A steel box girder three-dimensional model cost prediction method based on reinforcement learning

Abstract

This invention proposes a cost prediction method for 3D models of steel box girders based on reinforcement learning. It utilizes intelligent algorithms to deeply mine data from the 3D models of steel box girders, improving the accuracy and efficiency of cost prediction. The 3D model data of the steel box girder is preprocessed, converting the raw data into a standardized format and extracting multi-dimensional information such as geometric features, material properties, and structural complexity. A cost prediction model is constructed using a deep reinforcement learning algorithm. By optimizing the reward function and training strategy, the model can gradually approximate the actual manufacturing cost. The generalization ability and prediction accuracy are verified by comparing the model's prediction results with actual project data. Visualization tools are used to overlay the prediction results onto the 3D model, intuitively displaying high-cost areas and providing decision support for engineering design optimization and resource allocation. This method not only improves the accuracy of steel box girder manufacturing cost prediction but also provides a feasible technical solution for cost prediction of other complex structures.

Description

Technical Field

[0001] This invention relates to cost prediction technology in bridge engineering, and more particularly to a method for cost prediction of a three-dimensional model of a steel box girder based on reinforcement learning. Background Technology

[0002] Steel box girders, as an important load-bearing structure for bridges, are widely used in modern bridge engineering due to their advantages such as light weight, high torsional stiffness, and aesthetic appearance. However, the manufacturing process of steel box girders is complex, involving various materials, processes, and components. Their manufacturing cost is influenced by factors such as geometric characteristics, material properties, and structural complexity. Traditional cost estimation methods mainly rely on empirical formulas or expert judgment. These methods often struggle to balance prediction accuracy and efficiency when dealing with complex steel box girder designs, easily leading to cost overruns or resource waste.

[0003] With the popularization of Building Information Modeling (BIM) and Computer-Aided Design (CAD) technologies, 3D digital models have gradually become an important tool for bridge engineering design and construction. These models contain rich data on the geometric features, material properties, and connection relationships of steel box girders, providing a data foundation for cost prediction based on digital models. However, due to the diverse data formats, high model complexity, and non-linear characteristics of cost-influencing factors, how to efficiently and accurately utilize this data for cost prediction remains a pressing problem to be solved.

[0004] In recent years, artificial intelligence technology, especially reinforcement learning algorithms, has demonstrated powerful learning and prediction capabilities in complex decision-making problems. Reinforcement learning, by constructing an interaction model between the agent and its environment, can dynamically adjust strategies and approximate optimal solutions, making it highly suitable for solving multi-dimensional and nonlinear problems. In the field of steel box girder cost prediction, combining reinforcement learning with three-dimensional digital models can fully extract multi-dimensional information from the model, establishing a high-precision cost prediction model, thereby providing a scientific basis for engineering design optimization and resource allocation.

[0005] Therefore, it is necessary to provide a method for cost prediction of steel box girders using a three-dimensional model. This method uses intelligent algorithms to deeply mine and analyze the three-dimensional model data, thereby improving the accuracy and efficiency of the prediction and solving the aforementioned problems. Summary of the Invention

[0006] This invention aims to provide a cost prediction method for three-dimensional models of steel box girders based on reinforcement learning. By using intelligent algorithms to deeply mine the data in the three-dimensional model, it makes full use of multi-dimensional information such as geometric features, material information and structural complexity to achieve high-precision cost prediction.

[0007] To achieve the above objectives, this invention provides a method for cost prediction of a three-dimensional model of a steel box girder based on reinforcement learning. The main contents include preprocessing the three-dimensional model data of the steel box girder, extracting and analyzing features, constructing and training a reinforcement learning model to optimize and evaluate the model, predicting the cost of the model, displaying the prediction results, and periodically updating the model training data.

[0008] The preprocessing of the three-dimensional model data involves converting the three-dimensional model data of the steel box girder from the original format, namely BIM and CAD format, into a standardized neutral format, namely IFC standard format, to ensure the integrity of geometric information, material properties, and connection relationships.

[0009] The aforementioned feature extraction and analysis involves extracting the geometric features of the steel box girder, including plate thickness, dimensions, and volume; extracting material properties, including material type, density, and tensile strength; and extracting structural complexity, including weld length and node density. Furthermore, the model data is hierarchically stored in relational and graph databases according to the main structure (top plate, bottom plate, and web) and the secondary structure (U-ribs and stiffeners), generating multidimensional input feature vectors for reinforcement learning.

[0010] The construction and training of the reinforcement learning model refers to using a deep reinforcement learning algorithm to build the model, taking the features of each component of the steel box girder as the state space and the predicted cost of the component as the action space, and then designing a reward function to minimize the prediction error based on the difference between the actual manufacturing cost and the predicted cost. The reinforcement learning algorithm uses the deep Q-learning (DQN) algorithm to train the model, and optimizes the prediction path through experience replay, so that the model gradually approaches the actual cost.

[0011] The optimization and evaluation model refers to training the model using historical project data, generating predicted costs, calculating error metrics through the validation set, including mean squared error (MSE) and mean absolute error (MAE), and adjusting model parameters and reward functions if the error exceeds a threshold to continuously optimize the model.

[0012] The key to cost prediction using the model lies in inputting new 3D model data of steel box girders into the trained model, outputting the predicted manufacturing costs of each component and the whole, and then using visualization tools to overlay the prediction results onto the 3D model, visually displaying high-cost areas in a color-coded manner to assist in engineering design optimization and resource allocation. The model training data is periodically updated based on data from actual project feedback.

[0013] The 3D model data preprocessing process adopts a hierarchical storage structure, including a component table, which stores the geometric features, material information and complexity index of each component; a relationship table, which stores the connection relationships between components, including connection type and connection position; and a feature vector table, which vectorizes the features of all components to generate feature data for input reinforcement learning model.

[0014] The reinforcement learning model uses a deep Q-learning algorithm, which includes constructing a deep neural network consisting of an input layer, hidden layers, and an output layer. The input layer receives a multi-dimensional feature vector of the steel box girder component, including geometric features, material properties, and structural complexity. The output layer corresponds to the Q-value of each possible action, used to guide the prediction of component cost. A target network is introduced, maintaining a fixed parameter update interval with the online network. The parameters of the target network are updated using the soft update rule of the online network, with the following formula: Where θ represents the online network parameters, θ′ represents the target network parameters, and τ represents the soft update factor. Historical training data is stored in the experience pool. Training is performed by randomly sampling small batches of data to break the temporal correlation of data and improve training efficiency. Sampling probabilities are assigned according to the importance of historical samples, and samples with higher rewards or larger errors are prioritized for training to improve the model's ability to learn from key training data. An ε-greedy strategy is used to dynamically adjust the exploration rate ε. In the early stage, random actions are selected with a higher probability to increase exploration diversity, and in the later stage, ε is gradually reduced to select high Q-value actions to improve decision accuracy.

[0015] The reward function is designed to minimize the error between the actual cost and the predicted cost, and the specific formula is: R = -|C 预测 -C 实际 | where Cactual is the actual manufacturing cost, and Cpredicted is the model-predicted cost.

[0016] The following optimization steps are used during the model training process:

[0017] Step 1: Employ an experience replay mechanism by randomly sampling historical data to prevent data correlation during training from causing the model to get stuck in local optima.

[0018] Step two, exploration and utilization mechanism: initially set a high exploration rate to randomly explore different estimation paths, and gradually reduce the exploration rate in the later stage to improve the accuracy of cost prediction;

[0019] Step 3: Hyperparameter tuning, dynamically adjusting the learning rate and discount factor to balance learning speed and convergence stability at different training stages.

[0020] Compared with related technologies, the beneficial effects of the present invention are as follows:

[0021] (1) This invention deeply mines multi-dimensional information such as geometric features, material information and structural complexity of the three-dimensional model, and comprehensively analyzes the cost influencing factors of steel box girders, which can more accurately reflect the actual manufacturing cost and significantly improve the accuracy and reliability of cost prediction.

[0022] (2) The Deep Reinforcement Learning (DQN) algorithm is adopted, and the features of each component of the steel box girder are used as the state space and the predicted cost of the component is used as the action space. The prediction path is optimized through the reward function. The model has self-learning ability and can dynamically adjust parameters to gradually approach the actual cost, thereby achieving intelligent and efficient cost prediction.

[0023] (3) Using a hierarchical storage structure and feature vectorization technology, the geometric information, material properties and complexity index of steel box girder components are standardized and stored in a structured manner, which facilitates data management and model input, improves data processing efficiency, and ensures the integrity and consistency of information;

[0024] (4) Through continuous training and periodic updates of historical project data, the model has strong generalization ability, can adapt to the needs of different steel box girder design scenarios, maintain high prediction accuracy, and meet the diverse requirements in actual engineering applications.

[0025] (5) By adopting mechanisms such as experience replay, importance sampling and dynamic adjustment of exploration rate, the temporal correlation of training data is broken, key samples are learned first, the training efficiency and stability of the model are improved, and the model can converge quickly in different complex scenarios. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the cost prediction method for a three-dimensional model of a steel box girder based on reinforcement learning according to the present invention.

[0027] Figure 2 This is a diagram illustrating the preprocessing of the 3D model data in this invention.

[0028] Figure 3 This is a diagram illustrating the process of training the reinforcement learning model in this invention.

[0029] Figure 4 This is a flowchart illustrating the optimization and evaluation model process of this invention;

[0030] Figure 5 A diagram illustrating the application process of cost prediction using the model of this invention;

[0031] Figure 6 This is a schematic diagram of the hierarchical storage structure used in the data preprocessing process of this invention;

[0032] Figure 7 This is a schematic diagram illustrating the deep Q-learning algorithm used in the reinforcement learning model of this invention; Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0034] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0035] like Figure 1-2 As shown, the preprocessing of the 3D model data involves converting the 3D model data of the steel box girder from the original formats, namely BIM and CAD formats, into a standardized neutral format, namely the IFC standard format, to ensure the integrity of geometric information, material properties, and connection relationships. Through feature extraction and analysis, the geometric features of the steel box girder, including plate thickness, dimensions, and volume, are extracted; material properties, including material type, density, and tensile strength, are extracted; and structural complexity, including weld length and node density, is extracted. Furthermore, according to the main structure, including the top plate, bottom plate, and web plate, and also according to the secondary structure, including U-ribs and stiffeners, the model data is hierarchically stored in relational databases and graph databases to generate multidimensional input feature vectors for reinforcement learning.

[0036] like Figure 3 As shown, a reinforcement learning model for the construction and training of steel box girders is constructed using a deep reinforcement learning algorithm. The features of each component of the steel box girder are used as the state space, and the predicted cost of the components is used as the action space. A reward function is designed to minimize the prediction error based on the difference between the actual manufacturing cost and the predicted cost. The reinforcement learning algorithm uses the deep Q-learning (DQN) algorithm to train the model and optimizes the prediction path through experience replay, so that the model gradually approaches the actual cost.

[0037] like Figure 4 As shown, in the process of optimizing and evaluating the model, historical project data is used to train the model and generate the predicted cost. Then, the error index, including mean squared error (MSE) and mean absolute error (MAE), is calculated through the validation set. If the error exceeds the threshold, the model parameters and reward function are adjusted to continuously optimize the model so that it has strong generalization ability and the ability to adapt to different steel box girder designs.

[0038] like Figure 5As shown, in the process of applying the model to predict costs, new 3D model data of steel box girders are input into the trained model, and the predicted manufacturing costs of each component and the whole are output. The prediction results are superimposed on the 3D model through visualization tools, and high-cost areas are displayed intuitively in the form of color coding to assist in engineering design optimization and resource allocation. Based on the data feedback from actual projects, the model training data is periodically updated to further improve the prediction accuracy.

[0039] like Figure 6 As shown, a hierarchical storage structure is adopted in the 3D model data preprocessing process. There is a steel box girder component table, which stores the geometric features, material information and complexity index of each component. There is also a steel box girder relationship table, which stores the connection relationships between components, including connection type and connection position. There is also a steel box girder feature vector table, which vectorizes the features of all components to generate feature data for input reinforcement learning model.

[0040] like Figure 7 As shown, the reinforcement learning model uses the deep Q-learning algorithm, which includes constructing a deep neural network consisting of an input layer, hidden layers, and an output layer. The input layer receives the multi-dimensional feature vector of the steel box girder component, including geometric features, material properties, and structural complexity. The output layer corresponds to the Q-value of each possible action, used to guide the prediction of component cost. A target network is introduced, maintaining a fixed parameter update interval with the online network to reduce instability during training. The parameters of the target network are updated using the soft update rule of the online network, with the formula as follows: Where θ represents the online network parameters, θ′ represents the target network parameters, and τ represents the soft update factor. Historical training data is stored in the experience pool. Training is performed by randomly sampling small batches of data to break the temporal correlation of data and improve training efficiency. Sampling probabilities are assigned according to the importance of historical samples, and samples with higher rewards or larger errors are prioritized for training to improve the model's ability to learn from key training data. An ε-greedy strategy is used to dynamically adjust the exploration rate ε. In the early stage, random actions are selected with a higher probability to increase exploration diversity, and in the later stage, ε is gradually reduced to select high Q-value actions to improve decision accuracy.

[0041] The reward function before training is designed to minimize the error between the actual cost and the predicted cost. The specific formula is: R = -|C 预测 -C 实际 | where C is the actual manufacturing cost, and Cpredicted is the model prediction cost. The following optimization steps are used during the training process:

[0042] Step 1: Employ an experience replay mechanism by randomly sampling historical data to prevent data correlation during training from causing the model to get stuck in local optima.

[0043] Step two, exploration and utilization mechanism: initially set a high exploration rate to randomly explore different estimation paths, and gradually reduce the exploration rate in the later stage to improve the accuracy of cost prediction;

[0044] Step 3: Hyperparameter tuning, dynamically adjusting the learning rate and discount factor to balance learning speed and convergence stability at different training stages.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the descriptions in the foregoing specification are merely illustrative of the principles of the invention, and various changes and modifications can be made to the invention without departing from its spirit and scope; all such changes and modifications fall within the scope of the claimed invention.

[0046] The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for cost prediction of a three-dimensional model of a steel box girder based on reinforcement learning, characterized in that: The steel box girder 3D model data is preprocessed, features are extracted and analyzed, a reinforcement learning model is constructed and trained, the model is optimized and evaluated, the model is used to predict costs, the prediction results are displayed, and the model training data is updated periodically. The aforementioned feature extraction and analysis involves extracting the geometric features of the steel box girder, including plate thickness, dimensions, and volume; extracting material properties, including material type, density, and tensile strength; and extracting structural complexity, including weld length and node density. Furthermore, the model data is hierarchically stored in relational and graph databases according to the main structure (top plate, bottom plate, and web) and the secondary structure (U-ribs and stiffeners), generating multidimensional input feature vectors for reinforcement learning. The construction and training of the reinforcement learning model refers to the use of deep reinforcement learning algorithms to build the model, taking the features of each component of the steel box girder as the state space and the predicted cost of the component as the action space, and then designing a reward function to minimize the prediction error based on the difference between the actual manufacturing cost and the predicted cost. The reinforcement learning algorithm uses deep Q-learning algorithm to train the model and optimizes the prediction path through experience replay, so that the model gradually approaches the actual cost. The optimization and evaluation model refers to training the model using historical project data, generating predicted costs, calculating error metrics through the validation set, including mean squared error (MSE) and mean absolute error (MAE), and adjusting model parameters and reward functions to continuously optimize the model if the error exceeds the threshold. The key to the cost prediction using the model lies in inputting new 3D model data of steel box girders into the trained model, outputting the predicted manufacturing costs of each component and the whole, and then using visualization tools to overlay the prediction results onto the 3D model to intuitively display high-cost areas in a color-coded manner, thereby assisting in engineering design optimization and resource allocation. The model training data is periodically updated based on data from actual project feedback. The three-dimensional model data preprocessing process adopts a hierarchical storage structure, including a component table, which stores the geometric features, material properties and complexity index of each component; a relationship table, which stores the connection relationships between components, including connection type and connection position; and a feature vector table, which vectorizes the features of all components to generate feature data for input reinforcement learning model. The reinforcement learning model uses a deep Q-learning algorithm, which includes constructing a deep neural network consisting of an input layer, hidden layers, and an output layer. The input layer receives a multi-dimensional feature vector of the steel box girder component, including geometric features, material properties, and structural complexity. The output layer corresponds to the Q-value of each possible action, used to guide the prediction of component cost. A target network is introduced, maintaining a fixed parameter update interval with the online network. The parameters of the target network are updated using the soft update rule of the online network, with the following formula: Where θ is the online network parameter, θ′ is the target network parameter, and τ is the soft update factor. Historical training data is stored in the experience pool. Training is performed by randomly sampling small batches of data to break the temporal correlation of data and improve training efficiency. The sampling probability is assigned according to the importance of historical samples. Samples with higher rewards or larger errors are prioritized for training. The exploration rate ε is dynamically adjusted using an ε-greedy strategy. In the early stage, random actions are selected with a higher probability to increase exploration diversity. In the later stage, ε is gradually reduced to select high Q-value actions.

2. The method for cost prediction of a three-dimensional model of a steel box girder based on reinforcement learning as described in claim 1, characterized in that: The preprocessing of the three-dimensional model data involves converting the three-dimensional model data of the steel box girder from the original format, namely BIM and CAD format, into a standardized neutral format, namely IFC standard format, to ensure the integrity of geometric information, material properties, and connection relationships.

3. The method for cost prediction of a three-dimensional model of a steel box girder based on reinforcement learning as described in claim 2, characterized in that: The reward function is designed to minimize the error between the actual cost and the predicted cost, and the specific formula is: R = -|C 预测 -C 实际 |, where C 实际 For actual manufacturing costs, C 预测 To predict costs for the model.

4. The method for cost prediction of a three-dimensional model of a steel box girder based on reinforcement learning as described in claim 3, characterized in that: The following optimization steps are used during the model training process: Step 1: Employ an experience playback mechanism by randomly sampling historical data; Step two, exploration and utilization mechanism: initially set a high exploration rate and randomly explore different estimated paths, then gradually reduce the exploration rate in the later stages; Step 3: Hyperparameter tuning, dynamically adjusting the learning rate and discount factor to balance learning speed and convergence stability at different training stages.

Images