A method and system for optimizing assembly stresses of a steel web bridge member

Abstract

The application relates to a steel web bridge component assembly stress optimization method and system. The method comprises the following steps: acquiring bridge design parameters and component measured point cloud data, and modeling manufacturing error probability distribution; training a physical information proxy model based on high-fidelity finite element simulation, and training a time-space attention stress prediction model in combination with a dynamic graph sequence; constructing a reinforcement learning environment and a reward function, and training an intelligent agent through a proximal policy optimization algorithm; and fusing on-site strain monitoring and online calibration to drive the intelligent agent to rollingly optimize and output real-time process instructions. Through the fusion of manufacturing error probability modeling, a physical information proxy model, a time-space attention stress prediction and other means, the method effectively improves the control precision of steel web bridge component assembly stress, the global optimization effect of assembly process and construction reliability, and reduces the assembly residual stress concentration risk.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method and system for optimizing the assembly stress of steel web bridge components. Background Technology

[0002] Long-span steel web bridges, due to their advantages such as light weight, high bending stiffness, and convenient construction, have become one of the core bridge types in modern transportation infrastructure construction. Their prefabricated assembly construction mode is an important development direction in the current bridge engineering field. The stress distribution during assembly of steel web components directly determines the service safety and durability of the bridge structure. Residual stress concentration or uneven distribution generated during assembly can easily lead to weld cracking, component deformation, and other defects, which can seriously affect the overall load-bearing capacity of the bridge. Therefore, the control and dynamic optimization of assembly stress have become the core technologies for the construction of steel web bridges.

[0003] In the assembly and construction of steel web bridge components, manufacturing errors are a key factor affecting the stress distribution during assembly. Currently, static deviation analysis is commonly used to handle manufacturing errors, obtaining a single error value through a simple comparison between measured component data and the design model. However, this approach lacks a systematic statistical analysis of the probability distribution characteristics of manufacturing errors, making it difficult to reflect the error distribution patterns of batch components. Furthermore, high-fidelity finite element simulations based on this type of error data show significant deviations from actual assembly conditions, failing to provide reliable data support for subsequent stress prediction and process optimization. Simultaneously, traditional high-fidelity finite element simulation methods are computationally complex and time-consuming, making it difficult to meet the needs of real-time mechanical response analysis during assembly. Although some studies have attempted to use neural networks to construct simulation proxy models, most lack embedding fundamental equations of solid mechanics and physical constraints, resulting in insufficient generalization ability and prediction accuracy of the models, failing to accurately characterize the mechanical behavior of the steel web assembly process.

[0004] Furthermore, the assembly process of steel web bridges exhibits significant temporal and spatial coupling. The execution sequence of different assembly procedures, the positioning accuracy of components, and the tightness of connections all have a substantial impact on the evolution of assembly stress. However, current characterization methods for the assembly process primarily employ static models or discretized step-by-step analyses, failing to construct characterization models that reflect the dynamic changes in the construction sequence and structural spatial relationships. This makes it difficult to fully capture the spatiotemporal evolution of assembly stress, and stress prediction models trained based on such characterization results cannot accurately correlate the intrinsic relationship between the construction process and stress distribution, thus limiting the practicality of the prediction results. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and system for optimizing the assembly stress of steel web bridge components to address the above-mentioned technical problems, aiming to improve the accuracy and real-time performance of stress prediction for steel web bridge assembly.

[0006] Firstly, this application provides a method for optimizing the assembly stress of steel web bridge components, including:

[0007] The bridge design parameters and measured point cloud data of components are acquired. Based on the bridge design parameters and measured point cloud data of components, statistical analysis of manufacturing errors is performed to obtain a probability distribution model of manufacturing errors. Based on the probability distribution model of manufacturing errors, high-fidelity finite element simulation is performed to obtain simulation results. Based on the simulation results, a pre-set physical information neural network is trained to obtain a physical information proxy model.

[0008] A predefined assembly process sequence is obtained, and the assembly process is dynamically represented based on the bridge design parameters and the predefined assembly process sequence to obtain a dynamic graph sequence. A pre-defined temporal-spatial attention neural network is trained based on the physical information proxy model and the dynamic graph sequence to obtain a dynamic prediction model for assembly stress.

[0009] A reinforcement learning simulation environment is constructed based on a physical information agent model, and a reward function is constructed based on an assembly stress dynamic prediction model. Through the reinforcement learning simulation environment, a proximal policy optimization algorithm is adopted, and a policy is trained on a preset agent based on the reward function to obtain a deep reinforcement learning agent.

[0010] The system acquires actual component manufacturing errors, performs offline scheme calculations based on these errors and a deep reinforcement learning agent, generates and executes an initial assembly process scheme. During construction according to the initial assembly process scheme, it acquires on-site strain monitoring data and the current construction status. Based on the on-site strain monitoring data, it performs online calibration of the assembly stress dynamic prediction model to obtain an updated model. Based on the updated model and the current construction status, it drives the deep reinforcement learning agent to perform rolling optimization calculations and outputs real-time optimized assembly process instructions.

[0011] In one embodiment, manufacturing error statistical analysis is performed based on bridge design parameters and measured point cloud data of components to obtain a probability distribution model of manufacturing errors; high-fidelity finite element simulation is performed based on the probability distribution model of manufacturing errors to obtain simulation results; a pre-set physical information neural network is trained based on the simulation results to obtain a physical information proxy model, including:

[0012] The measured point cloud data of the component is registered with the theoretical model corresponding to the bridge design parameters to obtain the registered point cloud data; the registered point cloud data is then denoised to obtain the denoised point cloud data; the deviation between the denoised point cloud data and the theoretical model is calculated to obtain the component manufacturing error data.

[0013] The component manufacturing error data is fitted with a normal distribution to obtain an initial distribution model; the initial distribution model is then subjected to a significance test to obtain the test result. When the test result is qualified, the corresponding initial distribution model is used as the probability distribution model of the manufacturing error.

[0014] Based on the probability distribution model of manufacturing errors, random sampling is performed according to the preset sample size and error distribution range to generate multiple digital samples of components with random manufacturing errors.

[0015] Based on bridge design parameters and digital samples of components with random manufacturing errors, boundary constraints and load conditions corresponding to the assembly process are set; high-fidelity finite element simulation is performed according to the boundary constraints and load conditions to generate mechanical response simulation data of the assembly process.

[0016] The mechanical response simulation data of the assembly process is divided into a training set and a validation set according to a preset ratio. A preset solid mechanics fundamental equation is embedded into a preset physical information neural network to obtain a physical information neural network with physical constraints. The physical information neural network with physical constraints is trained based on the training set to obtain the trained neural network. The trained neural network is then validated using a validation set to obtain the predicted mechanical response data. The deviation between the actual mechanical response data and the predicted mechanical response data corresponding to the validation set is calculated to obtain the prediction error. When the prediction error is less than a preset error threshold, training is stopped, and the corresponding trained neural network is used as a physical information proxy model.

[0017] In one embodiment, the assembly process is dynamically represented based on bridge design parameters and a predefined assembly process sequence to obtain a dynamic graph sequence; a pre-defined temporal-spatial attention neural network is trained based on a physical information proxy model and the dynamic graph sequence to obtain a dynamic prediction model for assembly stress, including:

[0018] Based on the bridge design parameters, the geometric features and connection relationships of the assembled components are determined. Combined with the predefined assembly process sequence, the assembly structure of each step is defined as a graph structure of component nodes and connection relationship edges. The graph structure of component nodes and connection relationship edges is arranged according to the construction sequence to obtain a dynamic graph sequence.

[0019] Based on the physical information proxy model, the mechanical response of the assembly action corresponding to each step in the dynamic graph sequence is quickly calculated to obtain the component stress data of each step, and the component stress data of each step is used as the component mechanical state data.

[0020] The mechanical state data of the component is mapped to the node features of the graph structure of the component node-connection relationship edge. Based on the node features, the dynamic graph sequence is enhanced to obtain the enhanced dynamic graph sequence.

[0021] Based on a predefined assembly process sequence, bridge design parameters, and digital samples of components with random manufacturing errors, the corresponding assembly scheme is determined; the mechanical response of the assembly scheme is simulated based on a physical information proxy model to obtain the residual stress field distribution data after assembly.

[0022] A temporal-spatial attention neural network, including a temporal attention layer and a spatial attention layer, is constructed. The dynamic graph sequence after feature enhancement is used as input, and the residual stress field distribution data after assembly is used as the training target. The network parameters of the temporal-spatial attention neural network are optimized through backpropagation to obtain the optimized neural network. The optimized neural network is used as the dynamic prediction model for assembly stress.

[0023] In one embodiment, a reinforcement learning simulation environment is constructed based on a physical information agent model, and a reward function is constructed based on an assembly stress dynamic prediction model. Through the reinforcement learning simulation environment, a near-end policy optimization algorithm is used to train a pre-defined agent based on the reward function, resulting in a deep reinforcement learning agent, including:

[0024] The assembly state vector is obtained by vector encoding based on the node and edge features of the dynamic graph sequence; the action dimension is divided based on the predefined assembly process sequence, and the discrete-continuous hybrid action space is defined; the action dimension includes assembly sequence, fastening torque and positioning reference.

[0025] The assembly state vector and the action commands in the action space are input into the assembly stress dynamic prediction model, and the predicted stress field distribution data and the maximum residual stress are output.

[0026] The stress uniformity index is calculated based on the predicted stress field distribution data; a reward function is constructed by combining the maximum residual stress, the stress uniformity index, and the preset process execution efficiency threshold.

[0027] A reinforcement learning simulation environment is constructed based on the physical information agent model. The reinforcement learning simulation environment is used to receive action instructions in the action space and output the corresponding assembly state vector and instant reward value based on the reward function.

[0028] Initialize the policy network and value network of the preset agent to obtain the initial agent; use the proximal policy optimization algorithm to collect experience data through the interaction between the initial agent and the reinforcement learning simulation environment, and iteratively optimize the policy network parameters and value network parameters of the initial agent based on the experience data to obtain the deep reinforcement learning agent.

[0029] In one embodiment, a proximal policy optimization algorithm is employed. Through the interaction between a pre-set agent and a reinforcement learning simulation environment, experiential data is collected. Based on this experiential data, the policy network parameters and value network parameters of the initial agent are iteratively optimized to obtain a deep reinforcement learning agent, including:

[0030] S1. Through the initial agent, the agent performs actions in the action space in the reinforcement learning simulation environment according to the current policy network parameters. The agent collects the assembly state vector, action instruction, immediate reward value and the next assembly state vector corresponding to each step, and forms an experience data sequence. The experience data sequence is stored in the preset experience buffer.

[0031] S2. Randomly sample experience data from the preset experience buffer in preset batches, extract the assembly state vector and corresponding action instructions for each step from the experience data, and form a state-action pair; based on the generalized advantage estimation algorithm, combine the state value output by the value network and the immediate reward value, calculate the advantage function value corresponding to each state-action pair;

[0032] S3. Based on empirical data and advantage function values, construct an agent objective function based on a pruning mechanism, set a preset pruning coefficient, maximize the agent objective function through a stochastic gradient ascent algorithm according to the preset pruning coefficient, update the policy network parameters of the initial agent, and obtain the updated policy network parameters.

[0033] S4. Calculate the cumulative return value based on empirical data, use the mean squared error loss function, minimize the deviation between the predicted value output by the value network and the cumulative return value through the gradient descent algorithm, update the value network parameters of the initial agent, and obtain the updated value network parameters.

[0034] S5. Repeat steps S1 to S4. After each preset number of iterations, calculate the average reward value of the initial agent based on the instantaneous reward value obtained by the initial agent in the reinforcement learning simulation environment in each iteration. When the fluctuation range of the average reward value in consecutive preset number of iterations is less than the preset range threshold, the policy is determined to converge. Combine the corresponding updated policy network parameters and updated value network parameters to obtain the deep reinforcement learning agent.

[0035] In one embodiment, the current construction status includes the progress of the work process, the positioning accuracy of the components, and the connection and fastening status;

[0036] The assembly stress dynamic prediction model is calibrated online based on on-site strain monitoring data to obtain an updated assembly stress dynamic prediction model. Based on the updated assembly stress dynamic prediction model and the current construction state, a rolling optimization calculation is performed on the deep reinforcement learning agent to output real-time optimized assembly process instructions, including:

[0037] Obtain the process parameters of the assembly process that has been executed, input the current construction status and the process parameters of the assembly process that has been executed into the assembly stress dynamic prediction model, and output the predicted strain data.

[0038] The error between the on-site strain monitoring data and the predicted strain data is calculated. Based on the error, the network weights of the assembly stress dynamic prediction model are fine-tuned using an adaptive learning rate to obtain the updated assembly stress dynamic prediction model.

[0039] Input the current construction status into the updated assembly stress dynamic prediction model to obtain the stress field prediction results output by the updated assembly stress dynamic prediction model.

[0040] The current construction state is encoded as a state vector. Based on the state vector and the stress field prediction results, the deep reinforcement learning agent is driven to perform rolling optimization calculations on the unassembled processes, generating real-time optimized assembly process instructions including assembly sequence adjustment instructions, fastening torque correction parameters, and positioning benchmark fine-tuning schemes, and outputting the real-time optimized assembly process instructions.

[0041] In one embodiment, the mathematical expression for the proxy objective function is:

[0042]

[0043]

[0044] in, For the proxy objective function, These are the learnable parameters of the policy network; For time step The mathematical expectation is taken from the reinforcement learning experience data; The ratio of the probability of actions under the new and old strategies; For time step The generalized advantage estimate is used to characterize the value advantage of the current assembly action compared to the average action; For the clipping function, Preset clipping factor; For parameters The strategy network, in the assembly state Down Output Action The probability of; For the old policy network before the update, in the assembly state Down Output Action The probability of.

[0045] Secondly, this application also provides a stress optimization system for the assembly of steel web bridge components, comprising:

[0046] The digital twin construction module is used to acquire bridge design parameters and measured point cloud data of components. Based on the bridge design parameters and measured point cloud data of components, statistical analysis of manufacturing errors is performed to obtain a probability distribution model of manufacturing errors. Based on the probability distribution model of manufacturing errors, high-fidelity finite element simulation is performed to obtain simulation results. Based on the simulation results, a pre-set physical information neural network is trained to obtain a physical information proxy model.

[0047] The dynamic stress field prediction module is used to obtain a predefined assembly process sequence, and to perform dynamic graph representation of the assembly process based on the bridge design parameters and the predefined assembly process sequence to obtain a dynamic graph sequence; based on the physical information proxy model and the dynamic graph sequence, a preset temporal-spatial attention neural network is trained to obtain the assembly stress dynamic prediction model.

[0048] The deep reinforcement learning decision module is used to construct a reinforcement learning simulation environment based on the physical information agent model, construct a reward function based on the assembly stress dynamic prediction model, and train a pre-set agent through the reinforcement learning simulation environment using the near-end policy optimization algorithm and the reward function to obtain a deep reinforcement learning agent.

[0049] The closed-loop control execution module is used to acquire the actual component manufacturing error, perform offline scheme calculations based on the actual component manufacturing error and the deep reinforcement learning agent, generate and execute the initial assembly process scheme; during construction according to the initial assembly process scheme, acquire on-site strain monitoring data and the current construction status; perform online calibration of the assembly stress dynamic prediction model based on the on-site strain monitoring data to obtain an updated assembly stress dynamic prediction model; drive the deep reinforcement learning agent to perform rolling optimization calculations based on the updated assembly stress dynamic prediction model and the current construction status, and output real-time optimized assembly process instructions.

[0050] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0051] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0052] The aforementioned method and system for optimizing the assembly stress of steel web bridge components first acquires bridge design parameters and measured point cloud data of the components, constructs a manufacturing error probability distribution model, and trains a physical information agent model, laying the foundation for subsequent stress prediction and optimization. Secondly, by dynamically representing the assembly process and training a temporal-spatial attention neural network, the spatiotemporal evolution of assembly stress can be captured, overcoming the shortcomings of traditional static representation. Furthermore, a reinforcement learning environment and reward function are constructed based on the physical information agent model, and the agent is trained using a proximal policy optimization algorithm, improving the efficiency and specificity of assembly process optimization. Finally, an initial scheme is generated based on actual manufacturing errors, and the model is calibrated online using on-site data and continuously optimized, achieving dynamic adaptation of the assembly process and ensuring the quality of bridge assembly and structural safety. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart of a method for optimizing the assembly stress of steel web bridge components is provided as an exemplary embodiment of the present invention;

[0055] Figure 2 A flowchart of a method for obtaining a deep reinforcement learning agent is provided as an exemplary embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of a stress optimization system for steel web bridge components, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] In one embodiment, such as Figure 1 As shown, a method for optimizing the assembly stress of steel web bridge components is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0059] S101: Obtain bridge design parameters and measured point cloud data of components; perform statistical analysis of manufacturing errors based on bridge design parameters and measured point cloud data of components to obtain a probability distribution model of manufacturing errors; perform high-fidelity finite element simulation based on the probability distribution model of manufacturing errors to obtain simulation results; train a preset physical information neural network based on the simulation results to obtain a physical information proxy model.

[0060] Specifically, during the prefabrication of steel web bridge components, manufacturing errors exist due to factors such as fluctuations in material properties, limitations in the precision of processing equipment, and deviations in process operations. These errors exhibit randomness and dispersion. Subsequent analysis using only the measured error values ​​of a single component fails to reflect the error distribution patterns of a batch of components, leading to significant discrepancies between the mechanical analysis results based on this data and the actual assembly conditions. Therefore, it is necessary to first obtain bridge design parameters and measured point cloud data of the components. Bridge design parameters serve as the benchmark for assembly and construction, such as the geometric dimensions of the steel web components, material mechanical properties, component connection methods, and design load conditions. The measured point cloud data accurately reflects the actual geometric state of the components. Statistical analysis of manufacturing errors based on these two types of data can extract information on the deviation between the actual and design states of the components and establish a probability distribution model, quantifying the random characteristics of manufacturing errors and providing input conditions that closely approximate actual working conditions for subsequent simulation analysis. Furthermore, after obtaining the probability distribution model of manufacturing errors, high-fidelity finite element simulation can be used to accurately calculate the mechanical responses of components such as stress, strain, and displacement during assembly, based on the component's geometric model, material properties, and boundary conditions.

[0061] Furthermore, using the simulation results as training data to train a pre-defined physical information neural network yields a physical information proxy model. Traditional data-driven neural networks rely solely on sample data for training, lacking adherence to fundamental laws of solid mechanics and exhibiting poor generalization ability under unseen conditions. In contrast, the physical information neural network incorporates fundamental equations of solid mechanics (such as equilibrium, constitutive, and geometric equations) as constraints into the network training process, ensuring that the model output conforms to both data patterns and physical principles. Therefore, through this training process, the physical information proxy model can replace computationally complex and time-consuming high-fidelity finite element simulations, achieving rapid calculation of mechanical responses while maintaining prediction accuracy. This provides computational support for the dynamic characterization of subsequent assembly processes and the construction of reinforcement learning environments.

[0062] S102: Obtain a predefined assembly process sequence, perform dynamic graph representation of the assembly process based on bridge design parameters and the predefined assembly process sequence to obtain a dynamic graph sequence; train a preset temporal-spatial attention neural network based on the physical information proxy model and the dynamic graph sequence to obtain a dynamic prediction model of assembly stress.

[0063] Specifically, the predefined assembly process sequence is an initial process plan based on engineering experience and design requirements, including the assembly sequence of components, the execution priority of each process, and the construction sequence of key connection nodes. Through dynamic graph representation, the assembly structure corresponding to each construction step in the assembly process can be transformed into a graph structure of component nodes and connection relationship edges, quantitatively representing the structural composition and interaction relationships of the assembly at each step, generating a dynamic graph sequence. This sequence fully records the temporal changes and spatial correlation characteristics of the structural state during the assembly process.

[0064] After obtaining the dynamic graph sequence, a physical information proxy model can be used to quickly calculate the mechanical response of each assembly action corresponding to each step in the sequence. This outputs core mechanical state data such as stress and strain of each component after each assembly action, accurately reflecting the impact of the construction step on the mechanical state of the assembly. Using this mechanical state data as node features of the graph structure, feature enhancement is applied to the dynamic graph sequence. This allows the sequence to not only contain the spatiotemporal correlation information of the structure but also incorporate mechanical response features, providing richer and more targeted input data for training the subsequent stress prediction model. Finally, training based on a pre-defined temporal-spatial attention neural network yields a dynamic prediction model for assembly stress. This network possesses dual feature extraction capabilities: the temporal attention layer captures the temporal dependencies between different construction steps in the dynamic graph sequence, while the spatial attention layer captures the spatial coupling relationships between component nodes within the same construction step.

[0065] S103: Construct a reinforcement learning simulation environment based on a physical information agent model, construct a reward function based on an assembly stress dynamic prediction model, and train a deep reinforcement learning agent by using a near-end policy optimization algorithm and the reward function on a preset agent through the reinforcement learning simulation environment.

[0066] Specifically, since the physical information agent model can quickly and accurately output the mechanical response and structural state corresponding to assembly actions, it can be encapsulated into a reinforcement learning simulation environment. This environment can receive assembly action instructions output by the agent, quickly calculate and feedback the next structural state and mechanical response results after the action is executed, thus providing an efficient and reliable simulation scenario for the agent's interactive learning. A reward function is constructed based on the assembly stress dynamic prediction model. This function comprehensively considers stress concentration coefficient, stress distribution uniformity index, construction efficiency, and process feasibility constraints. Negative incentives are given when the predicted stress is close to the material's allowable stress or its distribution is severely uneven; positive incentives are given when the stress level is reasonable and the distribution is uniform, thereby guiding the agent to learn stress-optimized assembly strategies. Furthermore, the proximal policy optimization (PPO) algorithm is used for agent training. This algorithm can constrain the magnitude of policy updates by introducing a pruning mechanism, avoiding training instability caused by policy mutations during training. It also has the advantages of high sample utilization and fast convergence speed, making it suitable for complex decision-making scenarios in assembly process optimization. For example, during training, the agent can perform actions in a reinforcement learning simulation environment based on its current policy. The environment provides corresponding states and rewards. By collecting this interactive experience data, the agent iteratively optimizes its policy network and value network. The policy network outputs the optimal assembly action, while the value network evaluates the value of the current state. Through continuous interaction and iteration, when the agent's policy converges (e.g., the reward value stabilizes within a preset range, and the output assembly process policy can stably achieve the stress optimization objective), a well-trained deep reinforcement learning agent can be obtained.

[0067] S104: Obtain the actual component manufacturing error, perform offline scheme calculation based on the actual component manufacturing error and the deep reinforcement learning agent, generate and execute the initial assembly process scheme; during the construction process according to the initial assembly process scheme, obtain on-site strain monitoring data and the current construction status; perform online calibration of the assembly stress dynamic prediction model based on the on-site strain monitoring data to obtain the updated assembly stress dynamic prediction model; drive the deep reinforcement learning agent to perform rolling optimization calculation based on the updated assembly stress dynamic prediction model and the current construction status, and output real-time optimized assembly process instructions.

[0068] Specifically, in the offline initial scheme generation stage, the actual manufacturing error data of the components to be assembled can accurately reflect the actual state of the current batch of components. This actual manufacturing error data is input into a trained deep reinforcement learning agent. Based on its learned optimal strategy, the agent can quickly calculate and generate an initial assembly process scheme. This scheme is an optimal pre-plan tailored to the current actual component errors, providing a clear initial execution basis for construction, reducing the frequency of process adjustments in the early stages of construction, and improving construction efficiency. In the on-site dynamic optimization stage, during construction according to the initial assembly process scheme, strain sensors deployed in key stress areas of the components, such as fiber optic strain sensors and resistance strain gauges, can collect on-site strain monitoring data in real time. The current construction status (such as process execution progress, component positioning accuracy, and connection tightness) can be recorded through the construction monitoring system. This monitoring data is compared with the prediction results of the assembly stress dynamic prediction model. If there is a deviation, it indicates a difference between the model prediction and the actual working conditions. At this point, the parameters of the assembly stress dynamic prediction model can be calibrated online based on this deviation. The updated model can more accurately match the current on-site working conditions, providing a more reliable stress prediction basis for subsequent optimization. Based on the updated dynamic prediction model of assembly stress and the current construction status, the deep reinforcement learning agent can be re-driven to perform rolling optimization calculations on the remaining unexecuted assembly processes, generate real-time optimized assembly process instructions and output them to the construction control system to guide real-time adjustments in on-site construction.

[0069] The aforementioned method utilizes bridge design parameters and measured point cloud data of components to perform statistical analysis of manufacturing errors, constructing a probability distribution model that addresses the shortcomings of traditional methods, such as one-sided error representation and lack of statistical regularity. Secondly, simulation based on the probability distribution model and training of a physical information neural network overcomes the low computational efficiency of traditional simulations and the lack of physical constraints in neural networks. Furthermore, dynamic graphical representation of the assembly process and training of a temporal-spatial attention neural network effectively solve the problem of traditional static models failing to capture the spatiotemporal evolution of stress. Moreover, by constructing a reinforcement learning simulation environment and reward function, a deep reinforcement learning agent is trained, overcoming the limitations of traditional process optimization that relies on experience-based trial and error and lacks intelligent decision-making. Finally, an initial assembly process is generated through offline calculations, and the prediction model is calibrated and real-time process instructions are optimized using on-site monitoring data, achieving dynamic closed-loop control of assembly stress and improving the accuracy and real-time performance of stress management in steel web bridges.

[0070] In one embodiment, a statistical analysis of manufacturing errors is performed based on bridge design parameters and measured point cloud data of components to obtain a probability distribution model of manufacturing errors; high-fidelity finite element simulation is performed based on the probability distribution model of manufacturing errors to obtain simulation results; and a pre-defined physical information neural network is trained based on the simulation results to obtain a physical information proxy model, including:

[0071] The measured point cloud data of the component is registered with the theoretical model corresponding to the bridge design parameters to obtain the registered point cloud data; the registered point cloud data is then denoised to obtain the denoised point cloud data; the deviation between the denoised point cloud data and the theoretical model is calculated to obtain the component manufacturing error data.

[0072] The component manufacturing error data is fitted with a normal distribution to obtain an initial distribution model; the initial distribution model is then subjected to a significance test to obtain the test result. When the test result is qualified, the corresponding initial distribution model is used as the probability distribution model of the manufacturing error.

[0073] Based on the probability distribution model of manufacturing errors, random sampling is performed according to the preset sample size and error distribution range to generate multiple digital samples of components with random manufacturing errors; based on the bridge design parameters and the digital samples of components with random manufacturing errors, the boundary constraints and load conditions corresponding to the assembly process are set; high-fidelity finite element simulation is performed according to the boundary constraints and load conditions to generate mechanical response simulation data of the assembly process.

[0074] The mechanical response simulation data of the assembly process is divided into a training set and a validation set according to a preset ratio. A preset solid mechanics fundamental equation is embedded into a preset physical information neural network to obtain a physical information neural network with physical constraints. The physical information neural network with physical constraints is trained based on the training set to obtain the trained neural network. The trained neural network is then validated using a validation set to obtain the predicted mechanical response data. The deviation between the actual mechanical response data and the predicted mechanical response data corresponding to the validation set is calculated to obtain the prediction error. When the prediction error is less than a preset error threshold, training is stopped, and the corresponding trained neural network is used as a physical information proxy model.

[0075] Specifically, the iterative nearest-point algorithm can be used for registration. This algorithm iteratively finds the optimal correspondence between the measured point cloud and the surface points of the theoretical model, minimizing the cumulative distance error between corresponding points until the iteration converges to a preset accuracy. After registration, the measured point cloud data usually contains noise points caused by factors such as measurement equipment system errors, ambient light interference, and reflections from impurities on the component surface. This noise can seriously affect the accuracy of error calculation. Therefore, a combination of statistical filtering and radius filtering can be used for denoising. For example, statistical filtering is first used to calculate the average distance and standard deviation within the neighborhood of each point, and outliers with distances greater than the average distance plus three times the standard deviation are removed. Then, radius filtering is used to set the neighborhood radius and the minimum number of neighborhood points to remove isolated noise points with fewer than a threshold number of points in the neighborhood. Finally, the denoised point cloud data can be obtained. In addition, the point-by-point Euclidean distance calculation method can be used to calculate the straight-line distance between each sampling point in the denoised point cloud and the nearest point on the surface of the theoretical model. The magnitude and direction of this distance directly characterize the manufacturing error at that location. By summarizing the error data of all sampling points, component manufacturing error data can be constructed, which can reflect the deviation distribution between the actual geometric state and the design state of the component.

[0076] Furthermore, based on the statistical characteristics of manufacturing errors, influenced by a combination of independent random factors such as material performance fluctuations, processing equipment precision fluctuations, and process operation deviations, the manufacturing errors of mass-produced steel web bridge components typically follow a normal distribution. Therefore, the maximum likelihood estimation method can be used to fit a normal distribution. By calculating the sample mean and sample variance of the component manufacturing error data, the two core parameters of the normal distribution, namely the mean μ and the standard deviation σ, are determined, and an initial distribution model is constructed. To verify the reliability of the initial distribution model, the KS (Kolmogorov-Smirnov) test method can be used to test its significance. This method uses the maximum distance between the empirical distribution function of the error data and the theoretical distribution function of the initial distribution model as the test statistic. This statistic is compared with the critical value at a preset significance level to obtain the test result. When the test statistic is less than the critical value or the corresponding p-value is greater than the preset significance level (usually 0.05), the test result can be considered qualified, indicating that the initial distribution model can accurately characterize the actual distribution law of manufacturing errors. Therefore, this initial distribution model can be determined as the probability distribution model of manufacturing errors.

[0077] Based on the probability distribution model of manufacturing errors, Monte Carlo sampling can be used for random sampling. This method can generate statistically representative random samples based on the probability distribution model. The preset sample size can be set by comprehensively considering statistical significance and the efficiency of subsequent simulation calculations, while the error distribution range can be determined in conjunction with the error control boundaries of the actual manufacturing process to ensure that the sampled error values ​​are within a reasonable engineering reality range. Each random sampling process independently generates a set of error parameters, corresponding to random deviations in the component's geometric features, such as dimensional deviations, form and position deviations, and surface roughness deviations. This set of error parameters is embedded into the component's theoretical geometric model, forming a digital sample of the component containing independent random manufacturing errors. All samples can be aggregated to form a sample set for subsequent simulations.

[0078] Specifically, based on bridge design parameters and digital samples of components with random manufacturing errors, boundary constraints can be designed according to the actual assembly and construction conditions. For example, corresponding constraints can be applied in the finite element model based on the component connection methods and support forms specified in the bridge design parameters. Illustratively, fixed constraints can be applied to the fixed support parts of the components to restrict their translation and rotation in three dimensions, and hinge constraints can be applied to the hinged connections of the components to allow rotation about a specific axis, etc. Load conditions can be designed based on the material mechanical properties, design load conditions, and assembly process requirements in the bridge design parameters. These can include the component's own gravity load, preload applied during assembly (such as bolt tightening force), contact pressure between components, and temporary loads during construction, etc., and the magnitude, direction, and application location of each load are consistent with the actual assembly conditions. Subsequently, based on the aforementioned boundary constraints and load conditions, a high-fidelity finite element simulation method can be used to calculate the mechanical response of the assembly process. During the simulation, finite element types suitable for structural mechanics analysis (such as solid elements and shell elements) are used, and the mesh is refined for key areas of the component (such as connection nodes and stress concentration areas) to improve simulation accuracy. Furthermore, the simulation time step is matched with the execution rhythm of the assembly process. The mechanical response data such as stress field, strain field, and displacement field of the component are recorded after each assembly action is completed. Finally, the mechanical response simulation data of the assembly process can be obtained.

[0079] Specifically, the mechanical response simulation data of the assembly process can be randomly sampled and divided into training and validation sets according to a preset ratio, typically with the training set accounting for 70%-80% and the validation set accounting for 20%-30%. Then, a preset set of fundamental solid mechanics equations can be embedded into the preset physical information neural network to integrate data-driven approaches with physical constraints, addressing the problems of traditional data-driven neural networks lacking physical meaning and having poor generalization ability. The fundamental solid mechanics equations can be equilibrium equations, and their embedding method involves using the equation residuals as part of the network loss function, forcing the network output to satisfy fundamental mechanical laws. The core loss function expression can be... , This is the total loss function of the physical information neural network, used to guide the optimization of network parameters; The data loss term characterizes the deviation between the network's predicted mechanical response and the simulated actual value, and is calculated using the mean square error. These are the physical constraint weighting coefficients, used to balance the requirements of data fitting accuracy and physical consistency; their values ​​are determined through cross-validation. The physical constraint loss term characterizes the degree to which the network output satisfies the solid mechanics equilibrium equations, i.e., the equilibrium equations. The residuals, of which Let be the divergence of the stress tensor. For stress tensor, The external load vector per unit volume.

[0080] The physical information neural network with physical constraints is trained based on the training set, and the network parameters are updated using the gradient descent algorithm. During the iteration process, the total loss function is continuously minimized. The trained neural network is obtained. A validation set is used to verify the performance of the trained neural network. Input data from the validation set (such as component geometric parameters, error parameters, and boundary conditions) is input into the network to obtain predicted mechanical response data. The deviation between the actual mechanical response data (i.e., finite element simulation results) and the predicted mechanical response data corresponding to the validation set is calculated. The root mean square error (RMSE) is used as the evaluation index for prediction error, which comprehensively reflects the overall degree of deviation between the predicted and actual values. When the prediction error is less than a preset error threshold, it indicates that the network's prediction accuracy meets the requirements of engineering applications. The training process is then stopped, and the corresponding trained neural network is designated as the physical information proxy model. This model can quickly output the mechanical response of the assembly process while ensuring prediction accuracy, replacing the computationally complex and time-consuming high-fidelity finite element simulation.

[0081] In one embodiment, the assembly process is dynamically represented by a graph based on bridge design parameters and a predefined assembly process sequence to obtain a dynamic graph sequence; a pre-defined temporal-spatial attention neural network is trained based on a physical information proxy model and the dynamic graph sequence to obtain a dynamic prediction model for assembly stress, including:

[0082] Based on the bridge design parameters, the geometric features and connection relationships of the assembled components are determined. Combined with the predefined assembly process sequence, the assembly structure of each step is defined as a graph structure of component nodes and connection relationship edges. The graph structure of component nodes and connection relationship edges is arranged according to the construction sequence to obtain a dynamic graph sequence.

[0083] Based on the physical information proxy model, the mechanical response of the assembly action corresponding to each step in the dynamic graph sequence is quickly calculated to obtain the component stress data at each step, and the component stress data at each step is used as the component mechanical state data. The component mechanical state data is mapped to the node features of the graph structure of component nodes and connection relationship edges. Based on the node features, the dynamic graph sequence is enhanced to obtain the feature-enhanced dynamic graph sequence.

[0084] Based on a predefined assembly process sequence, bridge design parameters, and digital samples of components with random manufacturing errors, the corresponding assembly scheme is determined; the mechanical response of the assembly scheme is simulated based on a physical information proxy model to obtain the residual stress field distribution data after assembly.

[0085] A temporal-spatial attention neural network, including a temporal attention layer and a spatial attention layer, is constructed. The dynamic graph sequence after feature enhancement is used as input, and the residual stress field distribution data after assembly is used as the training target. The network parameters of the temporal-spatial attention neural network are optimized through backpropagation to obtain the optimized neural network. The optimized neural network is used as the dynamic prediction model for assembly stress.

[0086] Specifically, bridge design parameters include core attributes such as the three-dimensional geometric dimensions, cross-sectional shape, material elastic modulus, and Poisson's ratio of the steel web members. They also specify key information such as the connection type between members, the location of connection nodes, and the force transmission path. Based on these parameters, the unique geometric features of each assembly member can be precisely extracted, including the member's outline dimensions, key hole coordinates, and web thickness distribution. The connection relationships between members can be clearly defined, such as rigid connections, flexible connections, and hinged connections, each corresponding to different mechanical transmission characteristics. The predefined assembly process sequence includes the assembly order of each member, the construction sequence of key connection nodes, and the connection conditions between processes. Combining this sequence, the assembled structure after each construction step in the assembly process can be abstracted as a graph structure of member nodes and connection edges. The initial feature vector of a member node consists of the member's geometric and material characteristics, while the features of the connection edges consist of the connection type, connection stiffness, and construction execution priority. For each construction step, a corresponding graphical structure can be generated according to the above rules. By arranging the graphical structures corresponding to all construction steps in an orderly manner according to the chronological order of the assembly process, a dynamic graphical sequence can be formed. This sequence can fully depict the temporal evolution of the structural topology during assembly, as well as the dynamic changes in the spatial relationships between components.

[0087] Specifically, the physical information proxy model has been trained with high-fidelity finite element simulation data and embedded with fundamental constraints of solid mechanics. It possesses the ability to rapidly output the mechanical response of assembly actions. For each assembly action corresponding to a step in the dynamic diagram sequence, the geometric parameters of the assembly represented by the structure of that step, the component manufacturing error parameters, and the assembly action execution parameters can be input into the physical information proxy model. Based on the embedded mechanical constraints and data-driven principles, the model can quickly calculate and output the global stress data of all components after the completion of that assembly action, including core indicators such as principal stress, shear stress, and equivalent stress. During the extraction of component stress data, stress data from stress concentration areas of the steel web, component connection node areas, and areas of abrupt stress change can be collected. The stress data from these global and key areas are then integrated into component mechanical state data. Furthermore, a feature embedding layer can be used to map the mechanical state data of components to node features. For example, this feature embedding layer transforms high-dimensional mechanical state data into mechanical feature vectors with the same dimension as the initial feature vectors of component nodes through linear transformation. Then, after feature concatenation, a fully connected layer performs nonlinear mapping to fuse this mechanical feature vector with the initial geometric-material feature vectors of the component nodes, generating new node feature vectors. Replacing the node features in the original graph structure with the updated node feature vectors can enhance the features of each step of the graph structure in the dynamic graph sequence, resulting in a feature-enhanced dynamic graph sequence. This sequence not only retains the spatiotemporal topological features of the assembly but also incorporates the mechanical response features of each construction step, significantly improving the information dimensionality and relevance of the input data.

[0088] The illustrative example uses digital samples of components with random manufacturing errors, including random features such as geometric and positional deviations from batch components. Combined with the clearly defined assembly accuracy and load transfer requirements in bridge design parameters, as well as the process execution rules in the predefined assembly process sequence, a unique assembly scheme can be matched for each digital sample of a component with random manufacturing errors. This assembly scheme specifies the assembly sequence of the corresponding components, the construction parameters of the connection nodes, and the time nodes for process execution. The complete parameters of each assembly scheme are input into the physical information proxy model. The model can simulate the mechanical transmission process of the entire assembly process, such as the mechanical response evolution of all stages including component positioning, connection fastening, and load application. Finally, it outputs the residual stress field distribution data of the entire assembly after the assembly scheme is executed. The residual stress field distribution data can be indexed by the three-dimensional spatial coordinates of the component, storing the magnitude, direction, and type of residual stress at each spatial coordinate point. The data is regularized in the form of a gridded tensor to ensure a one-to-one correspondence between the data and the spatial structure of the assembly. This data can serve as the target label for subsequent training of the temporal-spatial attention neural network, providing a reference for the supervised training of the model.

[0089] Specifically, the temporal-spatial attention neural network can be composed of an input layer, a spatial attention layer, a temporal attention layer, a feature fusion layer, a decoder, and an output layer. The spatial attention layer captures the spatial coupling relationships between component nodes within the same construction step diagram structure, highlighting the feature contributions of key component nodes along the stress transmission path by calculating the attention weights between nodes. The temporal attention layer captures the temporal dependencies between different construction step diagram structures, quantifying the influence of preceding assembly actions on the stress state of subsequent assemblies by calculating the attention weights of each construction step diagram structure. The feature-enhanced dynamic graph sequence is input into the neural network. The input layer performs vector encoding on the dynamic graph sequence, transforming the graph structure into a tensor form recognizable by the neural network. The encoded tensor is first passed to the spatial attention layer for weighted extraction of spatial features, and then to the temporal attention layer for weighted extraction of temporal features. The feature fusion layer deeply fuses the extracted spatial and temporal features to generate a high-dimensional feature tensor that integrates spatiotemporal mechanical features. The decoder then performs dimensionality reduction and mapping on the high-dimensional feature tensor. Finally, the output layer outputs the predicted residual stress field distribution data.

[0090] Furthermore, the network training process can employ supervised learning, using the residual stress field distribution data after assembly as the training objective. The weights and bias parameters of the network are iteratively optimized through backpropagation. The core loss function can be the mean squared error loss function, whose mathematical expression is: , This is the training loss function for the assembly stress dynamic prediction model, used to measure the degree of deviation between the predicted value and the true value; The total number of training samples is 1, and each training sample corresponds to a set of feature-enhanced dynamic graph sequences and corresponding residual stress field distribution data. For neural networks targeting the first The predicted residual stress field distribution tensor output by each training sample contains the predicted residual stress values ​​at all spatial points in the entire assembly. For the first The true residual stress field distribution tensor corresponding to each training sample, i.e. the residual stress field distribution data output by the physical information proxy model; It is the L2 norm, used to calculate the Euclidean distance between two tensors, quantifying the overall deviation between the predicted and the true values.

[0091] During training, the loss function is continuously calculated iteratively and the network parameters are updated using the gradient descent algorithm until the loss function value converges to a preset stable range. At this point, the optimized neural network is obtained. This optimized neural network is determined as the assembly stress dynamic prediction model. This model can accurately predict the distribution of residual stress field after assembly based on the dynamic graph sequence of the assembly process, providing core stress prediction support for the subsequent training of reinforcement learning agents.

[0092] In one embodiment, a reinforcement learning simulation environment is constructed based on a physical information agent model, and a reward function is constructed based on an assembly stress dynamic prediction model. Through the reinforcement learning simulation environment, a proximal policy optimization algorithm is employed, and a pre-defined agent is trained based on the reward function to obtain a deep reinforcement learning agent, including:

[0093] The assembly state vector is obtained by vector encoding based on the node and edge features of the dynamic graph sequence; the action dimension is divided based on the predefined assembly process sequence, and the discrete-continuous hybrid action space is defined; the action dimension includes assembly sequence, fastening torque and positioning reference.

[0094] The assembly state vector and the action commands in the action space are input into the assembly stress dynamic prediction model, and the predicted stress field distribution data and the maximum residual stress are output.

[0095] The stress uniformity index is calculated based on the predicted stress field distribution data; a reward function is constructed by combining the maximum residual stress, the stress uniformity index, and the preset process execution efficiency threshold.

[0096] A reinforcement learning simulation environment is constructed based on the physical information agent model. The reinforcement learning simulation environment is used to receive action instructions in the action space and output the corresponding assembly state vector and instant reward value based on the reward function.

[0097] Initialize the policy network and value network of the preset agent to obtain the initial agent; use the proximal policy optimization algorithm to collect experience data through the interaction between the initial agent and the reinforcement learning simulation environment, and iteratively optimize the policy network parameters and value network parameters of the initial agent based on the experience data to obtain the deep reinforcement learning agent.

[0098] Specifically, each node feature in the dynamic graph sequence includes the geometric properties, material properties, and mechanical response characteristics of the component, while the edge features characterize the connection type, connection stiffness, and construction priority between components. Illustratively, a graph attention network can be used to encode the node and edge features. A multi-head attention mechanism is used to calculate the attention weights between nodes, and after weighted fusion, a feature vector for each node is obtained. Then, a global pooling layer aggregates all node feature vectors into a fixed-dimensional assembly state vector, which comprehensively represents the structural and mechanical state of the current assembly. Based on a predefined assembly process sequence, action dimensions are defined, and a discrete-continuous hybrid action space is defined. The discrete action dimension corresponds to the optional adjustment scheme of the assembly sequence, and its size is determined by the number of processes in the predefined assembly process sequence. Each discrete action represents a process execution sequence, while the continuous action dimension can correspond to the adjustable parameters of the fastening torque and the positioning reference. The value range of the fastening torque is determined by the design specifications and material strength of the bolt connection, and the adjustable range of the positioning reference is determined by the assembly accuracy requirements of the component and the on-site construction conditions. Furthermore, the discrete-continuous hybrid action space can cover the core decision variables of assembly process optimization, providing the agent with a flexible strategy selection space.

[0099] Specifically, the assembly state vector and action commands can be concatenated to form a unified input tensor, which is then input into the assembly stress dynamic prediction model. This model, based on an embedded spatiotemporal attention mechanism, can deeply model the temporal dependencies and spatial coupling relationships of the assembly process, outputting predicted stress field distribution data for the entire assembly. This data uses three-dimensional spatial coordinates as an index to store the stress magnitude and direction at each spatial point. Subsequently, the maximum residual stress can be extracted from the predicted stress field distribution data, i.e., by traversing the stress values ​​of all spatial points and selecting the maximum value as the maximum residual stress. This value directly reflects the stress concentration risk of the assembly. Furthermore, a stress uniformity index can be calculated based on the predicted stress field distribution data. This index quantifies the dispersion of the stress distribution in the assembly. For example, by calculating the mean and standard deviation of all stress values ​​in the predicted stress field distribution data, the stress variation coefficient is obtained. The smaller the value, the more uniform the stress distribution and the higher the assembly reliability. The reward function is constructed by comprehensively considering the three objectives of safety, reliability, and process efficiency. Its mathematical expression can be:

[0100]

[0101] in, This is an immediate reward value; the higher the value, the better the optimization effect of the current assembly action. , , These are weighting coefficients, and the sum of the three is 1. They are used to balance the optimization priorities of safety, reliability, and process efficiency, and their values ​​are determined through cross-validation. To predict the maximum residual stress; The allowable residual stress of a steel web member is determined by material strength and design specifications. The stress variation coefficient is the ratio of the standard deviation to the mean of the predicted stress field. The execution time of the process standard is determined by the predefined assembly process sequence; The process execution time corresponding to the current action is calculated from the process sequence and fastening torque parameters in the action instruction.

[0102] Specifically, the core function of the reinforcement learning simulation environment is to receive action commands output by the agent, simulate the execution process of assembly actions, and provide feedback on the corresponding assembly state and immediate reward. When the environment receives an action command, it first inputs the action command and the current assembly state vector into the physical information agent model. The model quickly calculates the mechanical response and structural state changes after the assembly action is executed, generates a new dynamic graph sequence, obtains a new assembly state vector through vector encoding, and can call the reward function to calculate the immediate reward value based on the new predicted stress field distribution data and process execution time. Furthermore, the state transition logic of this environment follows the physical laws of the assembly process to ensure a high degree of consistency between the simulated scenario and actual construction conditions, providing reliable simulation support for the agent's interactive learning.

[0103] Specifically, the policy network and value network of the pre-defined agent can be initialized first. The policy network can adopt a deep neural network structure with three fully connected layers and an output layer. The input layer dimension is consistent with the assembly state vector dimension. The hidden layer uses the ReLU activation function to achieve non-linear feature mapping. The output layer can be divided into two branches: one branch outputs the probability distribution of discrete actions (assembly order adjustment schemes) through the softmax activation function, and the other branch outputs the Gaussian distribution mean and variance of continuous actions (tightening torque, positioning reference) through two independent fully connected layers. The branch that outputs the variance can use the Softplus activation function to ensure that the output is non-negative. The value network can adopt a deep neural network with the same hidden layer structure as the policy network. The input is the assembly state vector, and the output is a single-dimensional continuous value, i.e., the value estimate of the current assembly state, used to quantify the cumulative expected reward of all subsequent assembly actions in this state. During training, the initial agent can execute actions in the reinforcement learning simulation environment based on the current policy, collecting the assembly state vector, action instructions, immediate reward value, and the next assembly state vector at each step, forming empirical data and storing it in the empirical buffer.

[0104] Once the experience buffer has accumulated a certain amount of data, experience data can be randomly sampled from the buffer. A generalized advantage estimation algorithm is used to calculate the advantage function value for each state-action pair. This value measures the value advantage of the current action compared to the average action. Subsequently, based on the advantage function value and experience data, a surrogate objective function for the proximal policy optimization algorithm is constructed. The gradient ascent algorithm is used to maximize the surrogate objective function to update the policy network parameters. Simultaneously, a mean squared error loss function is used to minimize the deviation between the state value output by the value network and the cumulative reward to update the value network parameters. By continuously repeating the iterative process of experience collection, advantage estimation, and parameter updates, and calculating the agent's average reward value after a preset number of iterations, policy convergence is determined when the fluctuation range of the average reward value for multiple consecutive iterations is less than a preset threshold. The resulting agent is a deep reinforcement learning agent capable of outputting the optimal assembly process strategy based on the assembly state, achieving global optimization of assembly stress.

[0105] In one embodiment, such as Figure 2 As shown, a proximal policy optimization algorithm is employed. Through the interaction between a pre-defined agent and a reinforcement learning simulation environment, empirical data is collected. Based on this empirical data, the policy network parameters and value network parameters of the initial agent are iteratively optimized to obtain a deep reinforcement learning agent, including:

[0106] S201. Through the initial agent, the agent performs actions in the action space in the reinforcement learning simulation environment according to the current policy network parameters. The agent collects the assembly state vector, action instruction, immediate reward value and the next assembly state vector corresponding to each step, and forms an experience data sequence. The experience data sequence is stored in a preset experience buffer.

[0107] Specifically, the initial agent's policy network, based on current parameters, can output the probability distribution of discrete actions and the Gaussian distribution of continuous actions. The agent samples the action commands according to the distribution and inputs them into the reinforcement learning simulation environment. This environment simulates the mechanical response and structural state changes after the action is executed through a physical information agent model, providing feedback on the immediate reward value and the next assembly state vector. Simultaneously, the agent collects the current assembly state vector, action command, immediate reward value, and next assembly state vector, forming a four-tuple empirical data sequence, which is stored in a pre-defined circular empirical buffer. Furthermore, the buffer capacity is fixed; when the data volume exceeds the capacity, it automatically overwrites the oldest stored empirical data to ensure the timeliness and diversity of the sampled data.

[0108] S202. Randomly sample experience data from the preset experience buffer according to the preset batch, extract the assembly state vector and corresponding action instruction for each step from the experience data, and form a state-action pair; based on the generalized advantage estimation algorithm, combine the state value output by the value network and the immediate reward value, calculate the advantage function value corresponding to each state-action pair.

[0109] Specifically, once the amount of experience data stored in the experience buffer reaches a preset batch threshold, experience data can be randomly sampled without replacement from the buffer according to a preset batch size. The assembly state vector and action command for each step are then extracted to form a state-action pair. Subsequently, based on the generalized dominance estimation algorithm, the temporal difference residual is first calculated, and then iteratively weighted and accumulated using a preset discount factor and dominance cutoff coefficient to obtain the dominance function value corresponding to each state-action pair. This value is used to measure the value gain of the current action relative to the average policy, balancing estimation bias and variance, and improving the stability of policy updates.

[0110] S203. Based on empirical data and advantage function values, construct a proxy objective function based on a pruning mechanism, set a preset pruning coefficient, maximize the proxy objective function through a stochastic gradient ascent algorithm according to the preset pruning coefficient, update the policy network parameters of the initial agent, and obtain the updated policy network parameters.

[0111] Specifically, based on sampled empirical data and calculated advantage function values, a surrogate objective function based on a pruning mechanism can be constructed. For example, its mathematical expression is:

[0112]

[0113]

[0114] in, For the surrogate objective function based on the pruning mechanism, These are the learnable parameters of the policy network, and the iterative update of the assembly policy is achieved by maximizing this function. For time step The mathematical expectation of the reinforcement learning experience data is used to characterize the average optimization effect of policy updates; The ratio of action probabilities between the new and old strategies is used to measure the difference in action selection between the new and old strategies. For time step The generalized advantage estimate is used to characterize the value advantage of the current assembly action compared to the average action; For the clipping function, A preset pruning factor is used to limit the magnitude of policy updates and avoid oscillations during the training process; For parameters The strategy network, in the assembly state Down Output Action The probability of; For the old policy network before the update, in the assembly state Down Output Action The probability of this is determined. After setting a preset pruning coefficient, the agent's objective function can be maximized using the stochastic gradient ascent algorithm. The weights and bias parameters of the policy network are then iteratively updated to obtain the updated policy network parameters.

[0115] S204. Calculate the cumulative reward value based on empirical data, use the mean squared error loss function, minimize the deviation between the predicted value output by the value network and the cumulative reward value through the gradient descent algorithm, update the value network parameters of the initial agent, and obtain the updated value network parameters.

[0116] Specifically, based on sampled empirical data, the immediate reward value of each step is iteratively accumulated according to a preset discount factor to obtain the cumulative reward value of the corresponding state-action pair. Using the mean squared error loss function, the deviation between the state prediction value output by the value network and the cumulative reward value can be calculated. Furthermore, by minimizing this deviation using the gradient descent algorithm, the weights and bias parameters of the value network can be iteratively updated to obtain the updated value network parameters, making the long-term revenue estimation of the assembly state by the value network more accurate.

[0117] S205. Repeat steps S201 to S204. After each preset number of iterations, calculate the average reward value of the initial agent based on the immediate reward value obtained by the initial agent in the reinforcement learning simulation environment in each iteration. When the fluctuation range of the average reward value in consecutive preset number of iterations is less than the preset range threshold, the policy is determined to converge. Combine the corresponding updated policy network parameters and updated value network parameters to obtain the deep reinforcement learning agent.

[0118] Specifically, the process of repeatedly executing experience collection, batch sampling, advantage estimation, policy network update, and value network update is performed. After each preset number of iterations, the instantaneous reward value obtained by the agent in executing the complete assembly process within that iteration can be statistically analyzed, and the average reward value can be calculated. When the fluctuation range of the average reward value over consecutive preset number of iterations is less than a preset range threshold, the policy can be determined to have converged. The agent formed by the updated policy network parameters and value network parameters at this point is the deep reinforcement learning agent.

[0119] In one embodiment, the current construction status includes the progress of the work process, the positioning accuracy of the components, and the connection and fastening status.

[0120] The assembly stress dynamic prediction model is calibrated online based on on-site strain monitoring data to obtain an updated assembly stress dynamic prediction model. Based on the updated assembly stress dynamic prediction model and the current construction state, a rolling optimization calculation is performed on the deep reinforcement learning agent to output real-time optimized assembly process instructions, including:

[0121] Obtain the process parameters of the assembly process that has been executed, input the current construction status and the process parameters of the assembly process that has been executed into the assembly stress dynamic prediction model, and output the predicted strain data.

[0122] The error between the on-site strain monitoring data and the predicted strain data is calculated. Based on the error, the network weights of the assembly stress dynamic prediction model are fine-tuned using an adaptive learning rate to obtain the updated assembly stress dynamic prediction model.

[0123] Input the current construction status into the updated assembly stress dynamic prediction model to obtain the stress field prediction results output by the updated assembly stress dynamic prediction model.

[0124] The current construction state is encoded as a state vector. Based on the state vector and the stress field prediction results, the deep reinforcement learning agent is driven to perform rolling optimization calculations on the unassembled processes, generating real-time optimized assembly process instructions including assembly sequence adjustment instructions, fastening torque correction parameters, and positioning benchmark fine-tuning schemes, and outputting the real-time optimized assembly process instructions.

[0125] Specifically, the current construction status can include the progress of process execution, component positioning accuracy, and connection tightness. The progress of process execution can be quantified by the ratio of completed processes to the total number of processes. Component positioning accuracy can be quantified by the deviation between measured 3D coordinates and design coordinates. Connection tightness can be quantified by the ratio of actual bolt tightening torque to design torque. Illustratively, the process parameters of executed assembly processes can also be obtained. These parameters include the assembly sequence of completed processes, the actual tightening torque values ​​of each connection node, the positioning reference coordinates of each step, and the process execution time, ensuring that the data fully reflects the actual state of the executed process. First, the process parameters of the current construction status and executed assembly processes can be standardized to eliminate dimensional differences between parameters of different dimensions. Then, they are converted into fixed-dimensional input tensors through a feature encoding layer and input into the assembly stress dynamic prediction model. This model, based on an embedded temporal-spatial attention mechanism, can combine the cumulative mechanical response of the executed process to output predicted strain data corresponding one-to-one with the location of on-site monitoring points. This data is consistent with the spatial distribution of on-site strain monitoring, providing a matching benchmark for subsequent error calculation.

[0126] Specifically, on-site strain monitoring data can be acquired in real time using sensors deployed in key stress areas of components, such as fiber optic strain sensors and resistance strain gauges. After acquisition, the data undergoes filtering and noise reduction to eliminate environmental interference and equipment errors, ensuring data accuracy. Subsequently, the root mean square error between the on-site strain monitoring data and the predicted strain data can be calculated to obtain the error value, which comprehensively reflects the prediction deviation across all monitoring points. Based on this error value, an adaptive learning rate can be used to fine-tune the network weights of the assembly stress dynamic prediction model. For example, the formula for calculating the adaptive learning rate can be:

[0127]

[0128] In the formula, An adaptive learning rate is used to dynamically adjust the magnitude of weight updates; The initial learning rate is set based on the optimal learning rate during the model training phase. This is the attenuation coefficient, used to control the sensitivity of the learning rate to changes in error; This represents the root mean square error between the on-site strain monitoring data and the predicted strain data. Furthermore, during fine-tuning, only the weights of the top fully connected layer of the model can be updated. A mini-batch stochastic gradient descent algorithm is used, with the error between the monitored and predicted data as the loss target. The weight parameters are iteratively adjusted until the loss value converges to a preset range, ultimately yielding an updated dynamic prediction model for assembly stress. This model can adapt to actual on-site conditions and improve the accuracy of stress prediction.

[0129] Specifically, the current construction state can be transformed into a model-recognizable input vector through feature encoding and fed into the updated assembly stress dynamic prediction model. This model can combine the cumulative effect of the mechanical response of the executed processes with the constraints of the current construction state to predict the stress evolution process corresponding to the subsequent unexecuted processes, and output the stress field prediction results for the entire assembly. The results clearly mark the stress distribution, maximum stress value, and stress concentration location in key areas, providing a core basis for the intelligent agent's optimization decision-making.

[0130] Furthermore, the current construction state can be transformed into a state vector recognizable by a deep reinforcement learning agent through vector encoding. This vector integrates the temporal characteristics of process execution progress, the spatial characteristics of component positioning accuracy, and the mechanical characteristics of connection and fastening states, comprehensively representing the actual state of the current assembly system. This state vector is then concatenated with the stress field prediction results output by the updated assembly stress dynamic prediction model. The concatenated result can be used as input to the agent, driving it to perform rolling optimization calculations on unassembled processes based on the optimization strategies learned during training. During optimization, the agent can make multi-dimensional decisions regarding the assembly sequence of unexecuted processes, the fastening torques at each node, and the positioning reference. It prioritizes adjusting the process parameters corresponding to stress concentration areas. For example, adjusting the assembly sequence can avoid the risk of stress superposition, correcting the fastening torque can balance the stress distribution, and fine-tuning the positioning reference can reduce stress concentration caused by geometric deviations. Finally, a real-time optimized assembly process instruction is generated, including assembly sequence adjustment instructions, fastening torque correction parameters, and positioning reference fine-tuning schemes. This instruction is output to the construction control system using a standardized data format, directly guiding the parameter adjustment and process execution of on-site construction equipment.

[0131] Based on the same inventive concept, this application also provides a system for optimizing the assembly stress of steel web bridge components to implement the aforementioned method for optimizing the assembly stress of steel web bridge components. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the system for optimizing the assembly stress of steel web bridge components provided below can be found in the limitations of the method for optimizing the assembly stress of steel web bridge components described above, and will not be repeated here.

[0132] In one exemplary embodiment, such as Figure 3 As shown, a stress optimization system 300 for the assembly of steel web bridge components is provided, comprising:

[0133] The digital twin construction module 301 is used to acquire bridge design parameters and measured point cloud data of components, perform statistical analysis of manufacturing errors based on the bridge design parameters and measured point cloud data of components to obtain a probability distribution model of manufacturing errors, perform high-fidelity finite element simulation based on the probability distribution model of manufacturing errors to obtain simulation results, and train a preset physical information neural network based on the simulation results to obtain a physical information proxy model.

[0134] The dynamic stress field prediction module 302 is used to obtain a predefined assembly process sequence, perform dynamic graph representation of the assembly process based on bridge design parameters and the predefined assembly process sequence to obtain a dynamic graph sequence; and train a preset temporal-spatial attention neural network based on the physical information proxy model and the dynamic graph sequence to obtain an assembly stress dynamic prediction model.

[0135] The deep reinforcement learning decision module 303 is used to construct a reinforcement learning simulation environment based on the physical information agent model, construct a reward function based on the assembly stress dynamic prediction model, and train a pre-set agent through the reinforcement learning simulation environment using the near-end policy optimization algorithm and the reward function to obtain a deep reinforcement learning agent.

[0136] The closed-loop control execution module 304 is used to acquire the actual component manufacturing error, perform offline scheme calculation based on the actual component manufacturing error and the deep reinforcement learning agent, generate and execute the initial assembly process scheme; during the construction process according to the initial assembly process scheme, acquire on-site strain monitoring data and the current construction status; perform online calibration of the assembly stress dynamic prediction model based on the on-site strain monitoring data to obtain the updated assembly stress dynamic prediction model; drive the deep reinforcement learning agent to perform rolling optimization calculation based on the updated assembly stress dynamic prediction model and the current construction status, and output real-time optimized assembly process instructions.

[0137] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method for optimizing the assembly stress of steel web bridge components according to the present application. A multi-core processor is preferred to improve the parallel processing capability of the system. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate a large amount of data and computational tasks.

[0138] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method for optimizing the assembly stress of steel web bridge components according to the present application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0139] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method of steel web bridge member assembly stress optimization, characterized by, The method includes: Obtain bridge design parameters and measured point cloud data of components; perform statistical analysis of manufacturing errors based on the bridge design parameters and the measured point cloud data of components to obtain a probability distribution model of manufacturing errors; perform high-fidelity finite element simulation based on the probability distribution model of manufacturing errors to obtain simulation results; train a preset physical information neural network based on the simulation results to obtain a physical information proxy model. A predefined assembly process sequence is obtained, and the assembly process is dynamically represented based on the bridge design parameters and the predefined assembly process sequence to obtain a dynamic graph sequence; a preset temporal-spatial attention neural network is trained based on the physical information proxy model and the dynamic graph sequence to obtain an assembly stress dynamic prediction model. A reinforcement learning simulation environment is constructed based on the physical information agent model, and a reward function is constructed based on the assembly stress dynamic prediction model. Through the reinforcement learning simulation environment, a near-end policy optimization algorithm is adopted, and a policy is trained on a preset agent based on the reward function to obtain a deep reinforcement learning agent. The system acquires actual component manufacturing errors, performs offline scheme calculations based on these errors and the deep reinforcement learning agent, generates and executes an initial assembly process scheme, and acquires on-site strain monitoring data and the current construction status during construction according to the initial assembly process scheme. Based on the on-site strain monitoring data, the system performs online calibration of the assembly stress dynamic prediction model to obtain an updated model. Based on the updated model and the current construction status, the system drives the deep reinforcement learning agent to perform rolling optimization calculations and outputs real-time optimized assembly process instructions. The process includes the following steps: statistical analysis of manufacturing errors based on the bridge design parameters and measured point cloud data of the components to obtain a probability distribution model of manufacturing errors; high-fidelity finite element simulation based on the probability distribution model of manufacturing errors to obtain simulation results; and training a pre-defined physical information neural network based on the simulation results to obtain a physical information proxy model. The measured point cloud data of the component is registered with the theoretical model corresponding to the bridge design parameters to obtain registered point cloud data; the registered point cloud data is denoised to obtain denoised point cloud data; the deviation between the denoised point cloud data and the theoretical model is calculated to obtain component manufacturing error data. The manufacturing error data of the component is fitted with a normal distribution to obtain an initial distribution model; the initial distribution model is then subjected to a significance test to obtain a test result. When the test result is qualified, the corresponding initial distribution model is used as the probability distribution model of the manufacturing error. Based on the probability distribution model of the manufacturing error, random sampling is performed according to the preset sample number and error distribution range to generate multiple component digital samples with random manufacturing errors. Based on the bridge design parameters and the digital samples of components with random manufacturing errors, boundary constraints and load conditions corresponding to the assembly process are set; high-fidelity finite element simulation is performed according to the boundary constraints and load conditions to generate mechanical response simulation data of the assembly process. The mechanical response simulation data of the assembly process is divided into a training set and a validation set according to a preset ratio; a preset solid mechanics fundamental equation is embedded into the preset physical information neural network to obtain a physical information neural network with physical constraints; the physical information neural network with physical constraints is trained based on the training set to obtain a trained neural network; the trained neural network is validated using the validation set to obtain predicted mechanical response data; the deviation between the actual mechanical response data corresponding to the validation set and the predicted mechanical response data is calculated to obtain the prediction error; when the prediction error is less than a preset error threshold, training is stopped, and the corresponding trained neural network is used as the physical information proxy model.

2. The method according to claim 1, characterized in that, The assembly process is dynamically represented by the bridge design parameters and the predefined assembly process sequence to obtain a dynamic diagram sequence. Based on the physical information proxy model and the dynamic graph sequence, a preset temporal-spatial attention neural network is trained to obtain an assembly stress dynamic prediction model, including: Based on the bridge design parameters, the geometric features and connection relationships of the assembled components are determined. Combined with the predefined assembly process sequence, the assembly structure of each step is defined as a graph structure of component nodes and connection relationship edges. The graph structure of component nodes and connection relationship edges is arranged according to the construction sequence to obtain the dynamic graph sequence. Based on the physical information proxy model, the mechanical response of the assembly action corresponding to each step in the dynamic graph sequence is quickly calculated to obtain the component stress data of each step, and the component stress data of each step is used as the component mechanical state data. The mechanical state data of the component is mapped to the node features of the graph structure of the component node-connection relationship edge. Based on the node features, the dynamic graph sequence is enhanced to obtain the enhanced dynamic graph sequence. Based on the predefined assembly process sequence, the bridge design parameters, and the component digital samples with random manufacturing errors, a corresponding assembly scheme is determined; based on the physical information proxy model, a mechanical response simulation is performed on the assembly scheme to obtain the residual stress field distribution data after assembly. A temporal-spatial attention neural network, including a temporal attention layer and a spatial attention layer, is constructed. The enhanced dynamic graph sequence is used as input, and the residual stress field distribution data after assembly is used as the training target. The network parameters of the temporal-spatial attention neural network are optimized through backpropagation to obtain an optimized neural network. The optimized neural network is used as the assembly stress dynamic prediction model.

3. The method according to claim 1, characterized in that, The process involves constructing a reinforcement learning simulation environment based on the physical information agent model, building a reward function based on the assembly stress dynamic prediction model, and training a pre-defined agent using a near-end policy optimization algorithm within the reinforcement learning simulation environment and the reward function to obtain a deep reinforcement learning agent. This process includes: Based on the node and edge features of the dynamic graph sequence, vector encoding is performed to obtain the assembly state vector; based on the predefined assembly process sequence, action dimensions are divided to define a discrete-continuous hybrid action space; the action dimensions include assembly sequence, fastening torque, and positioning reference. The assembly state vector and the action commands in the action space are input into the assembly stress dynamic prediction model, and the predicted stress field distribution data and the maximum residual stress are output. The stress uniformity index is calculated based on the predicted stress field distribution data; the reward function is constructed by combining the maximum residual stress, the stress uniformity index, and the preset process execution efficiency threshold. The reinforcement learning simulation environment is constructed based on the physical information agent model. The reinforcement learning simulation environment is used to receive action instructions in the action space and output the corresponding assembly state vector and instantaneous reward value based on the reward function. The policy network and value network of the preset agent are initialized to obtain the initial agent; the proximal policy optimization algorithm is used to collect experience data through the interaction between the initial agent and the reinforcement learning simulation environment, and the policy network parameters and value network parameters of the initial agent are iteratively optimized based on the experience data to obtain the deep reinforcement learning agent.

4. The method according to claim 3, characterized in that, The method employs a proximal policy optimization algorithm, which collects experience data through the interaction between the preset agent and the reinforcement learning simulation environment, and iteratively optimizes the policy network parameters and value network parameters of the initial agent based on the experience data to obtain the deep reinforcement learning agent, including: S1. Through the initial agent, the agent performs actions in the action space in the reinforcement learning simulation environment according to the current policy network parameters, and collects the assembly state vector, the action instruction, the immediate reward value and the next assembly state vector corresponding to each step to form an experience data sequence, and stores the experience data sequence in a preset experience buffer. S2. Randomly sample the experience data from the preset experience buffer in preset batches, extract the assembly state vector and the corresponding action instruction for each step from the experience data, and form a state-action pair; based on the generalized advantage estimation algorithm, combine the state value output by the value network and the instant reward value, calculate the advantage function value corresponding to the state-action pair for each step; S3. Based on the empirical data and the advantage function value, construct a proxy objective function based on the pruning mechanism, set a preset pruning coefficient, maximize the proxy objective function through the stochastic gradient ascent algorithm according to the preset pruning coefficient, update the policy network parameters of the initial agent, and obtain the updated policy network parameters. S4. Calculate the cumulative return value based on the empirical data, use the mean squared error loss function, minimize the deviation between the predicted value output by the value network and the cumulative return value through the gradient descent algorithm, update the value network parameters of the initial agent, and obtain the updated value network parameters. S5. Repeat steps S1 to S4. After each preset number of iterations, calculate the average reward value of the initial agent based on the instantaneous reward value obtained by the initial agent in the reinforcement learning simulation environment in each iteration. When the fluctuation range of the average reward value in the preset number of consecutive iterations is less than the preset range threshold, the policy is determined to converge. Combine the corresponding updated policy network parameters and the updated value network parameters to obtain the deep reinforcement learning agent.

5. The method according to claim 1, characterized in that, The current construction status includes the progress of the work process, the accuracy of component positioning, and the status of connection and fastening. The assembly stress dynamic prediction model is calibrated online based on the on-site strain monitoring data to obtain an updated assembly stress dynamic prediction model; based on the updated assembly stress dynamic prediction model and the current construction state, the deep reinforcement learning agent is driven to perform rolling optimization calculations, outputting real-time optimized assembly process instructions, including: Obtain the process parameters of the executed assembly process, input the current construction status and the process parameters of the executed assembly process into the assembly stress dynamic prediction model, and output the predicted strain data. Calculate the error between the on-site strain monitoring data and the predicted strain data, and fine-tune the network weights of the assembly stress dynamic prediction model based on the error value using an adaptive learning rate to obtain the updated assembly stress dynamic prediction model. The current construction status is input into the updated assembly stress dynamic prediction model to obtain the stress field prediction result output by the updated assembly stress dynamic prediction model. The current construction state is encoded as a state vector. Based on the state vector and the stress field prediction result, the deep reinforcement learning agent is driven to perform rolling optimization calculations on the unassembled process, generate the real-time optimized assembly process instruction including assembly sequence adjustment instructions, fastening torque correction parameters, and positioning reference fine-tuning scheme, and output the real-time optimized assembly process instruction.

6. The method according to claim 4, characterized in that, The mathematical expression for the proxy objective function is: ； ； in, Let the proxy objective function be... These are the learnable parameters of the policy network; For time step The mathematical expectation is taken from the reinforcement learning experience data; The ratio of the probability of actions under the new and old strategies; For time step The generalized advantage estimate is used to characterize the value advantage of the current assembly action compared to the average action; For the clipping function, Preset clipping factor; For parameters The strategy network, in the assembly state Down Output Action The probability of; For the old policy network before the update, in the assembly state Down Output Action The probability of.

7. A stress optimization system for the assembly of steel web bridge components, characterized in that, The system includes: A digital twin construction module is used to acquire bridge design parameters and measured point cloud data of components. Based on the bridge design parameters and the measured point cloud data of components, a statistical analysis of manufacturing errors is performed to obtain a probability distribution model of manufacturing errors. Based on the probability distribution model of manufacturing errors, a high-fidelity finite element simulation is performed to obtain simulation results. Based on the simulation results, a preset physical information neural network is trained to obtain a physical information proxy model. The dynamic stress field prediction module is used to obtain a predefined assembly process sequence, perform dynamic graph representation of the assembly process based on the bridge design parameters and the predefined assembly process sequence to obtain a dynamic graph sequence; and train a preset temporal-spatial attention neural network based on the physical information proxy model and the dynamic graph sequence to obtain an assembly stress dynamic prediction model. The deep reinforcement learning decision module is used to construct a reinforcement learning simulation environment based on the physical information agent model, construct a reward function based on the assembly stress dynamic prediction model, and train a preset agent through the reinforcement learning simulation environment using a proximal policy optimization algorithm and the reward function to obtain a deep reinforcement learning agent. The closed-loop control execution module is used to acquire the actual component manufacturing error, perform offline scheme calculation based on the actual component manufacturing error and the deep reinforcement learning agent, generate and execute an initial assembly process scheme; during the construction process according to the initial assembly process scheme, acquire on-site strain monitoring data and the current construction status; perform online calibration of the assembly stress dynamic prediction model based on the on-site strain monitoring data to obtain an updated assembly stress dynamic prediction model; drive the deep reinforcement learning agent to perform rolling optimization calculation based on the updated assembly stress dynamic prediction model and the current construction status, and output real-time optimized assembly process instructions. The probability distribution model of the manufacturing error and the physical information proxy model are obtained through the following steps: The measured point cloud data of the component is registered with the theoretical model corresponding to the bridge design parameters to obtain registered point cloud data; the registered point cloud data is denoised to obtain denoised point cloud data; the deviation between the denoised point cloud data and the theoretical model is calculated to obtain component manufacturing error data. The manufacturing error data of the component is fitted with a normal distribution to obtain an initial distribution model; the initial distribution model is then subjected to a significance test to obtain a test result. When the test result is qualified, the corresponding initial distribution model is used as the probability distribution model of the manufacturing error. Based on the probability distribution model of the manufacturing error, random sampling is performed according to the preset sample number and error distribution range to generate multiple component digital samples with random manufacturing errors. Based on the bridge design parameters and the digital samples of components with random manufacturing errors, boundary constraints and load conditions corresponding to the assembly process are set; high-fidelity finite element simulation is performed according to the boundary constraints and load conditions to generate mechanical response simulation data of the assembly process. The mechanical response simulation data of the assembly process is divided into a training set and a validation set according to a preset ratio; a preset solid mechanics fundamental equation is embedded into the preset physical information neural network to obtain a physical information neural network with physical constraints; the physical information neural network with physical constraints is trained based on the training set to obtain a trained neural network; the trained neural network is validated using the validation set to obtain predicted mechanical response data; the deviation between the actual mechanical response data corresponding to the validation set and the predicted mechanical response data is calculated to obtain the prediction error; when the prediction error is less than a preset error threshold, training is stopped, and the corresponding trained neural network is used as the physical information proxy model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images