
步骤206：完成多组图像集的训练过程，并提取训练后的卷积神经网络模型；
步骤300：基于步骤206对曲线加筋结构质量、屈曲特征值预测的卷积网络模型，利用遗传优化算法，选取曲线加筋结构屈曲特征值不大于8.40作为约束条件，开展结构质量轻量化曲线加筋布局优化设计，包括以下子步骤：
步骤301：搭建遗传算法优化框架，优化迭代起始首先设置初始种群数150个，遗传代数15代，最大优化次数50次，生成初始种群曲线加筋图像集N g，其中每张图像大小为64*64；
步骤302：将初始种群生成的图像集N g输入由步骤206中提取的卷积神经网络；
步骤303：利用卷积神经网络模型开展基于遗传算法的曲线加筋结构布局优化设计，寻优获得新的样本点K；
步骤304：由获得的样本点K建立有限元模型，并进行结构线性屈曲分析检验，完成对样本点图像的标记；
步骤305：由获得的样本点K生成64*64大小的图像，并补充到训练图像集，进一步将扩充后的图像集N g+k输入步骤206中的卷积神经网络进行重新训练；
步骤306：将重新训练后的卷积神经网络模型
Figure PCTCN2021077160-appb-000009
代入步骤302中的卷积神经网络模型F继续进行遗传算法优化；
步骤307：判断遗传算法优化过程是否达到最大迭代次数收敛，如果收敛，输出最优设计变量及结构屈曲特征值，否则，返回执行步骤305。
针对薄壁曲线加筋结构布局设计问题，本发明设计了曲线加筋路径表征图像特征学习方法，充分地挖掘了结构图像中的结构信息，搭建的卷积神经网络模型对结构质量及屈曲特征值响应的预测均方根误差均在5％左右，极大程度地保证了曲线加筋结构布局设计问题中的模型精度。利用卷积神经网络模型开展基于遗传算法的曲线加筋结构质量轻量化设计，对比样本中的最轻质量0.133，基于卷积神经网络的质量轻量化最优结果为0.100，减重比例达到24.8％，另外开展可变筋条的质量轻量化最优结果为0.0954，减重比例28.3％。本发明是基于 结构图像特征提取的深度学习方法，相较于传统代理模型优化方法，显著地提高多变量复杂结构优化问题中的模型精度，并且获得了结构力学承载效率更高的曲线加筋布局设计。
最后应说明的是：以上各实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述各实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其对前述各实施例所记载的技术方案进行修改，或者对其中部分或者全部技术特征进行等同替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims (5)

  1. 一种基于图像特征学习的曲线加筋结构布局智能设计方法，其特征在于，包括以下步骤：
    步骤100：选取曲线加筋路径函数生成图像集，输入自编码网络进行无监督学习训练，完成曲筋图像结构特征的提取，包括以下子步骤：
    步骤101：选取路径函数B(t)，并确定加筋薄壁结构路径函数设计变量，如式(1.1)所示；
    B(t)＝(1-t) 2P s(x s,y s)+2t(1-t)P m(x m,y m)+t 2P e(x e,y e),t∈[0,1]  (1.1)
    其中，B(t)为路径函数，t为路径函数控制变量，P s(x s，y s)为路径起点坐标，P m(x m，y m)为路径内一点坐标，P e(x e，y e)为路径终点坐标；
    步骤102：根据结构不同边界类型组合确定路径函数类型，并对曲线加筋结构路径函数的设计域空间进行约束；
    步骤103：确定每张曲线加筋结构图像的大小m*n，生成无监督学习训练所用的训练图像集N 0；
    步骤104：搭建对曲线加筋结构布局图像的解码网络模型E和编码网络模型D；
    步骤105：将图像解码网络模型E和编码网络模型D进行组合，形成自编码网络模型；
    步骤106：将曲线加筋布局图像集N 0输入自编码网络模型；
    步骤107：完成自编码网络模型对曲线加筋图像集N 0的训练过程；
    步骤108：提取自编码网络模型训练后的解码网络模型E；
    步骤200：建立曲线加筋结构力学响应的分析模型，形成用于有监督学习训练的数据集，进一步输入由步骤108解码网络模型和全连接层搭建的卷积神经网络模型，完成曲线加筋结构力学响应的学习，包括以下子步骤：
    步骤201：根据曲线路径函数B(t)建立曲线加筋结构模型；
    步骤202：设置结构位移载荷边界条件，进行结构力学响应分析；
    步骤203：确定每张曲线加筋结构图像的大小m*n，根据图像对应结构力学 响应，生成用于有监督学习模型训练及检验的训练集N 1、检验集N 2；另外设置用于模型优劣的评价标准，如式(1.2)所示，选取均方根(％RMSE)作为模型的误差评估；
    Figure PCTCN2021077160-appb-100001
    其中，n为样本数量，y i为结构响应值，
    Figure PCTCN2021077160-appb-100002
    为模型预测值；
    步骤204；由步骤108的解码网络模型E和两个全连接层搭建卷积神经网络模型F；
    步骤205：将含力学响应标签的训练集N 1输入卷积神经网络模型F中进行训练；
    步骤206：根据检验集N 2对卷积神经网络模型F的准确性进行判定，完成卷积神经网络对曲线加筋结构力学响应的训练过程；
    步骤300：基于步骤206对曲线加筋结构力学响应预测的卷积神经网络模型F，利用演化类算法完成曲线加筋结构布局的优化设计，包括以下子步骤：
    步骤301：搭建演化类算法优化框架，优化迭代起始首先生成初始曲线加筋图像集N g；
    步骤302：将图像集N g输入由步骤206提取的卷积神经网络模型F中；
    步骤303：利用演化类算法在所建立的卷积神经网络模型F上寻优获得新样本点K；
    步骤304：由获得的样本点K建立曲线加筋结构模型，并通过力学响应分析进行标记；
    步骤305：将新样本点K补充到训练图像集N g形成图像集N g+k，进一步输入步骤206中的卷积神经网络模型F进行重新训练；
    步骤306：将重新训练的卷积神经网络模型
    Figure PCTCN2021077160-appb-100003
    代替步骤302中的卷积神经网络模型F，继续开展演化类算法的优化过程；
    步骤307：判断当前优化过程是否达到算法收敛条件，如果收敛，输出最优 设计变量，否则，返回执行步骤301，其中所述的收敛条件为达到优化算法的最大迭代次数。
  2. 根据权利要求1所述的一种基于图像特征学习的曲线加筋结构布局智能设计方法，其特征在于，所述的步骤101中，选取的路径函数需要约束函数曲率不能过大且函数中间路径不能超出平板设计区域，包括但不限于与样条函数。
  3. 根据权利要求1所述的一种基于图像特征学习的曲线加筋结构布局智能设计方法，其特征在于，所述的步骤202中，结构的力学响应包括静力、动力或结构屈曲响应特征，所用到的分析方法包括有限元分析法、边界元分析法、等几何分析法和无网格法。
  4. 根据权利要求1所述的一种基于图像特征学习的曲线加筋结构布局智能设计方法，其特征在于，所述的步骤301中的演化类算法包括遗传算法、模拟退火算法、人工神经网络算法、粒子群算法、蚁群算法。
  5. 根据权利要求1所述的一种基于图像特征学习的曲线加筋结构布局智能设计方法，其特征在于，所述的步骤301至步骤307过程，需要对固定数目筋条和可变数目筋条分别进行优化，在可变数目筋条结构开展优化布局设计过程中，由于步骤100和步骤200形成的卷积神经网络已对曲筋图像完成结构特征力学响应的学习过程，无需额外生成可变筋条的训练集，只需基于可变筋条的程序代码重新开展步骤301至步骤307的优化过程，即可实现筋条数目动态可变的曲线加筋布局优化设计。
PCT/CN2021/077160 2020-07-08 2021-02-22 一种基于图像特征学习的曲线加筋结构布局智能设计方法 Ceased WO2022007409A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/424,785 US20220138582A1 (en) 2020-07-08 2021-02-22 Intelligent layout design method of curvilinearly stiffened structures based on image feature learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010649313.5A CN111859790B (zh) 2020-07-08 2020-07-08 一种基于图像特征学习的曲线加筋结构布局智能设计方法
CN202010649313.5 2020-07-08

Publications (1)

Publication Number Publication Date
WO2022007409A1 true WO2022007409A1 (zh) 2022-01-13

Family

ID=73151934

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/077160 Ceased WO2022007409A1 (zh) 2020-07-08 2021-02-22 一种基于图像特征学习的曲线加筋结构布局智能设计方法

Country Status (3)

Country Link
US (1) US20220138582A1 (zh)
CN (1) CN111859790B (zh)
WO (1) WO2022007409A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115099131A (zh) * 2022-06-10 2022-09-23 桂林电子科技大学 基于深度学习的极化转换超表面智能化逆向设计方法及系统
CN115320879A (zh) * 2022-10-14 2022-11-11 中国空气动力研究与发展中心低速空气动力研究所 一种环量控制翼型后缘柯恩达型面设计方法
CN115481488A (zh) * 2022-09-21 2022-12-16 中南大学 基于机器学习的方锥式吸能结构耐撞性能多目标优化方法
CN119808311A (zh) * 2024-12-23 2025-04-11 西安交通大学 一种复杂超结构的多类别强非线性曲线预测方法及系统
CN119989780A (zh) * 2025-01-08 2025-05-13 杭州电子科技大学 一种编织复合材料变曲率加筋薄壁结构一体化设计方法

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111859790B (zh) * 2020-07-08 2022-09-16 大连理工大学 一种基于图像特征学习的曲线加筋结构布局智能设计方法
CN112614109B (zh) * 2020-12-24 2024-06-07 四川云从天府人工智能科技有限公司 图像质量评估方法、装置以及计算机可读存储介质
CN113919071B (zh) * 2021-10-13 2022-07-05 武汉大学 一种平板加筋的布局及形状优化方法和装置
CN115830411B (zh) * 2022-11-18 2023-09-01 智慧眼科技股份有限公司 生物特征模型训练方法、生物特征提取方法及相关设备
CN116402995B (zh) * 2023-02-24 2026-01-02 北京建筑大学 基于轻量级神经网络的古建筑点云语义分割方法及系统
CN116822038B (zh) * 2023-03-09 2024-02-06 大连理工大学 基于数据驱动的异型封闭加筋拓扑优化方法
CN117610180B (zh) * 2023-11-16 2024-05-14 苏州科技大学 一种板壳加强筋生成式设计方法
CN117807823B (zh) * 2023-12-12 2024-10-01 大连理工大学 面向数字孪生建模的复杂曲面传感器布局方法
CN119729461B (zh) * 2024-12-10 2025-11-11 西北大学 可编程超表面的无线传感加密控制方法及无线传感系统
CN119919756B (zh) * 2025-01-17 2025-10-21 湖南大学 一种基于演化算法的文生图大模型提示词优化方法和系统
CN120688163B (zh) * 2025-08-25 2025-10-28 吉林大学 防撞梁的智能生成式设计方法
CN121030935B (zh) * 2025-10-27 2026-02-17 南京航空航天大学 刚度分布引导的航空薄壁构件定位布局智能设计方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102930137A (zh) * 2012-09-29 2013-02-13 华侨大学 一种加筋板结构的优化设计方法
CN104809469A (zh) * 2015-04-21 2015-07-29 重庆大学 一种面向服务机器人的室内场景图像分类方法
WO2017159638A1 (ja) * 2016-03-14 2017-09-21 オムロン株式会社 能力付与データ生成装置
CN108469375A (zh) * 2018-03-15 2018-08-31 中国航空工业集团公司沈阳飞机设计研究所 一种加筋板屈曲载荷判定方法及试验系统
CN111859790A (zh) * 2020-07-08 2020-10-30 大连理工大学 一种基于图像特征学习的曲线加筋结构布局智能设计方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107122809B (zh) * 2017-04-24 2020-04-28 北京工业大学 基于图像自编码的神经网络特征学习方法
CN109543745B (zh) * 2018-11-20 2021-08-24 江南大学 基于条件对抗自编码网络的特征学习方法及图像识别方法
CN109816661B (zh) * 2019-03-22 2022-07-01 电子科技大学 一种基于深度学习的牙齿ct图像分割方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102930137A (zh) * 2012-09-29 2013-02-13 华侨大学 一种加筋板结构的优化设计方法
CN104809469A (zh) * 2015-04-21 2015-07-29 重庆大学 一种面向服务机器人的室内场景图像分类方法
WO2017159638A1 (ja) * 2016-03-14 2017-09-21 オムロン株式会社 能力付与データ生成装置
CN108469375A (zh) * 2018-03-15 2018-08-31 中国航空工业集团公司沈阳飞机设计研究所 一种加筋板屈曲载荷判定方法及试验系统
CN111859790A (zh) * 2020-07-08 2020-10-30 大连理工大学 一种基于图像特征学习的曲线加筋结构布局智能设计方法

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115099131A (zh) * 2022-06-10 2022-09-23 桂林电子科技大学 基于深度学习的极化转换超表面智能化逆向设计方法及系统
CN115481488A (zh) * 2022-09-21 2022-12-16 中南大学 基于机器学习的方锥式吸能结构耐撞性能多目标优化方法
CN115481488B (zh) * 2022-09-21 2024-03-01 中南大学 基于机器学习的方锥式吸能结构耐撞性能多目标优化方法
CN115320879A (zh) * 2022-10-14 2022-11-11 中国空气动力研究与发展中心低速空气动力研究所 一种环量控制翼型后缘柯恩达型面设计方法
CN115320879B (zh) * 2022-10-14 2022-12-09 中国空气动力研究与发展中心低速空气动力研究所 一种环量控制翼型后缘柯恩达型面设计方法
CN119808311A (zh) * 2024-12-23 2025-04-11 西安交通大学 一种复杂超结构的多类别强非线性曲线预测方法及系统
CN119989780A (zh) * 2025-01-08 2025-05-13 杭州电子科技大学 一种编织复合材料变曲率加筋薄壁结构一体化设计方法

Also Published As

Publication number Publication date
CN111859790A (zh) 2020-10-30
CN111859790B (zh) 2022-09-16
US20220138582A1 (en) 2022-05-05

Similar Documents

Publication Publication Date Title
CN111859790B (zh) 一种基于图像特征学习的曲线加筋结构布局智能设计方法
Wu et al. Airfoil shape optimization using genetic algorithm coupled deep neural networks
CN102789539B (zh) 一种飞行器升力面结构优化设计方法
CN107341279B (zh) 一种针对高耗时约束的飞行器快速近似优化方法
CN117151205B (zh) 一种基于多先验策略的强化学习智能决策方法
CN112231839B (zh) 针对高耗时约束的飞行器追峰采样智能探索方法
CN109934330A (zh) 基于多样化种群的果蝇优化算法来构建预测模型的方法
KR102649283B1 (ko) 임베디드 플랫폼을 위한 심층신경망 최적화 장치 및 방법
CN114707652A (zh) 基于改进woa算法的rbf神经网络参数优化方法
CN114358197A (zh) 分类模型的训练方法及装置、电子设备、存储介质
Favilli et al. Geometric deep learning for statics-aware grid shells
Liu et al. A deep reinforcement learning optimization framework for supercritical airfoil aerodynamic shape design
CN119494058A (zh) 一种基于人工智能的数据标签分类方法及系统
CN114781207B (zh) 基于不确定性和半监督学习的热源布局温度场预测方法
CN114819091A (zh) 基于自适应任务权重的多任务网络模型训练方法及系统
Lin Optimizing Kernel Extreme Learning Machine based on a Enhanced Adaptive Whale Optimization Algorithm for classification task
Xu et al. Towards efficient filter pruning via adaptive automatic structure search
Zhan et al. A one-time training machine learning method for general structural topology optimization
CN114943837A (zh) 一种基于改进U-net的盐丘识别方法
CN114239330A (zh) 基于深度学习的大跨度网壳结构形态创建方法
CN119829840A (zh) 一种在游戏内容搜索联想领域多任务建模的方法
JP2022002087A (ja) 多目標最適化を行うための装置、方法及び記憶媒体
KR20250077737A (ko) 전이 최적화된 객체 탐지를 위한 신경망 구조 탐색 방법
CN111539306B (zh) 基于激活表达可替换性的遥感图像建筑物识别方法
Koratikere et al. Multi-Point Airfoil Shape Optimization Using Neural Network-Based Sequential Sampling

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21837719

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21837719

Country of ref document: EP

Kind code of ref document: A1