Net shell structure node parameter automatic optimization method based on neural network model

Abstract

The present application relates to the technical field of space structure design and parameter optimization, and particularly relates to a net shell structure node parameter automatic optimization method based on a neural network model, comprising: collecting "input-output" data pairs of a single-layer cylindrical net shell constructed by an aluminum alloy plate type node, and input parameters comprising design parameters and initial node parameters of the single-layer cylindrical net shell. The present application first obtains node optimization data under different parameters by a genetic algorithm combined with finite element software and programming tools, and supplements calculation examples to construct a training data set covering the range of common engineering design parameters, then trains a three-layer fully connected neural network model containing two layers of hidden layers, and cooperates with an early stop strategy and a learning rate decay strategy to ensure the model accuracy. When applied, the model can automatically output the optimized node parameters after inputting the design parameters of the target single-layer cylindrical net shell, without the repeated iteration and manual intervention of traditional methods, greatly shortening the optimization time consumption, and efficiently meeting the batch optimization demand under multiple working conditions.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of spatial structure design and parameter optimization, and particularly relates to a net shell structure node parameter automatic optimization method based on a neural network model. BACKGROUND

[0002] As a kind of efficient space bearing structure, the net shell structure is widely used in large-span buildings such as stadiums, transportation hubs and exhibition centers due to good mechanical properties and flexible architectural forms. Among them, the aluminum alloy single-layer cylindrical net shell becomes an important choice for modern green buildings due to the advantages of light weight, high strength, corrosion resistance and low life cycle maintenance cost. The core stress unit of this kind of net shell is the node and the rod, and especially the plate-type node as a commonly used semi-rigid node form, the node parameters (including node vertical coordinates, node bending stiffness, etc.) directly affect the overall stress state of the net shell, the node coordinates determine the surface form of the net shell, and then affect the load transfer path; the node stiffness determines the semi-rigid characteristics of the net shell, and if the parameters are not reasonably set, the ultimate bearing capacity of the net shell is easily reduced or local stress concentration is caused. In actual design, the node parameters need to be optimized so that the net shell can maintain the macro-aesthetic form while realizing the thin film stress dominance (improving the rod axial force and reducing the bending moment), thereby enhancing the overall stability performance.

[0003] At present, the optimization of the node parameters of the aluminum alloy single-layer cylindrical net shell is mostly realized by the interactive mode of the genetic algorithm combined with the finite element software (such as ANSYS) and the programming tool (such as MATLAB). Specifically, the node parameter adjustment scheme is generated by the genetic algorithm iteration, and then the finite element model of the net shell is established by using ANSYS to calculate the bearing capacity of the net shell, and MATLAB and ANSYS are repeatedly interacted to select the optimal scheme. However, this method has significant technical limitations: in order to ensure the optimization accuracy, a large population size and iteration generation (such as a population size of 50 and an iteration of 50 generations) are usually required, and the time consumption of a single optimization calculation is extremely long - even if the time is reduced by avoiding repeated gene calculation, it still takes about 74 hours to complete the optimization of a typical model; if the number of net shell nodes increases and the number of parameter variables (such as the rise-span ratio, the length-span ratio and the load ratio combination) increases, the time consumption will increase sharply, which makes it difficult to meet the demand of rapid optimization in engineering design. More importantly, the whole process needs manual intervention in parameter setting and software interaction, and cannot realize the automatic output of the node parameters, and when facing the batch optimization demand under multiple working conditions (such as different load distribution and support conditions), the problem of low efficiency is more prominent, which restricts the engineering application efficiency of the aluminum alloy net shell structure. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a net shell structure node parameter automatic optimization method based on a neural network model, which solves the problems of long optimization time, the need for manual intervention and the inability to automatically output parameters when compared with the prior art, and low efficiency of batch optimization under multiple working conditions.

[0005] To achieve the above object, the present application is implemented by the following technical solutions: a net shell structure node parameter automatic optimization method based on a neural network model, comprising:

[0006] S1, collecting an "input-output" data pair of a single-layer cylindrical net shell constructed by an aluminum alloy plate node, the input parameters including design parameters and initial node parameters of the single-layer cylindrical net shell, and the output parameters including optimized node vertical coordinate adjustment values and node bending stiffness adjustment values;

[0007] S2, performing normalization processing on the input parameters to eliminate dimensional differences and form a training data set of a neural network model;

[0008] S3, constructing a three-layer fully connected neural network model, the model including an input layer, a hidden layer and an output layer, and the hidden layer using a ReLU activation function to enhance the nonlinear fitting ability of the model;

[0009] S4, training the neural network model using the training data set, the training including dividing the training data set into a training set, a validation set and a test set, and using an early stopping strategy to avoid model overfitting;

[0010] S5, verifying the neural network model, the verification including comparing the optimized node parameters output by the model with traditional optimization results, and if the deviation of the ultimate bearing capacity of the single-layer cylindrical net shell corresponding to the optimized node parameters is less than or equal to 5% and the maximum displacement deviation is less than or equal to 8%, the model verification is passed;

[0011] S6, applying the neural network model for node parameter optimization, the application including inputting the design parameters of the target single-layer cylindrical net shell, and the neural network model automatically outputting optimized node vertical coordinate adjustment values and node bending stiffness adjustment coefficients;

[0012] S7, based on a cubic spline interpolation method, correcting the cross-sectional curve of the single-layer cylindrical net shell according to the node vertical coordinate adjustment values, so that the corrected curve is smooth and continuous;

[0013] S8, importing the optimized node parameters into finite element software for bearing capacity checking, and if the bearing capacity checking does not meet the engineering requirements, the parameters are input again for fine tuning and optimization.

[0014] Further, the input parameters include a rise-to-span ratio, a length-to-span ratio, a half-span load ratio, a support condition, an initial node x coordinate, an initial node y coordinate, an initial node bending stiffness, and a net shell span of the single-layer cylindrical latticed shell, wherein the rise-to-span ratio is set in a range of 1 / 6, 1 / 5, 1 / 4, and 1 / 3, the length-to-span ratio is set in a range of 1.4, 1.8, 2.2, and 2.6, the half-span load ratio is set in a range of 0, 1 / 4, 1 / 2, and 1, and the support condition includes two longitudinal edge supports and four edge supports.

[0015] Further, the input-output data pairs are derived from the optimized node data under different parameters obtained by genetic algorithm combined with the finite element software ANSYS and the programming tool MATLAB interactive optimization, and 100 groups of different parameter combinations of net shell optimization examples are supplemented to ensure that the training data set covers the design parameter range commonly used in engineering; the normalization processing includes converting the node coordinate parameters and the node bending stiffness parameters into relative values to eliminate the influence of dimensional differences on the training results of the neural network model.

[0016] Further, the number of nodes in the input layer is 8, and the 8 input parameters correspond to the rise-to-span ratio, the length-to-span ratio, the half-span load ratio, the support condition code value, the initial node x coordinate, the initial node y coordinate, the initial node bending stiffness, and the net shell span of the net shell in sequence.

[0017] Further, in the three-layer fully connected neural network model, the hidden layer is set to two layers, wherein the number of nodes in the first layer of the hidden layer is 64, and the number of nodes in the second layer of the hidden layer is 32.

[0018] Further, the number of nodes in the output layer is 2, corresponding to the adjusted value of the optimized node y coordinate and the adjusted coefficient of the node bending stiffness, the neural network model uses a mean square error function as a loss function to measure the deviation between the optimized parameters and the actual parameters of the model output, and uses an Adam optimizer, and the learning rate of the Adam optimizer is initially set to 0.001.

[0019] Further, the training process divides the complete data set into a training set, a validation set, and a test set in a ratio of 7:2:1, and the number of training rounds of the neural network model is set to 500 rounds.

[0020] Further, the condition of the early stopping strategy is to stop training when the loss of the validation set does not decrease for 20 consecutive rounds, so as to avoid overfitting of the neural network model; the Adam optimizer adopts a learning rate decay strategy, specifically, the learning rate is decayed to 0.9 of the original learning rate every 100 iterations, for improving the stability of the model training.

[0021] Further, the verification includes selecting a typical example in the test set, the input parameters of the typical example including the parameter combination of the 1 / 6 of the rise-to-span ratio, the 2.6 of the length-to-span ratio, the 0 of the load ratio, the four-edge supported net shell model, introducing the input parameters into the trained neural network model, outputting the optimized node parameters, and comparing with the optimization results of the traditional genetic algorithm; if the net shell ultimate bearing capacity deviation or the maximum displacement deviation exceeds the preset range, the hidden layer node number or the learning rate is adjusted, and the model is retrained until the accuracy requirement is met.

[0022] Further, when the neural network model is applied to optimize the node parameters, the design parameters of the target single-layer cylindrical net shell are input, including the rise-to-span ratio, the length-to-span ratio, the load ratio, the support condition, the initial node parameter and the net shell span information, and the neural network model automatically outputs the optimized node y coordinate adjustment value and the node bending stiffness adjustment coefficient.

[0023] Compared with the prior art, the beneficial effects of the present application are:

[0024] The present application first obtains the node optimization data under different parameters by genetic algorithm combined with finite element software and programming tools, supplements the calculation examples to construct the training data set covering the common design parameter range of engineering, trains the three-layer fully connected neural network model containing two hidden layers, cooperates with the early stop strategy and the learning rate decay strategy to ensure the model accuracy. When applied, the design parameters of the target single-layer cylindrical net shell are input, and the model can automatically output the optimized node parameters without repeated iteration and manual intervention of the traditional method, greatly shortening the optimization time and efficiently meeting the batch optimization demand under multiple working conditions. At the same time, based on the cubic spline interpolation method, the net shell cross section curve is corrected to be smooth and continuous, and then the optimized parameters are introduced into the finite element software for checking, so as to ensure that the net shell ultimate bearing capacity and the maximum displacement deviation meet the engineering requirements, avoid the unreasonable parameters leading to the structural performance problems, and effectively improve the overall stability performance and engineering design efficiency of the aluminum alloy single-layer cylindrical net shell. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The overall method flowchart of the present application is shown in the figure;

[0026] Figure 2 The neural network model architecture diagram of the present application is shown in the figure;

[0027] Figure 3 The training strategy flowchart of the present application is shown in the figure;

[0028] Figure 4 The node curve optimization schematic diagram of the present application is shown in the figure;

[0029] Figure 5 The engineering application flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0031] Please refer to Figures 1-5 The present application provides a method for automatically optimizing node parameters of a lattice shell structure based on a neural network model, comprising:

[0032] S1, collecting an "input-output" data pair of a single-layer cylindrical lattice shell constructed by an aluminum alloy plate node, the input parameters including design parameters and initial node parameters of the single-layer cylindrical lattice shell, and the output parameters including adjusted values of node vertical coordinates and node bending stiffness after optimization;

[0033] S2, performing normalization processing on the input parameters to eliminate dimensional differences and form a training data set of a neural network model;

[0034] S3, constructing a three-layer fully connected neural network model, the model including an input layer, a hidden layer and an output layer, and the hidden layer using a ReLU activation function to enhance the nonlinear fitting ability of the model;

[0035] S4, training the neural network model using the training data set, the training including dividing the training data set into a training set, a validation set and a test set, and using an early stopping strategy to avoid model overfitting;

[0036] S5, verifying the neural network model, the verification including comparing the optimized node parameters output by the model with traditional optimization results, and if the deviation of the ultimate bearing capacity of the single-layer cylindrical lattice shell corresponding to the optimized node parameters is less than or equal to 5% and the maximum displacement deviation is less than or equal to 8%, the model verification is passed;

[0037] S6, applying the neural network model to optimize node parameters, the application including inputting design parameters of a target single-layer cylindrical lattice shell, and the neural network model automatically outputting adjusted values of optimized node vertical coordinates and node bending stiffness adjustment coefficients;

[0038] S7, based on a cubic spline interpolation method, correcting a cross-sectional curve of the single-layer cylindrical lattice shell according to the adjusted values of the node vertical coordinates, so that the corrected curve is smooth and continuous;

[0039] S8, importing the optimized node parameters into a finite element software to perform bearing capacity checking, and if the bearing capacity checking does not meet engineering requirements, the parameters are input again for fine tuning and then optimized again.

[0040] Specifically, for the traditional optimization method relies on genetic algorithm and finite element software MATLAB interaction, single optimization time-consuming and need human intervention, while there is a smooth curve after optimization shell, material model simplification resulting in low accuracy problems, the embodiment is carried out as follows. First, the "input-output" data collection, to the aluminum alloy plate type node single-layer cylindrical shell as the object, according to the optimization objective function and constraint conditions to determine the collection logic, the objective function is:

[0041]

[0042] (Wherein is the vertical coordinate adjustment value of the independent control point, is the span of the shell, is the scaling factor and is recommended to be ≤1 / 100 to avoid macro shape change), through the genetic algorithm to write optimization program, call finite element software ANSYS to establish the model, combined with MATLAB to realize the automatic interaction iteration of the three, get a group of node optimization data, supplemented by 100 groups of examples covering different vector span ratio, long span ratio, load ratio and support conditions, to ensure that the data covers the range of common design parameters in engineering.

[0043] Then the input parameters are normalized, the node coordinate parameters are converted into relative values with respect to the span of the shell, and the bending stiffness parameters of the nodes are converted into relative values with respect to the initial bending stiffness of the nodes, so as to eliminate the dimensional differences and form the training data set. Then a three-layer fully connected neural network model is constructed, the input layer receives 8 input parameters, the hidden layer is set to two layers (64 nodes in the first layer and 32 nodes in the second layer), and the ReLU activation function is used to enhance the nonlinear fitting ability, and the output layer corresponds to the output node vertical coordinate adjustment value and node bending stiffness adjustment coefficient. In the training stage, the data set is divided into training set, validation set and test set according to the ratio of 7:2:1, the early stopping strategy (validation set loss continuous 20 rounds without decline then stop) is adopted to avoid overfitting, and the Adam optimizer (initial learning rate 0.001, decay to 0.9 of the original every 100 rounds) is selected to improve the training stability.

[0044] In the model verification, a typical example with vector span ratio 1 / 6, long span ratio 2.6, load ratio 0 and four-side support is selected from the test set, and the optimization parameters output by the model are compared with the results of the traditional genetic algorithm. In addition to checking the limit bearing capacity deviation ≤5% and the maximum displacement deviation ≤8%, the formula of bar internal force ratio is also used:

[0045]

[0046] is the axial force ratio, is the bending moment ratio, , is the actual axial force and bending moment, ​for non-proportional extension strength, for cross-sectional area, for moment of inertia, for cross-sectional height) to determine the rationality of force, if the axial force ratio is concentrated and the bending moment ratio is reduced, further evidence the effectiveness of the model.

[0047] When applying the model, input the target net shell design parameters, and the model automatically outputs the optimized node vertical coordinate adjustment value and node bending stiffness adjustment coefficient; based on the cubic spline interpolation method to correct the cross-sectional curve, using the formula:

[0048] for curve coefficient) to calculate the segmented curve, so that the corrected curve is smooth and continuous. Finally, the optimization parameters are imported into the finite element software for checking, and the Ramberg-Osgood material constitutive equation for strain, for stress, for elastic modulus, for strain hardening parameter) to set the material properties to ensure the accuracy of the checking; if it does not meet the engineering requirements, fine-tune the input parameters and optimize again.

[0049] Through this embodiment, the neural network model can automatically output optimization parameters without repeated iteration and manual intervention, greatly reducing the time consumption, and at the same time, with the help of cubic spline interpolation and accurate material model, the net shell curve is beautiful and the checking is reliable.

[0050] Further, the input parameters include the rise-to-span ratio, the length-to-span ratio, the half-span load ratio, the support condition, the initial node x coordinate, the initial node y coordinate, the initial node bending stiffness, and the net shell span of the single-layer cylindrical net shell, wherein the set range of the rise-to-span ratio includes 1 / 6, 1 / 5, 1 / 4 and 1 / 3, the set range of the length-to-span ratio includes 1.4, 1.8, 2.2 and 2.6, the set range of the half-span load ratio includes 0, 1 / 4, 1 / 2 and 1, and the support condition includes two longitudinal edge supports and four edge supports.

[0051] ​​​​​Specifically, the selection of input parameters needs to strictly follow the preset range, and be adapted to the input layer nodes of the neural network model and the engineering actual demand. The rise-to-span ratio is selected from 1 / 6, 1 / 5, 1 / 4 and 1 / 3, for example, a large-span exhibition center net shell can select 1 / 6 to balance space and stability; the length-to-span ratio is selected from 1.4, 1.8, 2.2 and 2.6, and a long and narrow type traffic hub net shell can select 2.2 or 2.6 to adapt to the layout; the half-span load ratio is determined according to the actual load distribution, and is 0 when only full-span constant load is borne, and is 1 / 4, 1 / 2 or 1 when half-span live load and full-span constant load are combined; the support condition is divided into two longitudinal side support (both sides have reliable constraint scene) and four side support (all around have constraint scene); the initial node x coordinate and the initial node y coordinate are determined according to the net shell grid division scheme, which needs to be matched with the span and the rise, and is converted into a relative value relative to the span in subsequent normalization; the initial node bending stiffness is calculated according to the node plate thickness, the number of bolts and other construction parameters, and the net shell span is determined according to the actual covered space size.

[0052] The determination of the value range of each parameter not only ensures that the input parameters meet the common engineering scenarios, but also provides regular input for subsequent data normalization and model training, avoids model fitting deviation caused by parameter confusion, and reduces the operation difficulty for engineers to select parameters directly, thereby indirectly improving the optimization efficiency.

[0053] Further, the "input-output" data pair is derived from the node optimization data under different parameters obtained by genetic algorithm combined with the interaction optimization of finite element software ANSYS and programming tool MATLAB, and 100 groups of net shell optimization examples with different parameter combinations are supplemented to ensure that the training data set covers the design parameter range commonly used in engineering; the normalization process includes converting the node coordinate parameters and node bending stiffness parameters into relative values to eliminate the influence of dimension difference on the training results of the neural network model.

[0054] Specifically, the acquisition of the "input-output" data pair is based on the traditional optimization method, combined with the optimization objective function and the constraint condition. By writing a program through genetic algorithm, the population initialization, selection, crossover and mutation logic are constructed in MATLAB, the net shell model is established by calling the ANSYS parameterized language, and the objective function Iterative selection of optimal node parameters, automatic interaction of MATLAB and ANSYS, and multiple sets of optimization data under different parameters are obtained; at the same time, 100 examples are supplemented to cover different rise-to-span ratios, length-to-span ratios, load ratios and support conditions, to ensure that the data cover common engineering scenarios.

[0055] In the normalization process, the node coordinate parameters are converted into relative values with respect to the shell span, and the node bending stiffness parameters are converted into relative values with respect to the initial node bending stiffness, so as to eliminate the dimensional difference. For example, when the node x coordinate is 5 m and the span is 30 m, the normalized value is 5 / 30≈0.167; when the initial bending stiffness of the node is 4000 kN·m / rad and the bending stiffness of an example is 4400 kN·m / rad, the normalized value is 4400 / 4000=1.1. This avoids the influence of different parameter magnitudes on the model training accuracy, and provides a reliable data basis for the mapping relationship between the subsequent model learning input and output.

[0056] Further, the number of input layer nodes is 8, and the 8 input parameters correspond to the aspect ratio, the length ratio, the half-span load ratio, the support condition coding value, the initial node x coordinate, the initial node y coordinate, the initial node bending stiffness and the shell span in turn.

[0057] Specifically, the 8 nodes of the input layer correspond to the key parameters of the shell in a fixed order, ensuring accurate analysis of the input information by the model. The first node corresponds to the aspect ratio (e.g. 1 / 5), the second node corresponds to the length ratio (e.g. 2.0), the third node corresponds to the half-span load ratio (e.g. 1 / 4), the fourth node corresponds to the support condition coding value (two longitudinal edge support coding is 1, four edge support coding is 2), the fifth node corresponds to the initial node x coordinate (normalized relative value), the sixth node corresponds to the initial node y coordinate (normalized relative value), the seventh node corresponds to the initial node bending stiffness (normalized relative value), and the eighth node corresponds to the shell span (normalized relative value).

[0058] This order is consistent with the parameter arrangement logic in the data collection and normalization stage. When used by engineering personnel, the order can be directly input without adjustment, avoiding model output deviation caused by incorrect order. At the same time, the clear node-parameter correspondence relationship enables the model to accurately extract the characteristics of each parameter, providing a correct basis for subsequent hidden layer feature processing and output layer parameter calculation, and ensuring the reliability of the optimization result.

[0059] Further, in the three-layer fully connected neural network model, the hidden layer is set to 2 layers, wherein the number of nodes of the first layer of hidden layer is 64, and the number of nodes of the second layer of hidden layer is 32.

[0060] Specifically, the hidden layer of the three-layer fully connected neural network model is set to two layers, which is determined based on the complexity of the parameter optimization problem of the reticulated shell node. The first hidden layer node number is set to 64, which can fully extract the complex correlation characteristics between the aspect ratio, the length ratio, the support condition and the node parameters in the input parameters, and avoid insufficient feature extraction due to too few nodes. The second hidden layer node number is set to 32, which can further filter and integrate the features extracted by the first layer, eliminate redundant information, and retain the key features to the optimization result, avoiding the increase of model complexity and the increase of training time due to too many nodes.

[0061] The two hidden layers cooperate with the ReLU activation function, which can enhance the fitting ability of the model to the nonlinear optimization problem, control the overall complexity of the model, make the training process quickly converge, quickly process the input parameters and output the results in the application stage, adapt to the optimization efficiency requirements of engineering, and avoid time-consuming or precision problems caused by unreasonable model structure.

[0062] Further, the output layer node number is 2, corresponding to the adjusted node y coordinate value and the node bending stiffness adjustment coefficient after optimization, the neural network model uses the mean square error function as the loss function to measure the deviation of the optimization parameters output by the model from the actual parameters, and uses the Adam optimizer, and the learning rate of the Adam optimizer is initially set to 0.001.

[0063] Specifically, the output layer node number is set to 2, corresponding to the adjusted node y coordinate value and the node bending stiffness adjustment coefficient after optimization, the adjusted node y coordinate value is directly used to correct the vertical position of the node, and the node bending stiffness adjustment coefficient needs to be multiplied by the initial node bending stiffness to obtain the final optimized stiffness, without additional conversion, improving the convenience of engineering application.

[0064] The loss function selects the mean square error function to quantify the deviation of the model output from the actual optimal parameters, providing a clear direction for parameter adjustment and ensuring that the model converges in the direction of error reduction; the optimizer selects the Adam optimizer, and the initial learning rate is set to 0.001, which can ensure the convergence speed in the initial training stage, avoid slow training caused by too small learning rate, and avoid model oscillation caused by too large learning rate, cooperate with the subsequent learning rate decay strategy, make the model adjust the parameters in the later stage, further improve the precision, and ensure that the output parameters can accurately reflect the optimal state of the reticulated shell.

[0065] Further, the training process divides the complete data set into a training set, a validation set and a test set in the ratio of 7:2:1, and the training rounds of the neural network model are set to 500 rounds.

[0066] Specifically, the training data set is divided into training set, validation set and test set in the ratio of 7:2:1. The 70% training set is used to learn the mapping relationship between input and output to ensure that the optimization rule is fully mastered. The 20% validation set is used to monitor the training state in real time to avoid model overfitting. The 10% test set is used for final inspection of generalization ability to ensure that the model is reliable in new scenarios.

[0067] The training round is set to 500 rounds, which is determined based on pre-training experiments. This can ensure that the model converges fully to achieve stable accuracy, avoid excessive training time caused by too many rounds, balance training effectiveness and efficiency, avoid the time-consuming problem of too many iterations in traditional optimization methods, and ensure that the model can be quickly put into engineering application.

[0068] Further, the condition of the early stopping strategy is that the validation set loss does not decrease for 20 consecutive rounds, and the training is stopped to avoid overfitting of the neural network model. The Adam optimizer adopts a learning rate decay strategy, specifically decaying the learning rate to 0.9 of the original every 100 iterations, to improve the stability of the model training.

[0069] Specifically, the early stopping strategy is set to stop training when the validation set loss does not decrease for 20 consecutive rounds. When the validation set loss does not decrease continuously, it means that the model has mastered the optimization rule, and continuing training will lead to overfitting. At this time, stopping can ensure accuracy and avoid unnecessary time consumption.

[0070] The Adam optimizer adopts a strategy of decaying the learning rate to 0.9 of the original every 100 iterations. A higher learning rate (0.001) is used in the early stage of training to accelerate convergence, and the learning rate is gradually reduced in the later stage to make parameter adjustment more precise, avoid shocks, improve training stability, ensure stable convergence of the model to an optimal state, and reduce accuracy fluctuations.

[0071] Further, the verification includes selecting a typical example in the test set, the input parameters of the typical example including the parameter combination of the 1 / 6 aspect ratio, the 2.6 aspect ratio, the 0 load ratio, and the four-edge supported lattice shell model, importing the input parameters into the trained neural network model, outputting the optimized node parameters, and comparing with the optimization results of the traditional genetic algorithm. If the lattice ultimate bearing capacity deviation or the maximum displacement deviation exceeds the preset range, adjust the number of hidden layer nodes or the learning rate, and retrain the model until the accuracy requirement is met.

[0072] Specifically, the model verification selects a typical example with a rise-to-span ratio of 1 / 6, a length-to-span ratio of 2.6, a load ratio of 0, and four-edge support in the test set. This example is an engineering foundation scenario. If the model meets the accuracy in this scenario, the reliability of other scenarios can also be guaranteed. After normalizing the example input parameters, the optimized node parameters are imported into the model. Meanwhile, the traditional genetic algorithm is combined with ANSYS and MATLAB to obtain the traditional optimization results according to the objective function The limit bearing capacity and maximum displacement deviation of the two groups of results are compared.

[0073] If the deviation exceeds the preset range (such as a limit bearing capacity deviation of 7%), the number of hidden layer nodes (such as changing from 64 to 70 in the first layer and from 32 to 35 in the second layer) or the learning rate (such as changing the initial value to 0.0008) is adjusted, and the model is retrained. During the verification process, the internal force ratio formula can also be used for auxiliary judgment. If the axial force ratio improves and the bending moment ratio decreases after optimization, it further proves the effectiveness of the model and ensures the overall reliability of the verification.

[0074] Further, when the neural network model is applied for node parameter optimization, the design parameters of the target single-layer cylindrical latticed shell are input, including the rise-to-span ratio, the length-to-span ratio, the load ratio, the support condition, the initial node parameter, and the latticed shell span information. The neural network model automatically outputs the optimized node y-coordinate adjustment value and the node bending stiffness adjustment coefficient.

[0075] Specifically, when the neural network model is applied, the input design parameters of the target single-layer cylindrical latticed shell need to be complete and meet the requirements, including the rise-to-span ratio (such as 1 / 4), the length-to-span ratio (such as 1.8), the load ratio (such as 0), the support condition (such as two longitudinal edge supports), the initial node x-coordinate (normalized relative value), the initial node y-coordinate (normalized relative value), the initial node bending stiffness (normalized relative value), and the latticed shell span (normalized relative value).

[0076] After the parameters are arranged in the input layer order and normalized, they are input into the model. The model automatically outputs the optimized node y-coordinate adjustment value and the node bending stiffness adjustment coefficient. After the output, the adjustment value can be preliminarily checked for reasonableness according to the constraint conditions to avoid abnormal shapes caused by excessively large adjustment values. This process does not require human intervention in software interaction and can quickly complete single-latticed shell optimization. When facing batch optimization under multiple working conditions, only the parameters of each latticed shell need to be input in sequence, which greatly improves the efficiency of engineering design.

[0077] To sum up, the application first obtains node optimization data under different parameters by genetic algorithm combined with finite element software and programming tools, supplements algorithm examples to construct a training data set covering the range of common design parameters in engineering, and then trains a three-layer fully connected neural network model with two hidden layers, cooperates with early stopping strategy and learning rate decay strategy to ensure the accuracy of the model. When applied, the design parameters of the target single-layer cylindrical shell are input, and the model can automatically output the optimized node parameters without the need for repeated iteration and manual intervention of traditional methods, greatly reducing the optimization time and efficiently meeting the demand for batch optimization under multiple working conditions. At the same time, based on the cubic spline interpolation method, the cross section curve of the shell is corrected to be smooth and continuous, and then the optimized parameters are imported into the finite element software for checking, so as to ensure that the ultimate bearing capacity and maximum displacement deviation of the shell meet the engineering requirements, avoid unreasonable parameters leading to structural performance problems, and effectively improve the overall stability performance and engineering design efficiency of the aluminum alloy single-layer cylindrical shell.

[0078] It should be noted that, in this document, the terms such as first and second are used merely to distinguish one entity or action from another, and do not necessarily require or imply that these entities or actions occur in any such actual relationship or order. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such process, method, article or device.

[0079] Although embodiments of the application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatic optimization of node parameters of a latticed shell structure based on a neural network model, characterized in that, The application relates to a method for optimizing node parameters of a single-layer cylindrical shell, and belongs to the technical field of structural optimization. The method comprises the following steps: S1, collecting an "input-output" data pair of a single-layer cylindrical shell constructed by an aluminum alloy plate node, wherein the input parameters include design parameters and initial node parameters of the single-layer cylindrical shell, and the output parameters include optimized node vertical coordinate adjustment values and node bending stiffness adjustment values; S2, performing normalization processing on the input parameters to eliminate dimensional differences and form a training data set of a neural network model; The "input-output" data pair is obtained by optimizing the node under different parameters through a genetic algorithm combined with finite element software ANSYS and a programming tool MATLAB, and 100 groups of different parameter combinations of the shell optimization examples are supplemented to ensure that the training data set covers the design parameter range commonly used in engineering; the normalization processing includes converting the coordinate parameters and the node bending stiffness parameters of the node into relative values to eliminate the influence of the dimensional differences on the training results of the neural network model; S3, constructing a three-layer fully connected neural network model, wherein the model comprises an input layer, a hidden layer and an output layer, and the hidden layer adopts a ReLU activation function to enhance the nonlinear fitting capability of the model; S4, training the neural network model by using the training data set, wherein the training includes dividing the training data set into a training set, a verification set and a test set, and an early stopping strategy is adopted to avoid model overfitting; S5, verifying the neural network model, wherein the verification includes comparing the optimized node parameters output by the model with traditional optimization results, and if the deviation of the ultimate bearing capacity of the single-layer cylindrical shell corresponding to the optimized node parameters is less than or equal to 5% and the maximum displacement deviation is less than or equal to 8%, the model verification is passed; S6, applying the neural network model to optimize node parameters, wherein the application includes inputting design parameters of a target single-layer cylindrical shell, and the neural network model automatically outputs optimized node vertical coordinate adjustment values and node bending stiffness adjustment coefficients; When the neural network model is applied to optimize node parameters, the design parameters of the target single-layer cylindrical shell are inputted, including a rise-span ratio, a length-span ratio, a load ratio, a support condition, initial node parameters and shell span information, and the neural network model automatically outputs optimized node y-coordinate adjustment values and node bending stiffness adjustment coefficients; S7, based on a cubic spline interpolation method, correcting a cross-section curve of the single-layer cylindrical shell according to the node vertical coordinate adjustment values, so that the corrected curve is smooth and continuous; 2. The method of claim 1, wherein the method is performed by a computer system. S8, inputting the optimized node parameters into finite element software to perform bearing capacity checking, and if the bearing capacity checking does not meet the engineering requirements, the parameters are inputted again for fine adjustment and optimization. The input parameters include a rise-span ratio, a length-span ratio, a half-span load ratio, a support condition, initial node x-coordinates, initial node y-coordinates, initial node bending stiffness and shell span information of the single-layer cylindrical shell, wherein the set range of the rise-span ratio includes 1 / 6, 1 / 5, 1 / 4 and 1 / 3, the set range of the length-span ratio includes 1.4, 1.8, 2.2 and 2.6, the set range of the half-span load ratio includes 0, 1 / 4, 1 / 2 and 1, and the support condition includes two longitudinal edge supports and four edge supports.

3. The method of claim 1, wherein the method further comprises: The input layer node number is 8, and the 8 input parameters correspond to the rise-span ratio, length-span ratio, half-span load ratio, support condition coding value, initial node x coordinate, initial node y coordinate, initial node bending stiffness and shell span of the shell in turn.

4. The method of claim 1, wherein the method further comprises: In the three-layer fully connected neural network model, the hidden layer is set to two layers, wherein the first layer hidden layer node number is 64, and the second layer hidden layer node number is 32.

5. The method of claim 4, wherein the method further comprises: The output layer node number is 2, corresponding to the optimized node y coordinate adjustment value and node bending stiffness adjustment coefficient, the neural network model adopts mean square error function as the loss function to measure the deviation of the optimized parameters of the model output and the actual parameters, and adopts Adam optimizer, and the learning rate of the Adam optimizer is initially set to 0.

001.

6. The method of claim 1, wherein the method further comprises: In S4, during the training process of the neural network model using the training data set, the complete data set is divided into a training set, a validation set and a test set in a ratio of 7:2:1, and the training round number of the neural network model is set to 500 rounds.

7. The method of claim 5, wherein the method further comprises: The condition of the early stopping strategy is that the validation set loss does not decrease for 20 consecutive rounds, and then the training is stopped, so as to avoid overfitting phenomenon of the neural network model; the Adam optimizer adopts a learning rate decay strategy, specifically, the learning rate is decayed to 0.9 of the original learning rate every 100 iterations, which is used to improve the stability of the neural network model training.

8. The method of claim 1, wherein the method further comprises: The verification includes selecting a typical example in the test set, the input parameters of the typical example include the parameter combination of the shell model with rise-span ratio 1 / 6, length-span ratio 2.6, load ratio 0 and four-side support, the input parameters are input into the trained neural network model, the optimized node parameters are output, and the output is compared with the optimization result of the traditional genetic algorithm; if the shell ultimate bearing capacity deviation or the maximum displacement deviation exceeds the preset range, the hidden layer node number of the three-layer fully connected neural network model or the learning rate of the Adam optimizer used in the training of the neural network model is adjusted, and the model is retrained until the accuracy requirement is met.

Images