Method for optimizing design of v-shaped cylinder of portal crane based on reinforcement learning

Abstract

The application belongs to the technical field of crane design, and particularly relates to a portal crane V-shaped cylinder optimization design method based on reinforcement learning, which comprises the following steps: S1. parameterized modeling and data set establishment; S2. graph neural network proxy model construction; S3. reinforcement learning multi-objective optimization based on action weight improvement of Sobol sensitivity analysis; and S4. optimal solution set output. The V-shaped combined structure changes the traditional single large cylinder force bearing mode, so that the bending moment and the torque can be converted into axial force through the diagonal bracing, which is beneficial to reduce the local stress, and under the premise of meeting the strength and stiffness, the design mode of simply relying on thickening cylinder wall and reinforcing rib can be reduced, and the structure weight and material cost are reduced.

Description

Technical Field

[0001] This application belongs to the field of crane design technology, specifically involving a reinforcement learning-based optimization design method for the V-shaped cylinder of a gantry crane. Background Technology

[0002] Gantry cranes are used in port operations and cargo transportation. The slewing section is not only connected to the lower support structure through the cylindrical structure, but also bears various loads such as the self-weight of the upper slewing structure, lifting load, and wind load. The reliability of the cylindrical structure determines the safety of the entire machine during operation. Currently, the most common cylindrical structure adopts a large-diameter cylindrical shell structure, with internal annular reinforcing ribs added to improve local stability.

[0003] When ports are in harsh environments such as strong winds, in order to enhance structural reliability, the common design of large-diameter cylindrical shell structures requires increasing the wall thickness of the cylinder or adding a large number of reinforcing ribs. However, this leads to an increase in the structural weight, increases material consumption and production costs, and reduces structural reliability. Summary of the Invention

[0004] Based on the above problems, this application proposes an optimized design method that ensures structural bending and torsional performance while reducing wind load and manufacturing costs. The technical solution is as follows:

[0005] A reinforcement learning-based optimization design method for the V-shaped cylinder of a gantry crane includes the following steps: S1. Parametric Modeling and Dataset Establishment: Select parameterized design variables and generate N sample points in the variable space; S2. Construction of Graph Neural Network Agent Model: The V-shaped structure is a parallel structure at the stress and physical connection points. The upper platform, square tube, left diagonal brace, right diagonal brace, lower platform, and local connection area are defined as nodes, and the force transmission relationship between them is defined as the edge construction graph structure. S3. Reinforcement learning multi-objective optimization based on action weights improved by Sobol sensitivity analysis: After obtaining the comprehensive sensitivity index through weighting, action weights are introduced to adjust the baseline adjustment vector for different samples, thereby obtaining the actual adjustment vector. S4. Output of the optimal solution set: Using the actual adjustment vector of S3, the proxy model described in S2 is called to evaluate and update the state and calculate the reward. The interaction data is stored in the experience pool for training. The process is iterated continuously. When the convergence criterion is reached, the optimal solution of the structural parameters is obtained.

[0006] Preferably, the parametric design variable is selected as: square tube wall thickness. diagonal brace wall thickness outer diameter of the upper platform upper platform aperture Angle of diagonal brace Angle between the lower platform Platform height Square tube height Lower platform height ; Generate N sample points in the variable space. This ensures that the sample points uniformly cover the entire design space, and each sample is a set of input parameters. ; For each sample Establish the corresponding V-shaped structural geometric model and calculate the structural self-weight corresponding to the sample parameters. And calculate the response of the V-shaped structure corresponding to each sample under the lifting load Q: maximum equivalent stress Maximum structural displacement and wind load factor The corresponding response vector is obtained. : ; in, It is the i-th sample The corresponding response vector of the V-shaped structure, It is the i-th sample The maximum equivalent stress It is the i-th sample The maximum structural displacement, It is the i-th sample The wind load factor, It is the i-th sample The structural self-weight; Therefore, the training dataset is constructed as follows: ; in, This is the training dataset, used for training the graph neural network described in S2. It is the i-th sample. It is the i-th sample The corresponding response vector of the V-shaped structure; This represents the total number of sample points in the training dataset.

[0007] Preferably, the graph neural network constructs a graph structure based on the structure of a V-shaped cylinder, where nodes represent individual components and edges represent the force transmission relationships between components. The network learns the mutual influence between components through a message passing mechanism. After multiple layers of message passing and feature aggregation, all node features are processed by the Readout function to output the predicted structural response value of the i-th sample. Thus, the network can learn arrive Nonlinear mapping; ; in, It is a node go through The final feature vector after layer message passing, It is the response prediction vector of the graph neural network for the i-th sample. This represents the total number of message passing layers. This is the graph-level average pooling function. The maximum equivalent stress after processing the i-th sample by the Readout function Maximum structural displacement Wind load factor Structural self-weight The predicted values ​​are used for subsequent multi-objective optimization.

[0008] Preferably, in step S3, the Sobol sensitivity analysis is performed. First, evaluate the impact of each design variable on each response objective using a training dataset. Given N samples, calculate the Sobol sensitivity of each design variable to each response objective, and obtain... For maximum equivalent stress Maximum structural displacement Wind load factor Structural self-weight global sensitivity , , , ; ; ; ; ; in, , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For maximum equivalent stress Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For the maximum structural displacement Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For wind load factor Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height Regarding the structural self-weight Sobol sensitivity; To evaluate the combined impact of each design variable on the multi-response objective, a weighted synthesis was performed to obtain the overall sensitivity. This provides a unified action reference for reinforcement learning: ; ; in, For overall sensitivity, , , , These are the maximum equivalent stresses. Sobol sensitivity, maximum structural displacement Sobol sensitivity, wind load factor Sobol sensitivity, structural weight Sobol sensitivity; , , and These are the weighting coefficients, ; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height The overall sensitivity.

[0009] Preferably, the state of the i-th sample is defined as : ; in, Let i be the state of the i-th sample, i = 1, 2, ..., N. For the i-th sample, This is the predicted structural response value of the i-th sample after processing by the Readout function. For overall sensitivity; Introducing action weights To adjust the amplitude of the action of different design variables, , … Mapped to respectively The interval is used to obtain the action weights of different design variables: ; ; ... ;

[0010] in , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height Action weights, This is the minimum threshold for the action weight. , The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height The overall sensitivity, Sobol sensitivity is the variance contribution rate, which is always greater than 0 and does not exceed 1, and each design variable has an impact on each response objective, therefore... Heng is established; The agent determines the state of the i-th sample. The baseline adjustment vector for the i-th sample is obtained through policy network calculation. ,Right now: ; Let Δ be the baseline adjustment vector for the i-th sample, and let Δ be the wall thickness of the square tube. diagonal brace wall thickness outer diameter of the upper platform upper platform aperture Angle of diagonal brace Angle between the lower platform Platform height Square tube height Lower platform height , The policy function for reinforcement learning is implemented by a policy network. Let i be the state of the i-th sample; Define Action Let be the actual adjustment vector for the i-th sample, i.e.: ; in, For action weights, Let be the baseline adjustment vector for the i-th sample. This indicates element-wise multiplication. This is the actual adjustment vector for the i-th sample; definition Let be the new set of input parameters for the i-th sample after actual adjustment, i.e.: ; in, Let i be the set of new input parameters after actual adjustment for the i-th sample. For the set of input parameters, Let be the actual adjustment vector for the i-th sample.

[0011] Preferred, After processing by the Readout function, the predicted values ​​of maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight are obtained. , , , To maintain a consistent expression for minimizing the predicted maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight, normalization is performed. The normalized maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight are denoted as follows: , , and Instant rewards are written as: ; in, For the immediate reward of the agent at step t, , , and These are the weighting coefficients, , The maximum equivalent stress predicted by the graph neural network The maximum equivalent stress after normalization The maximum structural displacement predicted by the graph neural network The maximum structural displacement after normalization Wind load coefficient predicted by graph neural network Normalized wind load factor The structural self-weight predicted by the graph neural network The normalized structural weight; The constraint violation penalty for the agent at step t is determined when the maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight exceed the limits. It is positive otherwise and is used to guide the agent away from infeasible areas; Using the actual adjustment vector in S3, the graph neural network surrogate model described in S2 is called to evaluate and update the state and calculate the reward. Data generated from each interaction is stored in an experience pool. During training, small batches of data are randomly sampled from this pool, and the weight parameters of the neural network are updated using these sampled data. If the convergence condition is not met, the process returns to continue the reinforcement learning iteration until the preset convergence condition is met. Then, the process proceeds to S4, where the optimal structure parameters are output. Sobol sensitivity is evaluated using the training dataset before the first iteration. This indicates that the range of values ​​for each design variable does not exceed the training dataset during subsequent iterations. When covering the design space, the Sobol sensitivity calculated initially is used. If the range of values ​​for each design variable exceeds the design space in subsequent iterations, the design space is re-planned with the current optimal sample point as the center, and a new variable space is generated within the variable space. A new training dataset is constructed from 1 sample point. The Sobol sensitivity was recalculated.

[0012] Preferably, the Pareto non-dominated solution set is used as the candidate design scheme, while retaining the optimal results under multiple different trade-offs, which facilitates subsequent screening for different operating conditions.

[0013] Compared to existing solutions, this application makes the following contributions: 1) The V-shaped composite structure changes the traditional single large cylinder load-bearing mode, so that bending moment and torque can be converted into axial force through diagonal bracing, which helps to reduce local stress. Under the premise of meeting strength and stiffness, it can reduce the design method of simply relying on thickening the cylinder wall and strengthening ribs, thereby reducing structural weight and material cost.

[0014] 2) The V-shaped structure is more conducive to controlling the windward projection area and reducing wind load.

[0015] 3) Graph neural network surrogate models can utilize the topological relationships of component connections to fit complex nonlinear relationships, and have higher modeling accuracy and better generalization performance than traditional regression methods.

[0016] 4) Sobol global sensitivity provides physical inspiration for reinforcement learning, enabling optimization actions to focus more on key parameters, reducing blind searches, and improving convergence efficiency and result interpretability. Attached Figure Description

[0017] Figure 1 This is a three-dimensional view of the cylinder. Figure 2This is a schematic diagram of a diagonal brace; Figure 3 This is a cross-sectional view of the square tube; Figure 4 To optimize the design flowchart; In the diagram: 1-Upper platform, 2-Diagonal brace, 3-Square tube, 4-Lower platform. Detailed Implementation

[0018] This invention uses the optimization of the V-shaped cylinder of a gantry crane as a specific implementation scenario for illustration.

[0019] This invention relates to a reinforcement learning-based optimization design method for the V-shaped cylinder of a gantry crane, achieving multi-objective optimization of the maximum equivalent stress, maximum structural displacement, wind load coefficient, and structural self-weight under lifting load.

[0020] A reinforcement learning-based optimization design method for the V-shaped cylinder of a gantry crane includes the following steps: S1. Parametric Modeling and Dataset Establishment: Select parameterized design variables and generate N sample points in the variable space; S2. Construction of Graph Neural Network Agent Model: The V-shaped structure is a parallel structure at the stress and physical connection points. The upper platform, square tube, left diagonal brace, right diagonal brace, lower platform, and local connection area are defined as nodes, and the force transmission relationship between them is defined as the edge construction graph structure. S3. Reinforcement learning multi-objective optimization based on action weights improved by Sobol sensitivity analysis: After obtaining the comprehensive sensitivity index through weighting, action weights are introduced to adjust the baseline adjustment vector for different samples, thereby obtaining the actual adjustment vector. S4. Output of the optimal solution set: Using the actual adjustment vector of S3, the proxy model described in S2 is called to evaluate and update the state and calculate the reward. The interaction data is stored in the experience pool for training. The process is iterated continuously. When the convergence criterion is reached, the optimal solution of the structural parameters is obtained.

[0021] Selecting parametric design variables: square tube wall thickness diagonal brace wall thickness outer diameter of the upper platform upper platform aperture Angle of diagonal brace Angle between the lower platform Platform height Square tube height Lower platform height ; Generate N sample points in the variable space. This ensures that the sample points uniformly cover the entire design space, and each sample is a set of input parameters. ; For each sample Establish the corresponding V-shaped structural geometric model and calculate the structural self-weight corresponding to the sample parameters. And calculate the response of the V-shaped structure corresponding to each sample under the lifting load Q: maximum equivalent stress Maximum structural displacement and wind load factor The corresponding response vector is obtained. : ; in, It is the i-th sample The corresponding response vector of the V-shaped structure, It is the i-th sample The maximum equivalent stress It is the i-th sample The maximum structural displacement, It is the i-th sample The wind load factor, It is the i-th sample The structural self-weight; Therefore, the training dataset is constructed as follows: ; in, This is the training dataset, used for training the graph neural network described in S2. It is the i-th sample. It is the i-th sample The corresponding response vector of the V-shaped structure; This represents the total number of sample points in the training dataset.

[0022] The graph neural network (GNN) constructs a graph structure based on a V-shaped cylinder, where nodes represent individual components and edges represent the force transmission relationships between components. It learns the mutual influence between components through a message passing mechanism. After multiple layers of message passing and feature aggregation, all node features are processed by the Readout function, outputting the predicted structural response value of the i-th sample. Thus, the network can learn arrive Nonlinear mapping; ; in, It is a node go through The final feature vector after layer message passing, It is the response prediction vector of the graph neural network for the i-th sample. This represents the total number of message passing layers. This is the graph-level average pooling function. The maximum equivalent stress after processing the i-th sample by the Readout function Maximum structural displacement Wind load factor Structural self-weight The predicted values ​​are used for subsequent multi-objective optimization.

[0023] Sobol sensitivity analysis in step S3: First, evaluate the impact of each design variable on each response objective using a training dataset. Given N samples, calculate the Sobol sensitivity of each design variable to each response objective, and obtain... For maximum equivalent stress Maximum structural displacement Wind load factor Structural self-weight global sensitivity , , , ; ; ; ; ; in, , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For maximum equivalent stress Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For the maximum structural displacement Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For wind load factor Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height Regarding the structural self-weight Sobol sensitivity.

[0024] To evaluate the combined impact of each design variable on the multi-response objective, a weighted synthesis was performed to obtain the overall sensitivity. This provides a unified action reference for reinforcement learning: ; ; in, For overall sensitivity, , , , These are the maximum equivalent stresses. Sobol sensitivity, maximum structural displacement Sobol sensitivity, wind load factor Sobol sensitivity, structural weight Sobol sensitivity; , , and These are the weighting coefficients, ; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height The overall sensitivity.

[0025] The state of the i-th sample is defined as : ; in, Let i be the state of the i-th sample, i = 1, 2, ..., N. For the i-th sample, This is the predicted structural response value of the i-th sample after processing by the Readout function. For overall sensitivity; Introducing action weights To adjust the amplitude of the action of different design variables, , … Mapped to respectively The interval is used to obtain the action weights of different design variables: ; ; ... ; ; in , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height Action weights, This is the minimum threshold for the action weight. , The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height The overall sensitivity, Sobol sensitivity is the variance contribution rate, which is always greater than 0 and does not exceed 1, and each design variable has an impact on each response objective, therefore... Heng was established.

[0026] The agent determines the state of the i-th sample. The baseline adjustment vector for the i-th sample is obtained through policy network calculation. ,Right now: ; Let Δ be the baseline adjustment vector for the i-th sample, and let Δ be the wall thickness of the square tube. diagonal brace wall thickness outer diameter of the upper platform upper platform aperture Angle of diagonal brace Angle between the lower platform Platform height Square tube height Lower platform height , The policy function for reinforcement learning is implemented by a policy network. Let be the state of the i-th sample.

[0027] Define Action Let be the actual adjustment vector for the i-th sample, i.e.: ; in, For action weights, Let be the baseline adjustment vector for the i-th sample. This indicates element-wise multiplication. This is the actual adjustment vector for the i-th sample; definition Let be the new set of input parameters for the i-th sample after actual adjustment, i.e.: ; in, Let i be the set of new input parameters after actual adjustment for the i-th sample. For the set of input parameters, Let be the actual adjustment vector for the i-th sample.

[0028] After processing by the Readout function, the predicted values ​​of maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight are obtained. , , , To maintain a consistent expression for minimizing the predicted maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight, normalization is performed. The normalized maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight are denoted as follows: , , and Instant rewards are written as: ; in, For the immediate reward of the agent at step t, , , and These are the weighting coefficients, , The maximum equivalent stress predicted by the graph neural network The maximum equivalent stress after normalization The maximum structural displacement predicted by the graph neural network The maximum structural displacement after normalization Wind load coefficient predicted by graph neural network Normalized wind load factor The structural self-weight predicted by the graph neural network The normalized structural weight; The constraint violation penalty for the agent at step t is determined when the maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight exceed the limits. It is positive if it is positive, otherwise it is 0, and is used to guide the agent away from infeasible areas.

[0029] Using the actual adjustment vector in S3, the graph neural network surrogate model described in S2 is called to evaluate and update the state and calculate the reward. Data generated from each interaction is stored in an experience pool. During training, small batches of data are randomly sampled from this pool, and the weight parameters of the neural network are updated using these sampled data. If the convergence condition is not met, the process returns to continue the reinforcement learning iteration until the preset convergence condition is met. Then, the process proceeds to S4, where the optimal structure parameters are output. Sobol sensitivity is evaluated using the training dataset before the first iteration. This indicates that the range of values ​​for each design variable does not exceed the training dataset during subsequent iterations. When covering the design space, the Sobol sensitivity calculated initially is used. If the range of values ​​for each design variable exceeds the design space in subsequent iterations, the design space is re-planned with the current optimal sample point as the center, and a new variable space is generated within the variable space. A new training dataset is constructed from 1 sample point. The Sobol sensitivity was recalculated.

[0030] Pareto non-dominated solution sets are used as candidate design schemes, while retaining the optimal results under multiple different trade-offs, which facilitates subsequent screening for different operating conditions.

Claims

1. A method for optimizing the design of a V-shaped cylinder of a gantry crane based on reinforcement learning, characterized in that, Includes the following steps: S1. Parametric Modeling and Dataset Establishment: Select parameterized design variables and generate N sample points in the variable space; S2. Construction of Graph Neural Network Agent Model: The V-shaped structure is a parallel structure at the stress and physical connection points. The upper platform, square tube, left diagonal brace, right diagonal brace, lower platform, and local connection area are defined as nodes, and the force transmission relationship between them is defined as the edge construction graph structure. S3. Reinforcement learning multi-objective optimization based on action weights improved by Sobol sensitivity analysis: After obtaining the comprehensive sensitivity index through weighting, action weights are introduced to adjust the baseline adjustment vector for different samples, thereby obtaining the actual adjustment vector. S4. Output of the optimal solution set: Using the actual adjustment vector of S3, the proxy model described in S2 is called to evaluate and update the state and calculate the reward. The interaction data is stored in the experience pool for training. The process is iterated continuously. When the convergence criterion is reached, the optimal solution of the structural parameters is obtained.

2. The reinforcement learning-based V-shaped cylinder optimization design method for gantry cranes according to claim 1, characterized in that, Selecting parametric design variables: square tube wall thickness diagonal brace wall thickness outer diameter of the upper platform upper platform aperture Angle of diagonal brace Angle between the lower platform Platform height Square tube height Lower platform height ; Generate N sample points in the variable space. This ensures that the sample points uniformly cover the entire design space, and each sample is a set of input parameters. ; For each sample Establish the corresponding V-shaped structural geometric model and calculate the structural self-weight corresponding to the sample parameters. And calculate the response of the V-shaped structure corresponding to each sample under the lifting load Q: maximum equivalent stress Maximum structural displacement and wind load factor The corresponding response vector is obtained. : ; in, It is the i-th sample The corresponding response vector of the V-shaped structure, It is the i-th sample The maximum equivalent stress It is the i-th sample The maximum structural displacement, It is the i-th sample The wind load factor, It is the i-th sample The structural self-weight; Therefore, the training dataset is constructed as follows: ； in, This is the training dataset, used for training the graph neural network described in S2. It is the i-th sample. It is the i-th sample The corresponding response vector of the V-shaped structure; This represents the total number of sample points in the training dataset.

3. The reinforcement learning-based V-shaped cylinder optimization design method for gantry cranes according to claim 2, characterized in that, The graph neural network (GNN) constructs a graph structure based on a V-shaped cylinder, where nodes represent individual components and edges represent the force transmission relationships between components. It learns the mutual influence between components through a message passing mechanism. After multiple layers of message passing and feature aggregation, all node features are processed by the Readout function to output the predicted structural response value of the i-th sample. Thus, the network can learn arrive Nonlinear mapping; ； in, It is a node go through The final feature vector after layer message passing, It is the response prediction vector of the graph neural network for the i-th sample. This represents the total number of message passing layers. This is the graph-level average pooling function. The maximum equivalent stress after processing the i-th sample by the Readout function Maximum structural displacement Wind load factor Structural self-weight The predicted values ​​are used for subsequent multi-objective optimization.

4. The reinforcement learning-based V-shaped cylinder optimization design method for gantry cranes according to claim 3, characterized in that, Sobol sensitivity analysis in step S3: First, evaluate the impact of each design variable on each response objective using a training dataset. Given N samples, calculate the Sobol sensitivity of each design variable to each response objective, and obtain... For maximum equivalent stress Maximum structural displacement Wind load factor Structural self-weight global sensitivity , , , ; ； ； ； ； in, , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For maximum equivalent stress Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For the maximum structural displacement Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height For wind load factor Sobol sensitivity; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height Regarding the structural self-weight Sobol sensitivity; To evaluate the combined impact of each design variable on the multi-response objective, a weighted synthesis was performed to obtain the overall sensitivity. This provides a unified action reference for reinforcement learning: ； ； in, For overall sensitivity, , , , These are the maximum equivalent stresses. Sobol sensitivity, maximum structural displacement Sobol sensitivity, wind load factor Sobol sensitivity, structural weight Sobol sensitivity; , , and These are the weighting coefficients, ; , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height The overall sensitivity.

5. The reinforcement learning-based V-shaped cylinder optimization design method for gantry cranes according to claim 4, characterized in that, The state of the i-th sample is defined as : ； in, Let i be the state of the i-th sample, i = 1, 2, ..., N. For the i-th sample, This is the predicted structural response value of the i-th sample after processing by the Readout function. For overall sensitivity; Introducing action weights To adjust the amplitude of the action of different design variables, , … Mapped to respectively The interval is used to obtain the action weights of different design variables: ； ； ⋮ ； ； in , … The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height Action weights, This is the minimum threshold for the action weight. , The wall thickness of the square tube is respectively diagonal brace wall thickness ..., lower platform height The overall sensitivity, Sobol sensitivity is the variance contribution rate, which is always greater than 0 and does not exceed 1, and each design variable has an impact on each response objective, therefore... Hengcheng was established; The agent determines the state of the i-th sample. The baseline adjustment vector for the i-th sample is obtained through policy network calculation. ,Right now: ； Let Δ be the baseline adjustment vector for the i-th sample, and let Δ be the wall thickness of the square tube. diagonal brace wall thickness outer diameter of the upper platform upper platform aperture Angle of diagonal brace Angle between the lower platform Platform height Square tube height Lower platform height , The policy function for reinforcement learning is implemented by a policy network. Let i be the state of the i-th sample; Define Action Let be the actual adjustment vector for the i-th sample, i.e.: ； in, For action weights, Let be the baseline adjustment vector for the i-th sample. This indicates element-wise multiplication. This is the actual adjustment vector for the i-th sample; definition Let be the new set of input parameters for the i-th sample after actual adjustment, i.e.: ； in, Let i be the set of new input parameters after actual adjustment for the i-th sample. For the set of input parameters, Let be the actual adjustment vector for the i-th sample.

6. The reinforcement learning-based V-shaped cylinder optimization design method for gantry cranes according to claim 5, characterized in that, After processing by the Readout function, the predicted values ​​of maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight are obtained. , , , To maintain a consistent expression for minimizing the predicted maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight, normalization is performed. The normalized maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight are denoted as follows: , , and Instant rewards are written as: ； in, For the immediate reward of the agent at step t, , , and These are the weighting coefficients, , The maximum equivalent stress predicted by the graph neural network The maximum equivalent stress after normalization The maximum structural displacement predicted by the graph neural network The maximum structural displacement after normalization Wind load coefficient predicted by graph neural network Normalized wind load factor The structural self-weight predicted by the graph neural network The normalized structural weight; The constraint violation penalty for the agent at step t is determined when the maximum equivalent stress, maximum structural displacement, wind load factor, and structural self-weight exceed the limits. It is positive otherwise and is used to guide the agent away from infeasible areas; Using the actual adjustment vector in S3, the graph neural network surrogate model described in S2 is called to evaluate and update the state and calculate the reward. Data generated from each interaction is stored in an experience pool. During training, small batches of data are randomly sampled from this pool, and the weight parameters of the neural network are updated using these sampled data. If the convergence condition is not met, the process returns to continue the reinforcement learning iteration until the preset convergence condition is met, at which point the process proceeds to S4 to output the optimal structure parameters. The Sobol sensitivity is evaluated using the training dataset before the first iteration. This indicates that the range of values ​​for each design variable does not exceed the training dataset during subsequent iterations. When covering the design space, the Sobol sensitivity calculated initially is used. If the range of values ​​for each design variable exceeds the design space in subsequent iterations, the design space is re-planned with the current optimal sample point as the center, and a new variable space is generated within the variable space. A new training dataset is constructed from 1 sample point. The Sobol sensitivity was recalculated.

7. The reinforcement learning-based V-shaped cylinder optimization design method for gantry cranes according to claim 6, characterized in that, Pareto non-dominated solution sets are used as candidate design schemes, while retaining the optimal results under multiple different trade-offs, which facilitates subsequent screening for different operating conditions.

Images