An intelligent design method and system for an eccentrically compressed column of an externally wrapped UHPC and a medium

Abstract

This disclosure relates to an intelligent design method, system, and medium for outsourced UHPC eccentrically compressed columns, belonging to the field of structural engineering technology. The method includes constructing a deep reinforcement learning network model; obtaining training samples and inputting them into the model; calculating the bearing capacity, cost, and compliance with engineering specifications using combinations of predicted design parameters obtained from the training samples; calculating a reward signal for training using the reward function and updating the network parameters; obtaining the load parameters of the outsourced UHPC eccentrically compressed column to be designed, and obtaining multiple sets of candidate design parameter combinations using the trained deep reinforcement learning network model; calculating the bearing capacity and checking against engineering specifications for the candidate design parameter combinations; and selecting the final design scheme based on the reward signal, the bearing capacity calculation results, and the engineering specification check results. This disclosure realizes intelligent design of outsourced UHPC eccentrically compressed columns, enabling rapid and efficient acquisition of the safe and economical optimal design.

Description

Technical Field

[0001] This disclosure relates to the fields of deep reinforcement learning and structural engineering technology, and in particular to an intelligent design method, system and medium for outsourced UHPC eccentrically compressed columns. Background Technology

[0002] Bridge engineering is a crucial component of the transportation engineering system, and the safety and economy of its structure directly impact the life-cycle cost of the entire road network and traffic safety. Bridge piers are vital load-bearing components of the bridge superstructure, and their design rationality is paramount. In recent years, with the continuous development of Ultra-High Performance Concrete (UHPC), UHPC-encased eccentrically compressed columns, consisting of a core concrete, an outer UHPC casing, and reinforcing steel, have gained widespread application due to their excellent mechanical and durability properties.

[0003] However, the current design of outsourced UHPC eccentric compression columns mainly relies on manual design, which is an extremely complex process and faces the following technical challenges: The mechanical calculations are complex, and the trial calculation process is slow: Calculating the bearing capacity of an eccentrically compressed UHPC-encased column under eccentric compression involves complex nonlinear equilibrium equations. As a novel structure, designers need to pre-guess the cross-sectional dimensions, reinforcement size, and thickness of the UHPC layer during trial calculations, then substitute these into empirical formulas to calculate the ultimate bearing capacity. If the requirements are not met, the parameters must be readjusted and the calculation repeated. This iterative trial calculation process is time-consuming and labor-intensive. The design variables of the UHPC eccentrically compressed column are complex, including column diameter, outer casing thickness, rebar diameter, rebar quantity, core concrete strength grade, and ultra-high performance concrete strength grade. These design variables are interconnected and are all discrete. Traditional optimization algorithms (such as genetic algorithms and particle swarm optimization) are prone to getting trapped in local optima when solving discrete and high-dimensional search problems, resulting in slow convergence and difficulty in quickly finding the optimal balance between safety and economic efficiency.

[0004] Although machine learning is often used in structural design, there is no intelligent design scheme for outsourced UHPC eccentric compression columns. Therefore, there is an urgent need to design a method and system for intelligent design of outsourced UHPC eccentric compression columns. Summary of the Invention

[0005] Therefore, it is necessary to provide an intelligent design method, system, and medium for outsourced UHPC eccentric compression columns to address the problem that the lack of intelligent design solutions for outsourced UHPC eccentric compression columns makes the design process time-consuming and labor-intensive, and makes it difficult to quickly find the optimal design that is both safe and economical.

[0006] To solve the above problems, the present disclosure adopts the following technical solution: In a first aspect, this disclosure provides a smart design method for outsourced UHPC eccentrically compressed columns, including: A deep reinforcement learning network model is constructed, and the state space, action space, and reward function of the network model are defined. The load parameters of the state space include the computational length, eccentricity, and axial force design value. The design parameters of the action space include the column cross-section diameter, the thickness of the outer UHPC, the diameter of the reinforcing bars, the number of reinforcing bars, the core concrete strength, and the outer UHPC strength. Obtain training samples, The training samples are input into the deep reinforcement learning network model, which is used to output a combination of design parameter prediction values. The network model includes a shared feature extraction layer for extracting shared features from the load parameters and an independent action output layer whose number is equal to the number of design parameters in the action space. The independent action output layer corresponds one-to-one with the number of design parameters in the action space and is used to process the shared features and output the design parameter prediction value of one design parameter. Using the combination of predicted design parameters obtained from training samples, the bearing capacity and cost of the external UHPC eccentrically compressed column are calculated, and the first calculation and inspection results are obtained by checking whether the external UHPC eccentrically compressed column meets the engineering specifications. Based on the first calculation and inspection results, the reward signal used for training is calculated using the reward function, and the network parameters of the network model are updated based on the reward signal used for training. Obtain the load parameters of the outsourced UHPC eccentrically compressed column to be designed, and use the trained deep reinforcement learning network model to obtain multiple sets of candidate design parameter combinations; For each combination of candidate design parameters, the bearing capacity of the enclosed UHPC eccentrically compressed column is calculated and checked against engineering specifications. Based on the reward signals corresponding to the candidate design parameter combinations, the bearing capacity calculation results, and the engineering specification inspection results, the optimal candidate design parameter combination among all candidate design parameter combinations is selected as the final design scheme.

[0007] In a preferred embodiment, the step of checking whether the external UHPC eccentric compression column meets the engineering specifications specifically involves checking whether the slenderness ratio, reinforcement ratio, and steel bar spacing of the external UHPC eccentric compression column all meet the engineering specifications. The calculation of the bearing capacity and cost of the externally packaged UHPC eccentrically compressed column specifically includes: Calculate the eccentricity amplification factor at the ultimate limit state of bearing capacity; Establish the moment equilibrium equation and axial force equilibrium equation for the eccentrically compressed UHPC column, and use the bisection method to solve for the half-pressure angle; Based on the semi-compression angle, the design value of the core concrete strength, the design value of the steel reinforcement strength, and the section parameters, the axial bearing capacity of the enclosed UHPC eccentrically compressed column is calculated. Calculate the cost of core concrete, external UHPC, and steel reinforcement. Based on the costs of core concrete, external UHPC, and steel reinforcement, calculate the cost of the external UHPC eccentrically compressed column.

[0008] In a preferred embodiment, the moment balance equation is: The axial force balance equation is: The bisection method for solving the half-pressure angle involves: simultaneously solving the equilibrium equations and calculating the residual function. , Find the solution using the cyclic binary search method. The half pressure angle; in, It is the half-pressure angle; This represents the balance error between the eccentricity of internal forces and the eccentricity of external forces within the cross section. Representing the moment equilibrium equation, Represent the axial force equilibrium equation; This is the eccentricity amplification factor; The eccentricity of the axial force about the centroidal axis of the cross section; The core concrete radius; The radius of the circumference where the reinforcing bars are distributed; For reinforcement ratio, This represents the total cross-sectional area of ​​the reinforcing bars. The core concrete cross-sectional area; For the external reinforcement ratio, The cross-sectional area of ​​the UHPC casing; For the outer center radius, For the outer casing thickness; This refers to the design value of the core concrete compressive strength. This is the design value for the strength of the reinforcing steel. This is the design value for the compressive strength of the outer UHPC. This is the design value for the tensile strength of the outer UHPC. It is the height coefficient of the tension zone reinforcement relative to the limiting compression zone.

[0009] In a preferred embodiment, the formula for the reward function is: in, As a reward signal; As a reward for carrying capacity; As a reward for cost-effectiveness; For compressive bearing capacity index, To enhance the axial bearing capacity of the UHPC eccentrically compressed column, This is the design value for axial force; As a cost-performance indicator; Cost of outsourcing UHPC eccentric compression columns; When the result of checking whether the outsourced UHPC eccentrically compressed column meets the engineering specifications is that it meets the engineering specifications, a reward signal is given. The calculation is performed using the formula for the reward function; otherwise, let... ; The carrying capacity reward According to the compressive bearing capacity index The distribution calculation formula is as follows: In satisfying and At that time, the aforementioned cost-effectiveness reward The calculation formula is: .

[0010] In a preferred embodiment, the specific steps of updating the network parameters of the network model based on the reward signal used for training include: The reward samples during training are stored in an experience replay pool. If the reward signal corresponding to a reward sample is greater than a preset high-quality threshold, the reward sample is stored multiple times in the experience replay pool. The reward sample includes the current state vector. Action command vector Reward signals Next state vector and termination mark ; Randomly sample batches of data from the experience replay pool and calculate the total loss between the network model output value and the target value of the target network model. The calculation formula is: in, Batch size; Number the sample; Output the branch number for the policy network; For policy network output branches medium sample of value; For policy network output branches medium sample goal value; Calculation target value The formula is: in, For the sample The reward signal; Discount factor; Number the next action. Indicates all possible next actions Select the maximum value from the list; For target network samples Next state vector and the next action Output target value; For the sample The termination marker; Minimize the total loss using the Adam algorithm. Update the policy network parameters in the model, and synchronize the policy network to the target network every first number of training rounds.

[0011] In a preferred embodiment, during the training process of the deep reinforcement learning network model, the following is employed: - Greedy strategies select actions, specifically including: In the initial exploration phase, using probability Randomly select an action from all available design parameters; As the number of training rounds increases, the probability... Decreasing according to the exponential decay law, satisfying... in, To find the maximum value function, The attenuation coefficient is... This represents the minimum exploration probability.

[0012] In a preferred embodiment, the shared feature extraction layer includes three fully connected layers arranged sequentially, each followed by a ReLU activation function layer; specifically, the shared feature extraction layer is used to process the input state vector... Mapped to a high-dimensional vector The calculation formula is: in, This is the weight matrix of the first fully connected layer. for The corresponding bias vector; This is the weight matrix for the second fully connected layer. for The corresponding bias vector; This is the weight matrix for the third fully connected layer. for The corresponding bias vector; The high-dimensional vector is the ReLU activation function. As the shared feature; The independent action output layer is specifically used to process shared features to obtain the design parameters corresponding to the independent action output layer. value vector , The formula is: in, The number of the independent action output layer; For independent action output layer The weights; For independent action output layer The bias; The deep reinforcement learning network model is used to select each value vector The maximum value is used to generate the combination of predicted design parameter values.

[0013] In a preferred embodiment, the step of selecting the optimal candidate design parameter combination from all candidate design parameter combinations as the final design scheme based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation result, and the engineering specification check result includes: Step 6.1: Filter out all candidate design parameter combinations that do not meet the engineering specifications; Step 6.2: For the retained candidate design parameter combinations, if there is a candidate design parameter combination whose compressive bearing capacity index meets the first compressive bearing capacity index range, then all candidate design parameter combinations that meet the first compressive bearing capacity index range are retained, and the candidate design parameter combination with the largest reward signal among all retained candidate design parameter combinations is selected as the final design scheme; if there is no candidate design parameter combination whose compressive bearing capacity index meets the first compressive bearing capacity index range, then all candidate design parameter combinations that meet the second compressive bearing capacity index range are retained, and the candidate design parameter combination with the largest reward signal among all retained candidate design parameter combinations is selected as the final design scheme.

[0014] Secondly, this disclosure provides an intelligent design system for outsourced UHPC eccentric compression columns, including: The model building module is used to construct a deep reinforcement learning network model, defining the state space, action space, and reward function of the network model; the load parameters of the state space include the computational length, eccentricity, and axial force design value; the design parameters of the action space include the column cross-section diameter, the thickness of the outer UHPC, the diameter of the reinforcing bars, the number of reinforcing bars, the core concrete strength, and the outer UHPC strength. The acquisition and input module is used to acquire training samples and input the training samples into the deep reinforcement learning network model. The network model is used to output a combination of design parameter prediction values, including a shared feature extraction layer for extracting shared features from load parameters and an independent action output layer whose number is equal to the number of design parameters in the action space. The independent action output layer and the number of design parameters in the action space correspond one-to-one, and are used to process the shared features and output the design parameter prediction value of one design parameter. The calculation and update module is used to calculate the bearing capacity and cost of the eccentrically compressed UHPC column using the combination of predicted design parameter values ​​obtained based on training samples, and to check whether the eccentrically compressed UHPC column meets the engineering specifications to obtain the first calculation and check result; it is used to calculate the reward signal for training based on the first calculation and check result and the reward function, and to update the network parameters of the network model based on the reward signal for training. The acquisition and prediction module is used to acquire the load parameters of the eccentrically compressed UHPC column to be designed, and to obtain multiple sets of candidate design parameter combinations using a trained deep reinforcement learning network model. The calculation module is used to perform bearing capacity calculations and engineering code checks on the external UHPC eccentrically compressed columns corresponding to each set of candidate design parameter combinations. The selection module is used to select the optimal candidate design parameter combination from all candidate design parameter combinations as the final design scheme, based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation result, and the engineering specification check result.

[0015] Thirdly, this disclosure provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent design method for an outsourced UHPC eccentrically compressed column described in the first aspect.

[0016] The aforementioned intelligent design method, system, and medium for outsourced UHPC eccentric compression columns realizes the intelligent design of outsourced UHPC eccentric compression columns, solving the problems of time-consuming and labor-intensive processes, and enabling the rapid and efficient acquisition of a safe and economical optimal design. Specifically, by constructing and training a deep reinforcement learning network model, which includes a shared feature extraction layer and multiple independent action output layers, the design parameters of the outsourced UHPC eccentric compression column can be designed autonomously and efficiently. By calculating the bearing capacity and cost, and checking whether the outsourced UHPC eccentric compression column meets engineering specifications, and then using a reward function to calculate the reward signal used for training to update the network parameters of the network model, the intelligent design takes into account multiple objective requirements such as bearing capacity, cost, and engineering specifications, and can obtain the candidate design parameter combination that best balances safety and economy. Based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation result, and the engineering specification check result, the output result of the deep reinforcement learning network model is evaluated to select the optimal design from the candidate design parameter combinations. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method in one embodiment of the present disclosure; Figure 2 This is a perspective view of an externally enclosed UHPC eccentrically compressed column member in one embodiment of this disclosure; Figure 3 This is a top view of an externally packaged UHPC eccentrically compressed column member in one embodiment of this disclosure; Figure 4 This is a scatter plot of the cost-bearing capacity ratio in one embodiment of this disclosure; Figure 5 A load-bearing capacity distribution diagram is designed for one embodiment of this disclosure; Figure 6 This is a schematic diagram of the system structure in one embodiment of the present disclosure. Detailed Implementation

[0018] The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0019] See Figure 1 This embodiment provides a smart design method for an externally packaged UHPC eccentrically compressed column, including: A deep reinforcement learning network model is constructed, and the state space, action space, and reward function of the network model are defined. The load parameters of the state space include the computational length, eccentricity, and axial force design value. The design parameters of the action space include the column cross-section diameter, the thickness of the outer UHPC, the diameter of the reinforcing bars, the number of reinforcing bars, the core concrete strength, and the outer UHPC strength. Acquire training samples and input them into the deep reinforcement learning network model. The network model is used to output a combination of design parameter prediction values. The network model includes a shared feature extraction layer for extracting shared features from load parameters and an independent action output layer whose number is equal to the number of design parameters in the action space. The number of independent action output layers corresponds one-to-one with the number of design parameters in the action space. The independent action output layer is used to process the shared features and output the design parameter prediction value of one design parameter. Using the combination of predicted design parameters obtained from training samples, the bearing capacity and cost of the external UHPC eccentrically compressed column are calculated, and the first calculation and inspection results are obtained to check whether the external UHPC eccentrically compressed column meets the engineering specifications. Based on the first calculation and inspection results, the reward signal is calculated using the reward function, and the network parameters of the network model are updated based on the reward signal. Obtain the load parameters of the outsourced UHPC eccentrically compressed column to be designed, and use the trained deep reinforcement learning network model to obtain multiple sets of candidate design parameter combinations; For each combination of candidate design parameters, the bearing capacity of the enclosed UHPC eccentrically compressed column is calculated and checked against engineering specifications. Based on the reward signal, bearing capacity calculation results, and engineering specification constraint check results, the optimal candidate design parameter combination among all candidate design parameter combinations is selected as the final design scheme.

[0020] In this embodiment, the number of design parameters in the independent action output layer and the action space correspond one-to-one. Each independent action output layer is used to predict one design parameter; that is, multiple independent action output layers are used one-to-one to predict their corresponding design parameter. The independent action output layers process shared features and output predicted design parameter values ​​for the corresponding design parameter items. Specifically, the deep reinforcement learning network model includes six independent action output layers used to process shared features and output predicted design parameter values. These six independent action output layers correspond one-to-one to the prediction of six design parameters: column cross-section diameter, outer UHPC thickness, rebar diameter, number of rebars, core concrete strength, and outer UHPC strength.

[0021] As a specific embodiment, the above method is described in detail below, and the method includes the following steps: Step 1: Construct a deep reinforcement learning network model and define the model's state space, action space, and reward function. The state space (load parameters of the state space) includes the computational length, eccentricity, and axial force design value. The action space (design parameters of the action space) includes the column cross-section diameter, outer UHPC thickness, rebar diameter, number of rebars, core concrete strength, and outer UHPC strength.

[0022] The state space vector consists of three load parameters: the calculated length of the outer UHPC eccentrically compressed column, the eccentricity of the outer UHPC eccentrically compressed column, and the design value of the axial force of the outer UHPC eccentrically compressed column. Preferably, the load parameters are normalized load parameters.

[0023] The action space consists of an index of 6 design parameters, including the column section diameter of the UHPC eccentrically compressed column, the thickness of the UHPC eccentrically compressed column, the diameter of the reinforcing bars of the UHPC eccentrically compressed column, the number of reinforcing bars of the UHPC eccentrically compressed column, the core concrete strength of the UHPC eccentrically compressed column, and the strength of the UHPC eccentrically compressed column.

[0024] Understandably, the state space refers to the design requirements of the external UHPC eccentric compression column, while the motion space refers to the design scheme of the external UHPC eccentric compression column.

[0025] like Figure 2 and Figure 3 The diagram illustrates the structure of the externally encapsulated UHPC eccentrically compressed column in this embodiment. Figure 2 and Figure 3 The example in the text shows reinforced concrete, core concrete, and external UHPC. Figure 3 The core concrete radius was also specifically marked. Circumferential radius of reinforcing bar distribution Outer center radius Outer packaging thickness and the diameter of the reinforcing bar .

[0026] Step 2: Generate random training samples to simulate the design requirements of different engineering conditions. Input the training samples into the deep reinforcement learning network model for prediction. The deep reinforcement learning network model outputs the combination of predicted design parameters corresponding to the action space. In this embodiment, uniform sampling is used to randomly generate training samples to ensure that the model can cover working conditions such as high and low axial forces and large and small eccentricities, avoiding insufficient generalization ability of the model due to data bias. Specifically, simple random sampling is performed within a preset range for each load parameter to generate training samples. The load parameters in the training samples are normalized results to eliminate the influence of their dimensions. It can be understood that the training samples include the axial force design value, calculated length, and eccentricity of the eccentrically compressed UHPC column.

[0027] In this embodiment, the network structure of the deep reinforcement learning network model (DQN) includes a shared feature extraction layer and multiple independent action output layers.

[0028] The shared feature extraction layer is used to process the input state vector Mapped to a high-dimensional vector The calculation formula is: The shared feature extraction layer comprises three fully connected layers arranged sequentially, each followed by a ReLU activation function layer. The first and second fully connected layers are used for data dimensionality upscaling, corresponding to 3D to 128D and 128D to 256D respectively. The third fully connected layer is used for data dimensionality downscaling, specifically reducing the 256D dimension to 128D. In the above formula, This is the weight matrix of the first fully connected layer. for The corresponding bias vector; This is the weight matrix for the second fully connected layer. for The corresponding bias vector; This is the weight matrix for the third fully connected layer. for The corresponding bias vector; The ReLU activation function is used. (High-dimensional vector) As shared features, these shared features serve as inputs to the output layer of each individual action.

[0029] The multiple independent action output layers correspond to 6 design parameters. Here, there are a total of 6 action branches, i.e. 6 independent action output layers. This is the number of the action branch, which is also the number of the independent action output layer, and subsequently the number of the policy network output branch. Action Branch Used to output the diameter of the column cross section value vector Action branches Used to output the thickness of the outer UHPC. value vector Action branches Used to output the diameter of reinforcing bars value vector Action branches Used to output the number of steel bars value vector Action branches Used to output the core concrete strength value vector Action branches Used for outputting the strength of the outer UHPC. value vector That is, each action branch It is possible to obtain the corresponding design parameter values. value vector The independent action output layer is specifically used to process shared features to obtain the design parameters corresponding to the independent action output layer. value vector , The formula is: in, For the first Each action branch (i.e., an independent action output layer) The weight of ); For the first Bias of each action branch.

[0030] Select each action branch value vector The index corresponding to the maximum value in each action branch determines the action instruction. In other words, the maximum value in each action branch serves as the design parameter value for that branch. Based on the outputs of the six action branches, the combination of design parameters is determined. This approach decomposes the combinatorial optimization problem into multiple independent classification problems. Essentially, each action branch independently selects the optimal value, breaking down the originally complex multi-parameter combinatorial optimization problem into six independent classification problems, significantly reducing the decision-making difficulty.

[0031] In this embodiment, action branch To action branch It is used in a one-to-one correspondence to reduce 128 dimensions to 10, 3, 6, 13, 3, and 3 dimensions.

[0032] In one embodiment, the independent action output layer includes not only multiple action branches but also a maximum value selection module. The number of maximum value selection modules is not limited; for example, the number of maximum value selection modules can be the same as the number of action branches. Each action branch and maximum value selection module is configured in a one-to-one correspondence. The maximum value selection module is used to select the output of its corresponding action branch. The maximum value in the value vector represents the predicted design parameters of the final output of the deep reinforcement learning network model; another example is a maximum value selection module used to select the output of all action branches. The maximum value in the value vector. That is, the independent action output layer is also used to select each... value vector The maximum value is used as the design parameter predicted by the deep reinforcement learning network model, and the maximum value is used to generate the combination of predicted design parameter values. It can be understood that the combination of predicted design parameter values ​​includes the predicted column cross-section diameter, the predicted outer UHPC thickness, the predicted rebar diameter, the predicted number of rebars, the predicted core concrete strength, and the predicted outer UHPC strength.

[0033] In another embodiment, the maximum value selection module is independent of the independent action output layer, that is, it is set after the independent action output layer, and the maximum value selection module is used to select the action branch output. The maximum value in the value vector is the predicted design parameter output by the deep reinforcement learning network model. In addition to the maximum value selection module, the deep reinforcement learning network model also includes a combination generation module, which uses the maximum value to generate a combination of the predicted design parameter values.

[0034] It is understandable that the two embodiments of the maximum value selection module described above are essentially the same.

[0035] In other words, the deep reinforcement learning network model is used to select each value vector The maximum value, which is used as a design parameter for prediction by a deep reinforcement learning network model (e.g., the deep reinforcement learning network model is used to) generate a combination of predicted values ​​for the design parameters using the maximum value.

[0036] Step 3: Based on the output of Step 2, calculate the bearing capacity and cost of the corresponding external UHPC eccentric compression column, and check whether the corresponding external UHPC eccentric compression column meets the engineering specifications, and obtain the first calculation and check results.

[0037] In this embodiment, the calculation steps for bearing capacity (i.e., axial bearing capacity) are as follows: (1) Calculate the eccentricity amplification factor for the ultimate limit state of bearing capacity, taking into account the influence of the slenderness ratio of the member and the initial eccentricity. The calculation formula is as follows: in, This is the eccentricity amplification factor; The calculated length of the component; The effective height of the cross section, The core concrete radius, The radius of the circumference where the reinforcing bars are distributed; For the cross-sectional height, The thickness of the outer UHPC casing; The eccentricity of the axial force about the centroidal axis of the cross section; This is the coefficient representing the influence of load eccentricity on the cross-sectional curvature. This is the coefficient representing the influence of the slenderness ratio of the component on the curvature of the cross section.

[0038] (2) Establish the moment equilibrium equation and axial force equilibrium equation for the eccentrically compressed UHPC column. These equations are nonlinear equilibrium equations. The half-pressure angle in the nonlinear equilibrium equations is solved by the bisection method. The calculation formula is as follows: Solving the equilibrium equations simultaneously, the residual function is calculated as follows: In the formula: in, It is the half-pressure angle; This represents the balance error between the eccentricity of internal forces and the eccentricity of external forces within the cross section. Representing the moment equilibrium equation, Represent the axial force equilibrium equation; This is the eccentricity amplification factor; The eccentricity of the axial force about the centroidal axis of the cross section; The core concrete radius; The radius of the circumference where the reinforcing bars are distributed; For reinforcement ratio, This represents the total cross-sectional area of ​​the reinforcing bars. The core concrete cross-sectional area; For the external reinforcement ratio, The cross-sectional area of ​​the UHPC casing; For the outer center radius, For the outer casing thickness; This refers to the design value of the core concrete compressive strength. This is the design value for the strength of the reinforcing steel. This is the design value for the compressive strength of the outer UHPC. This is the design value for the tensile strength of the outer UHPC. It is the height coefficient of the tension zone reinforcement relative to the limiting compression zone.

[0039] make The solution is obtained by using the cyclic bisection method, and the half-pressure angle that satisfies the convergence condition is returned. .

[0040] (3) Calculation of axial bearing capacity Based on the semi-compression angle, the design values ​​of the core concrete strength, the design values ​​of the reinforcement strength, and the section parameters, the axial bearing capacity of the eccentrically compressed UHPC column, contributed by the core concrete, the outer UHPC, and the reinforcement, is calculated using the following formula: in, The core concrete cross-sectional area; The design value for the core concrete (compressive strength); The half-pressure angle is obtained by the bisection method; This is the height coefficient of the tension zone reinforcement relative to the limiting compression zone. This refers to the reinforcement ratio; This is the design value for the strength of the reinforcing steel. For the reinforcement ratio of the external UHPC; For the compressive strength of the outer UHPC, The tensile strength of the outer UHPC.

[0041] In this embodiment, calculating the cost of the externally enclosed UHPC eccentrically compressed column includes calculating the cost of the core concrete, the cost of the external UHPC, and the cost of the reinforcing steel. In other words, the cost of the externally enclosed UHPC eccentrically compressed column is calculated based on these three costs. In a specific embodiment, the cost of the externally enclosed UHPC eccentrically compressed column is the sum of the core concrete cost, the cost of the external UHPC cost, and the cost of the reinforcing steel.

[0042] In this embodiment, checking whether the external UHPC eccentric compression column meets the engineering specifications includes: checking whether the slenderness ratio meets the engineering specifications, checking whether the reinforcement ratio meets the engineering specifications, and checking whether the rebar spacing meets the engineering specifications. When the slenderness ratio, reinforcement ratio, and rebar spacing all meet the corresponding engineering specifications, the external UHPC eccentric compression column meets the engineering specifications.

[0043] The calculation for the slenderness ratio inspection is as follows: in, The slenderness ratio of the eccentrically compressed UHPC column; The calculated length of the component; The radius of the core concrete.

[0044] The calculation for the reinforcement ratio check is as follows: in, For reinforcement ratio, This represents the total cross-sectional area of ​​the reinforcing bars. The core concrete cross-sectional area.

[0045] The calculation for checking the rebar spacing is as follows: in, This refers to the net spacing between the reinforcing bars; Minimum clear spacing between reinforcing bars; The radius of the circumference where the reinforcing bars are distributed; This represents the total number of steel bars. The diameter of the reinforcing bar.

[0046] If both of the following conditions are met: ① 、② and ③ If the external UHPC eccentrically compressed column meets the engineering specifications, it can also be said to comply with the engineering specifications constraints.

[0047] Step 4: Based on the first calculation and inspection results, calculate the reward signal using the reward function, and update the network parameters of the network model based on the reward signal. Specifically, based on the bearing capacity, cost, and whether the engineering specifications are met as calculated in Step 2, calculate the reward signal for training (model training) using the reward function, construct a loss function using the reward signal for training, perform backpropagation on the network output based on the loss function, update the parameters of the deep reinforcement neural network, and obtain a trained deep reinforcement learning network model after training. The trained deep reinforcement learning network model can be called the outsourced UHPC eccentric compression column design model, including the following steps: (1) Calculate the reward signal Here, the reward signal used for training is calculated; In this embodiment, the formula for calculating the reward function is as follows: in, This is an index of compressive bearing capacity, also known as the bearing capacity ratio. To enhance the axial bearing capacity of the UHPC eccentrically compressed column, This is the design value for axial force; As a cost-performance indicator; Cost of outsourcing UHPC eccentric compression columns; As a reward for carrying capacity; This is a reward for cost-effectiveness.

[0048] When the design meets engineering specifications, a reward signal is given. The calculation is performed using the aforementioned reward function formula. When the design does not meet engineering specifications, let... Severe penalties will be imposed.

[0049] Carrying capacity bonus According to the compressive bearing capacity index The distribution is calculated using the following formula: Value-for-money reward In satisfying and The calculation is as follows: The calculation of the reward signal corresponding to the candidate design parameter combination in step 6 can also be done using or by referring to this method.

[0050] (2) Experience replay The reward samples from the training process are stored in an experience replay pool, where the reward samples include the current state vector. Action command vector Reward signals Next state vector and termination mark If the reward signal of the current sample is greater than the preset high-quality threshold... If so, the sample is stored repeatedly to increase its sampling probability in subsequent training. (3) Loss Calculation Randomly sample batches of data from the experience replay pool and calculate the total loss between the output value of the policy network and the target value of the target network. The calculation formula is as follows: in, Batch size; Number the sample; The number is the output branch number of the policy network. It can be understood that the policy network output branch corresponds to the independent action output layer. For policy network output branches medium sample of value, For policy network output branches medium sample goal value.

[0051] In this embodiment, the batch size is set to 512.

[0052] (4) Objectives Value Calculation Calculation target value The formula is as follows: in, For the sample The reward signal; Discount factor; Number the next action. Indicates all possible next actions Select the maximum value from the list; For target network samples Next state vector and the next action Output target value; For the sample The termination marker; (5) Backpropagation Minimize the total loss using the Adam algorithm. The policy network parameters in the model are updated, and the policy network is synchronized to the target network every first number (e.g., 100) training rounds, which updates the network parameters of the network model.

[0053] In this embodiment, the learning rate =0.0003, the weight decay coefficient is .

[0054] In the training process of this embodiment, the following is adopted: - The Epsilon-Greedy Algorithm selects actions, including: In the initial exploration phase, using probability Randomly select an action from all available design parameters; As the number of training rounds increases, the probability... Decreasing according to the exponential decay law: in, To find the maximum value function, The attenuation coefficient is... To minimize the exploration probability, we set it to 0.05; During the engineering design phase, take Directly select the policy network The combination of design parameters that yields the highest value.

[0055] Step 5: Obtain the load parameters of the UHPC eccentrically compressed column to be designed, and use the trained deep reinforcement learning network model to obtain multiple sets of candidate design parameter combinations. In this embodiment, 300 candidate design parameter combinations are generated.

[0056] Step 6: Based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation result corresponding to the candidate design parameter combination, and the engineering specification check result corresponding to the candidate design parameter combination, select the optimal candidate design parameter combination among all candidate design parameter combinations as the final design scheme, and output the final design scheme.

[0057] The obtained candidate design parameter combinations undergo secondary bearing capacity calculations and are checked for compliance with engineering specifications. Based on the evaluation indicators, the optimal candidate design parameters are selected as the final design scheme output. Furthermore, the scheme is only output to the user if the secondary bearing capacity calculation is compliant and the engineering specification check is passed.

[0058] In this embodiment, the evaluation indicators include reward signals and design quality levels.

[0059] The evaluation of reward signals is usually determined by the magnitude of the reward signal value.

[0060] The design quality level is determined by the compressive bearing capacity index.

[0061] In this embodiment, the design quality level is determined only when the engineering specifications are met. It is understandable that if the engineering specifications are not met, the design quality level is equivalent to a design failure.

[0062] Design quality levels include insufficient load-bearing capacity design, excellent design, acceptable design, and excessive design, with the following classification criteria: The specific method for selecting the optimal combination of candidate design parameters based on reward signals and design quality levels is not limited.

[0063] As an example rather than a limitation, if an excellent design exists, the excellent design with the largest reward signal is selected from the candidate design parameter combinations of the excellent design; if no excellent design exists but an acceptable design exists, the candidate design parameter combination with the largest reward signal is selected from the acceptable design as the optimal candidate design parameter combination; if there is neither an excellent design nor an acceptable design, but an over-design exists, the over-design with the largest reward signal is selected or a design failure message is output.

[0064] In another embodiment, all candidate design parameter combinations that do not meet the engineering specifications are first eliminated. For the remaining combinations, they are first screened based on the compressive bearing capacity index, and then the optimal candidate design parameters are selected based on the reward signal. Specifically, all candidate design parameter combinations that do not meet the engineering specifications are first eliminated. For the remaining combinations, all candidate design parameter combinations that meet the first compressive bearing capacity index range are retained, and the candidate design parameter combination with the largest reward signal among all retained candidate design parameter combinations is selected as the final design scheme. If no candidate design parameter combination has a compressive bearing capacity index that meets the first compressive bearing capacity index range, then all candidate design parameter combinations that meet the second compressive bearing capacity index range are retained, and the candidate design parameter combination with the largest reward signal among all retained candidate design parameter combinations is selected as the final design scheme.

[0065] The intelligent design method for outsourced UHPC eccentrically compressed columns can obtain a design model for outsourced UHPC eccentrically compressed columns. The design model includes a deep reinforcement learning network model, which is used to obtain the load parameters of the outsourced UHPC eccentrically compressed column to be designed and output multiple sets of candidate design parameter combinations. The design model also includes an optimal candidate design parameter combination screening module, which is used to perform bearing capacity calculation and engineering code check on the outsourced UHPC eccentrically compressed column corresponding to each set of candidate design parameter combinations. Based on the reward signal, bearing capacity calculation results and engineering code check results corresponding to the candidate design parameter combinations, the optimal candidate design parameter combination among all candidate design parameter combinations is selected as the final design scheme.

[0066] Furthermore, the outsourced UHPC eccentric compression column intelligent design method, following step 6 above, also includes the following steps: The performance of deep reinforcement learning network models is evaluated using model evaluation metrics to verify their generalization ability. 5000 test samples were randomly generated using a uniform distribution. The test samples were then input into the pre-trained deep reinforcement learning network model. After a second check, the model output the optimal candidate design (i.e. the optimal combination of candidate design parameters) and the evaluation result of the design.

[0067] In this embodiment, the design quality level distribution is shown in Table 1.

[0068] Table 1. Distribution of Design Quality Levels

[0069] like Figure 4 As shown, this is a scatter plot of the cost-bearing capacity ratio of the output of the deep reinforcement learning network model, which shows that the model achieves a balance between safety and economy.

[0070] like Figure 5 As shown, the histogram of the load-bearing ratio distribution of the deep reinforcement learning network model output demonstrates that the design of the deep reinforcement learning network model mainly focuses on the excellent design range. Figure 5 The target value = 1.0 and the upper limit = 1.2 correspond to... .

[0071] In this embodiment, the formula for calculating the success rate index is as follows: in, To measure the success rate of deep reinforcement learning network models; The number of samples for an excellent design; The acceptable sample size; This represents the total number of test samples.

[0072] Key performance evaluation indicators are shown in Table 2: Table 2 Key Performance Evaluation Metrics for Deep Reinforcement Learning Network Models

[0073] See Figure 6 This embodiment provides an intelligent design system for an outsourced UHPC eccentric compression column, comprising: The model building module is used to construct a deep reinforcement learning network model, defining the state space, action space, and reward function of the network model; the load parameters of the state space include the computational length, eccentricity, and axial force design value; the design parameters of the action space include the column cross-section diameter, the thickness of the outer UHPC, the diameter of the reinforcing bars, the number of reinforcing bars, the core concrete strength, and the outer UHPC strength. The acquisition and input module is used to acquire training samples and input the training samples into the deep reinforcement learning network model. The network model is used to output a combination of design parameter prediction values, including a shared feature extraction layer for extracting shared features from load parameters and an independent action output layer whose number is equal to the number of design parameters in the action space. The independent action output layer and the number of design parameters in the action space correspond one-to-one, and are used to process the shared features and output the design parameter prediction value of one design parameter. The calculation and update module is used to calculate the bearing capacity and cost of the eccentrically compressed UHPC column using the combination of predicted design parameter values ​​obtained based on training samples, and to check whether the eccentrically compressed UHPC column meets the engineering specifications to obtain the first calculation and check result; it is used to calculate the reward signal for training based on the first calculation and check result and the reward function, and to update the network parameters of the network model based on the reward signal for training. The acquisition and prediction module is used to acquire the load parameters of the eccentrically compressed UHPC column to be designed, and to obtain multiple sets of candidate design parameter combinations using a trained deep reinforcement learning network model. The calculation module is used to perform bearing capacity calculations and engineering code checks on the external UHPC eccentrically compressed columns corresponding to each set of candidate design parameter combinations. The selection module is used to select the optimal candidate design parameter combination from all candidate design parameter combinations as the final design scheme, based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation result, and the engineering specification check result.

[0074] The deep reinforcement learning network model described in this embodiment includes a shared feature extraction layer for extracting shared features from load parameters and six independent action output layers for processing the shared features and outputting predicted values ​​of design parameters. The six independent action output layers correspond one-to-one with six design parameters (column cross-section diameter, outer UHPC thickness, steel bar diameter, number of steel bars, core concrete strength, and outer UHPC strength). In specific implementation, the intelligent design system for an outsourced UHPC eccentric compression column can be implemented by referring to the intelligent design method for an outsourced UHPC eccentric compression column in any of the above embodiments. The specific implementation steps will not be repeated here.

[0075] An electronic device can be implemented according to the method of this disclosure, the electronic device comprising: a memory; one or more processors; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for executing an outsourced UHPC eccentric compression column intelligent design method according to any of the above embodiments.

[0076] This disclosure also provides a storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the intelligent design method for an outsourced UHPC eccentric compression column.

[0077] The advantages of this disclosed intelligent design method, system, and medium for outsourced UHPC eccentrically compressed columns are as follows: This disclosure realizes the intelligent design of outsourced UHPC eccentrically loaded columns, solving the problems of time-consuming and labor-intensive design, and enabling the rapid and efficient acquisition of a safe and economical optimal design. This disclosure constructs and trains a deep reinforcement learning network model, which includes a shared feature extraction layer and multiple independent action output layers, enabling efficient and intelligent autonomous design of the design parameters for outsourced UHPC eccentrically loaded columns. By calculating the bearing capacity and cost, and checking whether the outsourced UHPC eccentrically loaded columns meet engineering specifications, and then using a reward function to calculate the reward signal used for training to update the network parameters of the network model, the intelligent design takes into account multiple objectives such as bearing capacity, cost, and engineering specifications, obtaining the candidate design parameter combination that best balances safety and economy. Based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation results, and the engineering specification check results, the output of the deep reinforcement learning network model is evaluated to select the optimal design from the candidate design parameter combinations. This disclosure offers good design performance and high design efficiency.

[0078] Specifically, it has the following effects: Improved design efficiency of outsourced UHPC eccentric compression columns: This disclosure provides an intelligent design method for outsourced UHPC eccentric compression columns based on deep reinforcement learning neural networks. Through the trained deep reinforcement learning neural network, the parameters of the outsourced UHPC eccentric compression column are designed autonomously, which greatly improves the design efficiency of this complex structure and reduces the reliance on the designer's personal experience.

[0079] The safety of the design is strictly ensured: This disclosure severely punishes designs that do not comply with engineering specifications or fail to meet load-bearing requirements, forcing the model to automatically avoid non-compliant and unsafe designs. The model output also undergoes secondary verification to ensure the safety and reliability of the design and reduce the workload of review. Furthermore, the design based on a network model with a reward function does not rely on designers or require manual steps such as labeling training samples.

[0080] This invention improves the rationality of the design of outsourced UHPC eccentrically compressed columns: Unlike traditional design methods that rely on empirical formulas, this invention uses a rigorous mechanical solution process, considers the sharing of the tensile strength of ultra-high performance concrete during calculation, and introduces a cost-effectiveness bonus to guide the model to find a more economical design scheme. This effectively solves the problem of the traditional design scheme being too conservative and improves the economy.

[0081] To address the discreteness issue of high-dimensional design parameters and improve model convergence speed, this disclosure constructs a deep reinforcement learning network model comprising multiple independent action output layers. This decomposes the combinatorial optimization problem into independent classification dimensions, reducing the action space dimension of the design parameters. Simultaneously, an experience replay pool is used to store reward samples during training. Batch data is randomly sampled from the experience replay pool, and the probability of high-quality samples being sampled is increased by repeatedly storing them. This improves the convergence speed and stability of the neural network and avoids getting trapped in local optima.

[0082] This disclosure allows for flexible provision of reward functions and does not require a massive number of well-labeled design samples.

[0083] This disclosure offers high security, good economic efficiency, and comprehensive consideration of both security and economy. The design results are highly reliable, highly usable, have a high probability of yielding an excellent design, and have high practical value.

[0084] This disclosure presents a method for designing an externally-enclosed UHPC eccentrically compressed column that can deeply integrate mechanical mechanisms, engineering constraint checks, and cost estimation models, and achieve multi-objective global optimization through self-learning. This method enables the rapid and accurate generation of externally-enclosed UHPC eccentrically compressed column design schemes from upper load parameters, and has significant engineering application value and practical significance.

[0085] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0086] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems and methods according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0087] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0088] The embodiments described above are merely illustrative of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these all fall within the protection scope of this disclosure. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A smart design method for an externally packaged UHPC eccentrically compressed column, characterized in that, include: A deep reinforcement learning network model is constructed, and the state space, action space, and reward function of the network model are defined. The load parameters of the state space include the computational length, eccentricity, and axial force design value. The design parameters of the action space include the column cross-section diameter, the thickness of the outer UHPC, the diameter of the reinforcing bars, the number of reinforcing bars, the core concrete strength, and the outer UHPC strength. Acquire training samples and input them into the deep reinforcement learning network model. The network model is used to output a combination of design parameter prediction values, including a shared feature extraction layer for extracting shared features from load parameters and an independent action output layer whose number is equal to the number of design parameters in the action space. The independent action output layer corresponds one-to-one with the number of design parameters in the action space and is used to process the shared features and output the design parameter prediction value of one design parameter. Using the combination of predicted design parameters obtained from training samples, the bearing capacity and cost of the external UHPC eccentrically compressed column are calculated, and the first calculation and inspection results are obtained by checking whether the external UHPC eccentrically compressed column meets the engineering specifications. Based on the first calculation and inspection results, the reward signal used for training is calculated using the reward function, and the network parameters of the network model are updated based on the reward signal used for training. Obtain the load parameters of the outsourced UHPC eccentrically compressed column to be designed, and use the trained deep reinforcement learning network model to obtain multiple sets of candidate design parameter combinations; For each combination of candidate design parameters, the bearing capacity of the enclosed UHPC eccentrically compressed column is calculated and checked against engineering specifications. Based on the reward signals corresponding to the candidate design parameter combinations, the bearing capacity calculation results, and the engineering specification inspection results, the optimal candidate design parameter combination among all candidate design parameter combinations is selected as the final design scheme.

2. The intelligent design method for an externally packaged UHPC eccentrically compressed column according to claim 1, characterized in that, The specific steps for checking whether the external UHPC eccentric compression column meets the engineering specifications are as follows: check whether the slenderness ratio, reinforcement ratio, and steel bar spacing of the external UHPC eccentric compression column all meet the engineering specifications; The calculation of the bearing capacity and cost of the externally packaged UHPC eccentrically compressed column specifically includes: Calculate the eccentricity amplification factor at the ultimate limit state of bearing capacity; Establish the moment equilibrium equation and axial force equilibrium equation for the eccentrically compressed UHPC column, and use the bisection method to solve for the half-pressure angle; Based on the semi-compression angle, the design value of the core concrete strength, the design value of the steel reinforcement strength, and the section parameters, the axial bearing capacity of the enclosed UHPC eccentrically compressed column is calculated. Calculate the cost of core concrete, external UHPC, and steel reinforcement. Based on the costs of core concrete, external UHPC, and steel reinforcement, calculate the cost of the external UHPC eccentrically compressed column.

3. The intelligent design method for an externally packaged UHPC eccentrically compressed column according to claim 2, characterized in that, The moment balance equation is: The axial force balance equation is: The bisection method for solving the half-pressure angle involves: simultaneously solving the equilibrium equations and calculating the residual function. , Find the solution using the cyclic binary search method. The half pressure angle; in, It is the half-pressure angle; This represents the balance error between the eccentricity of internal forces and the eccentricity of external forces within the cross section. Representing the moment equilibrium equation, Represent the axial force equilibrium equation; This is the eccentricity amplification factor; The eccentricity of the axial force about the centroidal axis of the cross section; The core concrete radius; The radius of the circumference where the reinforcing bars are distributed; For reinforcement ratio, This represents the total cross-sectional area of ​​the reinforcing bars. The core concrete cross-sectional area; For the external reinforcement ratio, The cross-sectional area of ​​the UHPC casing; For the outer center radius, For the outer casing thickness; This refers to the design value of the core concrete compressive strength. This is the design value for the strength of the reinforcing steel. This is the design value for the compressive strength of the outer UHPC. This is the design value for the tensile strength of the outer UHPC. It is the height coefficient of the tension zone reinforcement relative to the limiting compression zone.

4. The intelligent design method for an externally packaged UHPC eccentrically compressed column according to claim 1, characterized in that, The formula for the reward function is: in, As a reward signal; As a reward for carrying capacity; As a reward for cost-effectiveness; For compressive bearing capacity index, To enhance the axial bearing capacity of the UHPC eccentrically compressed column, This is the design value for axial force; As a cost-performance indicator; Cost of outsourcing UHPC eccentric compression columns; When the result of checking whether the outsourced UHPC eccentrically compressed column meets the engineering specifications is that it meets the engineering specifications, a reward signal is given. The calculation is performed using the formula for the reward function; otherwise, let... ; The carrying capacity reward According to the compressive bearing capacity index The distribution calculation formula is as follows: In satisfying and At that time, the aforementioned cost-effectiveness reward The calculation formula is: 。 5. The intelligent design method for an externally packaged UHPC eccentrically compressed column according to claim 4, characterized in that, The specific steps for updating the network parameters of the network model based on the reward signal used for training include: The reward samples during training are stored in an experience replay pool. If the reward signal corresponding to a reward sample is greater than a preset high-quality threshold, the reward sample is stored multiple times in the experience replay pool. The reward sample includes the current state vector. Action command vector Reward signals Next state vector and termination mark ; Randomly sample batches of data from the experience replay pool and calculate the total loss between the network model output value and the target value of the target network model. The calculation formula is: in, Batch size; Number the sample; Output the branch number for the policy network; For policy network output branches medium sample of value; For policy network output branches medium sample goal value; Calculation target value The formula is: in, For the sample The reward signal; Discount factor; Number the next action. Indicates all possible next actions Select the maximum value from the list; For target network samples Next state vector and the next action Output target value; For the sample The termination marker; Minimize the total loss using the Adam algorithm. Update the policy network parameters in the model, and synchronize the policy network to the target network every first number of training rounds.

6. The intelligent design method for an externally packaged UHPC eccentrically compressed column according to claim 4, characterized in that, During the training process of the deep reinforcement learning network model, the following methods are employed: - Greedy strategies select actions, specifically including: In the initial exploration phase, using probability Randomly select an action from all available design parameters; As the number of training rounds increases, the probability... Decreasing according to the exponential decay law, satisfying... in, To find the maximum value function, The attenuation coefficient is... This represents the minimum exploration probability.

7. The intelligent design method for an externally packaged UHPC eccentrically compressed column according to claim 1, characterized in that, The shared feature extraction layer comprises three sequentially arranged fully connected layers, each followed by a ReLU activation function layer; specifically, the shared feature extraction layer is used to process the input state vector... Mapped to a high-dimensional vector The calculation formula is: in, This is the weight matrix of the first fully connected layer. for The corresponding bias vector; This is the weight matrix for the second fully connected layer. for The corresponding bias vector; This is the weight matrix for the third fully connected layer. for The corresponding bias vector; The high-dimensional vector is the ReLU activation function. As the shared feature; The independent action output layer is specifically used to process shared features to obtain the design parameters corresponding to the independent action output layer. value vector , The formula is: in, The number of the independent action output layer; For independent action output layer The weights; For independent action output layer The bias; The deep reinforcement learning network model is used to select each value vector The maximum value is used to generate the combination of predicted design parameter values.

8. The intelligent design method for an externally packaged UHPC eccentrically compressed column according to claim 1, characterized in that, The process of selecting the optimal candidate design parameter combination as the final design scheme based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation results, and the engineering specification check results includes: Step 6.1: Filter out all candidate design parameter combinations that do not meet the engineering specifications; Step 6.2: For the retained candidate design parameter combinations, if there is a candidate design parameter combination whose compressive bearing capacity index meets the first compressive bearing capacity index range, then all candidate design parameter combinations that meet the first compressive bearing capacity index range are retained, and the candidate design parameter combination with the largest reward signal among all retained candidate design parameter combinations is selected as the final design scheme; if there is no candidate design parameter combination whose compressive bearing capacity index meets the first compressive bearing capacity index range, then all candidate design parameter combinations that meet the second compressive bearing capacity index range are retained, and the candidate design parameter combination with the largest reward signal among all retained candidate design parameter combinations is selected as the final design scheme.

9. A smart design system for an outsourced UHPC eccentric compression column, characterized in that, include: The model building module is used to construct a deep reinforcement learning network model, defining the state space, action space, and reward function of the network model; the load parameters of the state space include the computational length, eccentricity, and axial force design value; the design parameters of the action space include the column cross-section diameter, the thickness of the outer UHPC, the diameter of the reinforcing bars, the number of reinforcing bars, the core concrete strength, and the outer UHPC strength. The acquisition and input module is used to acquire training samples and input the training samples into the deep reinforcement learning network model. The network model is used to output a combination of design parameter prediction values, including a shared feature extraction layer for extracting shared features from load parameters and an independent action output layer whose number is equal to the number of design parameters in the action space. The independent action output layer and the number of design parameters in the action space correspond one-to-one, and are used to process the shared features and output the design parameter prediction value of one design parameter. The calculation and update module is used to calculate the bearing capacity and cost of the outer UHPC eccentric compression column by using the combination of design parameter prediction values ​​obtained based on training samples, and to check whether the outer UHPC eccentric compression column meets the engineering specifications to obtain the first calculation and check results. Used to calculate a reward signal for training based on the first calculation and inspection results and the reward function, and to update the network parameters of the network model based on the reward signal for training; The acquisition and prediction module is used to acquire the load parameters of the eccentrically compressed UHPC column to be designed, and to obtain multiple sets of candidate design parameter combinations using a trained deep reinforcement learning network model. The calculation module is used to perform bearing capacity calculations and engineering code checks on the external UHPC eccentrically compressed columns corresponding to each set of candidate design parameter combinations. The selection module is used to select the optimal candidate design parameter combination from all candidate design parameter combinations as the final design scheme, based on the reward signal corresponding to the candidate design parameter combination, the bearing capacity calculation result, and the engineering specification check result.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent design method for an outsourced UHPC eccentrically compressed column as described in any one of claims 1-8.

Images