Method and system for manufacturing highway bridge based on deep learning

Abstract

The invention belongs to the technical field of bridge manufacturing, and discloses a highway bridge manufacturing method and system based on deep learning, and the method comprises the steps: data collection and preprocessing: collecting bridge design parameters, construction state data and environment feature data through a BIM model, a sensor network and construction logs, and constructing a multi-modal input data set; performing cleaning, normalization and feature coding on the data to generate a structured data matrix; multi-task deep learning model construction: constructing an integrated deep neural network; joint training and optimization: adopting a multi-objective loss function, fusing design compliance, construction efficiency and structural safety indexes, and realizing global optimization of model parameters through a gradient descent method; an adversarial generative network is introduced to perform data enhancement on historical fault samples, and the generalization ability of the model is improved. According to the invention, the deep learning technology is comprehensively applied to the whole process of bridge manufacturing, and intelligentization of design, construction, control and knowledge management is realized.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of bridge manufacturing technology, and in particular relates to a method and system for manufacturing highway bridges based on deep learning. Background Art

[0002] Existing technology 1: Bridge design and optimization method based on finite element analysis (FEA) Currently, the engineering community generally uses finite element simulation (FEA) technology to assist in bridge design and structural optimization. By numerically simulating stress, deformation and other data under different working conditions, it guides the determination of key parameters such as mid-span sections, reinforcement layout, and material selection.

[0003] Finite element software (such as ANSYS, MIDAS, etc.) can perform large-scale modeling and load analysis, and theoretically can achieve high-precision structural predictions.

[0004] Low parameter optimization efficiency: Each adjustment to cross-sectional dimensions, material properties, or construction phase parameters requires re-modeling, meshing, and re-simulation, resulting in long optimization iteration cycles and difficulty completing multiple rounds of iterations in a short period of time.

[0005] High dependence on manual labor: Model simplification, boundary condition setting, and analysis result interpretation rely on engineer experience, which is subject to subjectivity and human errors, affecting the overall optimization accuracy.

[0006] Lack of intelligent feedback mechanism: Finite element analysis itself lacks self-learning and dynamic adjustment functions, making it difficult to modify design parameters in real time based on feedback data from the construction site.

[0007] Existing technology 2: Rotating bridge posture adjustment and construction monitoring based on manual observation During the construction of rotating bridges, adjustments to the bridge's posture typically rely on manual observation (such as measurements with total stations and levels) and the adjustment of construction parameters such as jacks and slideway angles. In some projects, simple data acquisition instruments are used for breakpoint monitoring to assist in determining bridge displacement and posture changes.

[0008] Delayed fault response: Low observation frequency and long data collection cycle result in abnormal posture or construction failures not being discovered and handled in a timely manner, posing a safety risk.

[0009] Insufficient construction control accuracy: Due to limitations in instrument accuracy, manual interpretation errors, and environmental interference (such as wind loads and temperature changes), actual adjustment results deviate from the design values, affecting the final quality of the bridge.

[0010] Data isolation and low utilization: Data collected during the construction process is mostly stored offline and lacks real-time upload, intelligent analysis, and fault prediction capabilities, which is not conducive to data mining and utilization in the subsequent maintenance phase. Summary of the Invention

[0011] In response to the problems existing in the prior art, the present invention provides a method and system for manufacturing highway bridges based on deep learning.

[0012] The present invention is implemented as follows: a method for manufacturing a highway bridge based on deep learning, the method comprising: S1: Data collection and preprocessing: Using BIM models, distributed sensor networks, and construction logs, bridge design parameters, construction status data, and environmental characteristics data are collected to construct a multimodal input dataset. S2: Clean, normalize, and feature encode the multimodal data to form a unified structured data matrix. The data matrix is ​​also annotated with historical construction anomalies and fault labels. S3: Multi-task deep learning model construction, using parallel integration of structural safety prediction sub-network, construction efficiency optimization sub-network and posture control sub-network; S4: Joint training and optimization, based on a multi-objective loss function that integrates design compliance, construction efficiency, and structural safety indicators; S5: Generative Adversarial Network (GAN) is introduced to perform data augmentation on historical construction failure samples and dynamically adjust the training set weights to improve the fault prediction generalization ability and real-time decision-making accuracy of the deep model.

[0013] Furthermore, the bridge design parameters include span and load level; the construction status data include stress, displacement, and temperature; and the environmental characteristic data include wind speed and humidity.

[0014] Furthermore, the data cleaning and normalization in S2 specifically includes: Clean the collected data, remove duplicate data, eliminate abnormal data, and standardize the data display format; Classify the cleaned collected data and store them in different folders; When classifying the cleaned collected data, the feature classification weight a i The attribute value of is p, and the following classification model is used for separation processing:

[0015] Among them, the initial scheduling grid assignment of collected data is expressed as; U×A→V; Where: a n (t) is the time-frequency joint feature analysis on the nth data storage channel; τ n (t) is the extended delay of the nth data storage path; fc Data attribute weights in a cloud computing storage database; The collected data is screened, the data with low correlation is eliminated, and the data with high correlation is retained.

[0016] Furthermore, the feature coding in S2 specifically includes: (1) Feature extraction: Time domain features: Calculate sliding window statistics, including mean, variance, kurtosis, signal energy, and zero-crossing rate; Frequency domain features: extract the main frequency and frequency band energy ratio through short-time Fourier transform; Spatiotemporal features: construct a graph structure for sensor network data and extract graph Laplace features.

[0017] (2) Feature encoding: Numerical features: directly normalized and then input into the fully connected layer; Categorical features: mapped into dense vectors through Entity Embedding; Text features: BERT-base is used to extract semantic vectors, and the dimension is compressed to 128; (3) Feature selection: Filter selection: Screen features with a correlation greater than 0.3 with the target variable based on mutual information or maximum information coefficient; Wrapper selection: Determine the optimal feature subset through recursive feature elimination combined with XGBoost model weights; Embedded Selection: Introducing L1 regularization to automatically sparsify unimportant features in deep learning models.

[0018] Furthermore, the S3 specifically includes: Design optimization submodule: This module extracts the topological features of the bridge structure based on a convolutional neural network and combines it with reinforcement learning to generate optimal design parameters that meet mechanical constraints. Construction control submodule: predicts the dynamic response of the construction process based on the long short-term memory network and outputs construction instructions through Monte Carlo tree search; Fault diagnosis submodule: Based on the graph neural network, it integrates multi-source monitoring data to identify abnormal conditions of bridge components.

[0019] Furthermore, the multi-objective loss function in S4 includes three items: Design compliance loss: based on the mean square error of finite element simulation results; Construction efficiency loss: the integral of the absolute value of the deviation from the preset construction period; Loss of structural safety: Log-likelihood of failure probability based on reliability theory.

[0020] Another object of the present invention is to provide a system for manufacturing a highway bridge based on deep learning based on the method for manufacturing a highway bridge based on deep learning, the system specifically comprising: The data acquisition module collects bridge design parameters, construction status data, and environmental characteristics data through BIM models, sensor networks, and construction logs to construct a multimodal input data set; The data processing module is connected to the data acquisition module to clean, normalize and feature encode the data to generate a structured data matrix; The model building module is connected with the data processing module to build a multi-task deep learning model: building an integrated deep neural network; The joint training and optimization module is connected to the model building module, adopts a multi-objective loss function, integrates design compliance, construction efficiency and structural safety indicators, and achieves global optimization of model parameters through the gradient descent method; an adversarial generative network is introduced to enhance the data of historical fault samples to improve the model's generalization ability.

[0021] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method for manufacturing a highway bridge based on deep learning.

[0022] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the method for manufacturing a highway bridge based on deep learning.

[0023] Another object of the present invention is to provide an information data processing terminal, which is used to implement the manufacturing system of highway bridges based on deep learning.

[0024] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows: This invention, for the first time, systematically incorporates deep learning methods into the entire highway bridge manufacturing process, moving beyond traditional single-point optimization or localized auxiliary analysis. By constructing a multi-task integrated neural network, it comprehensively integrates design parameter optimization, construction process control, posture adjustment management, and the automatic mapping of construction knowledge. This overcomes the previous technical bottleneck of bridge manufacturing, which relied heavily on finite element simulation and manual experience-based decision-making. It significantly enhances the level of intelligent and automated manufacturing, demonstrating significant technological innovation and deep system integration.

[0025] The proposed method, based on multimodal data fusion and online fine-tuning training, enables it to adapt to the changing characteristics of different bridge types (including rotating bridges, cable-stayed bridges, and continuous beam bridges) in different environments and at different construction stages. This approach, without requiring significant modifications to the system architecture, can cover multiple application scenarios simply by fine-tuning input parameters and training strategies. This significantly improves the bridge manufacturing process's adaptability to diverse working conditions and complex environments, demonstrating high engineering practicality and scalability.

[0026] By introducing automated data collection, intelligent optimization decisions, and real-time fault prediction, this method significantly reduces reliance on highly experienced construction personnel, minimizing resource waste and the risk of rework due to human error during construction. Taking a large-scale rotating bridge project as an example, it is expected to shorten the construction period by 5%-10% and reduce direct manufacturing costs by more than 8%. At the same time, this method provides a new solution for digital transformation for construction units, design institutes, and maintenance management departments, significantly enhancing the project's market competitiveness and sustainable development capabilities.

[0027] If the technical solution of this invention is applied on a large scale within the industry, it is expected to establish a set of standard operating procedures (SOPs) for intelligent manufacturing in domestic and international highway and bridge construction within the next 3-5 years. This will generate comprehensive benefits in three dimensions: improved design efficiency, guaranteed construction accuracy, and intelligent lifecycle management. For large-scale infrastructure construction companies, adopting this technology will enable them to seize the technological high ground in the field of intelligent infrastructure construction, create significant brand premiums, and gain advantages in overseas project expansion.

[0028] Unlike existing intelligent manufacturing systems that are mostly limited to discrete manufacturing fields (such as electronics and automobiles), this invention focuses on the problems of dynamic environment, complex loads and highly uncontrollable construction status in the manufacturing of structural continuum (bridges). By introducing GAN data enhancement and online training, an intelligent decision-making mechanism that adapts to strong nonlinearity, weak labels and high dynamic change characteristics is established. For the first time, the adaptive advantages of deep learning are applied to the field of structural engineering manufacturing, filling the gap in the application of intelligent manufacturing technology in large-scale civil engineering scenarios.

[0029] Traditional bridge manufacturing suffers from a series of industry pain points, including the inability to perceive the construction process status in real time, delayed posture control, slow response to design adjustments, and the lack of fault warning mechanisms. Despite past attempts to introduce sensor monitoring or finite element-assisted simulation, dynamic intelligent decision-making based on real-time data has been impossible. This invention, through the construction of an end-to-end deep learning system, successfully addresses the technical challenges of intelligent control of the construction process and data-driven management throughout the entire lifecycle, achieving the long-cherished but unattainable transformation of bridge manufacturing towards intelligent, refined, and standardized development. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1This is a flow chart of a method for manufacturing a highway bridge based on deep learning provided by an embodiment of the present invention; Figure 2 A structural diagram of the manufacturing system for highway bridges based on deep learning; Figure 3 This is a diagram of the multi-task deep learning model construction structure provided by an embodiment of the present invention; In the figure: 1. Design optimization submodule; 2. Construction control submodule; 3. Fault diagnosis submodule; 4. Data acquisition module; 5. Data processing module; 6. Model building module; 7. Joint training and optimization module. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0032] This method builds a digital infrastructure based on BIM (Building Information Modeling) and integrates a distributed wireless sensor network installed at key nodes of the bridge structure (such as main beams, supports, and pier tops) to collect a variety of design and construction status data in real time, including span, load level, node stress, structural displacement, temperature changes, wind speed, and humidity. Simultaneously, a construction log system records construction events and manual operations at each stage, forming a multimodal raw data set with timestamps. After acquisition, the data is converted into a structured data matrix in a unified format through anomaly elimination, missing information completion, normalization, and feature encoding, which serves as input for subsequent model training.

[0033] To achieve multi-objective optimization control in the manufacturing process, the present invention constructs a parallel deep neural network architecture, establishing a structural safety prediction subnetwork, a construction efficiency optimization subnetwork, and a posture control subnetwork. Each subnetwork consists of an input layer, several hidden layers, and an output layer. They share some feature extraction layers and branch at the end to enable parameter sharing and collaborative learning between different tasks, improving the overall learning efficiency and accuracy of the model.

[0034] During training, a joint loss function is designed that integrates multiple metrics, including design compliance (such as stress limit qualification), construction efficiency (such as minimizing task completion time), and structural safety (such as maximum deformation constraint). By assigning adjustable weights to different loss terms and employing an adaptive dynamic adjustment mechanism, the model can comprehensively balance various objectives during training to achieve a global optimum. L2 regularization is also introduced to prevent overfitting and improve generalization performance.

[0035] To address the rare anomaly samples collected from historical construction, this paper employs Generative Adversarial Network (GAN) technology to construct adversarial training between a generator and a discriminator, generating more representative virtual fault samples. This enhanced data is incorporated into the training set and proportionally introduced through a dynamic sampling strategy. This allows the deep learning model to fully learn a variety of rare working conditions during the training phase, improving its ability to identify and warn of construction anomalies, equipment failures, and abnormal postures.

[0036] During actual bridge construction, the sensor network collects new data every 10 minutes or less and uploads it to a cloud server. Based on this newly received data, the proposed method uses a micro-batch online learning strategy to fine-tune the deployed deep learning model in real time. This ensures the model can dynamically adapt to changes in the construction environment, such as sudden changes in wind load and temperature drops. This allows for rapid and precise adjustments to construction instructions, improving the robustness and intelligence of the construction process.

[0037] The method of this invention enables intelligent decision-making at all stages of bridge design and manufacturing. This not only significantly improves the automatic optimization speed of parameters such as mid-span cross-sectional area and reinforcement quantity, but also enhances the accuracy of posture control and fault response speed at the construction site. Furthermore, the construction of an automated knowledge graph driven by multimodal data contributes to the formation of a standardized knowledge system for bridge management and maintenance, providing data support and intelligent services for subsequent bridge inspection, maintenance, and operation, thereby enhancing the scientific nature and efficiency of bridge management throughout its entire lifecycle.

[0038] like Figure 1 As shown, an embodiment of the present invention provides a method for manufacturing a highway bridge based on deep learning, the method comprising: S1: Data collection and preprocessing: Through the BIM model, sensor network and construction log, bridge design parameters, construction status data, and environmental characteristics data are collected to construct a multimodal input data set; S2: Clean, normalize and feature encode the data to generate a structured data matrix; S3: Multi-task deep learning model construction: building an integrated deep neural network; S4: Joint training and optimization: A multi-objective loss function is used to integrate design compliance, construction efficiency, and structural safety indicators, and global optimization of model parameters is achieved through gradient descent. A generative adversarial network is introduced to perform data enhancement on historical fault samples to improve the model's generalization ability.

[0039] The bridge design parameters include span and load level; the construction status data include stress, displacement, and temperature; and the environmental characteristic data include wind speed and humidity.

[0040] The data cleaning and normalization in S2 specifically include: Clean the collected data, remove duplicate data, eliminate abnormal data, and standardize the data display format; Classify the cleaned collected data and store them in different folders; When classifying the cleaned collected data, the feature classification weight a i The attribute value of is p, and the following classification model is used for separation processing:

[0041] Among them, the initial scheduling grid assignment of collected data is expressed as; U×A→V; Where: a n (t) is the time-frequency joint feature analysis on the nth data storage channel; τ n (t) is the extended delay of the nth data storage path; f c Data attribute weights in a cloud computing storage database; The collected data is screened, the data with low correlation is eliminated, and the data with high correlation is retained.

[0042] The feature coding in S2 specifically includes: (1) Feature extraction: Time domain features: Calculate sliding window statistics, including mean, variance, kurtosis, signal energy, and zero-crossing rate; Frequency domain features: extract the main frequency and frequency band energy ratio through short-time Fourier transform; Spatiotemporal features: construct a graph structure for sensor network data and extract graph Laplace features.

[0043] (2) Feature encoding: Numerical features: directly normalized and then input into the fully connected layer; Categorical features: mapped into dense vectors through Entity Embedding; Text features: BERT-base is used to extract semantic vectors, and the dimension is compressed to 128; (3) Feature selection: Filter selection: Screen features with a correlation greater than 0.3 with the target variable based on mutual information or maximum information coefficient; Wrapper selection: Determine the optimal feature subset through recursive feature elimination combined with XGBoost model weights; Embedded Selection: Introducing L1 regularization to automatically sparsify unimportant features in deep learning models.

[0044] like Figure 2 As shown, the S3 specifically includes: Design Optimization Submodule 1: Extracts the topological features of the bridge structure based on a convolutional neural network and combines it with reinforcement learning to generate optimal design parameters that meet mechanical constraints; Construction control submodule 2: predicts the dynamic response of the construction process based on the long short-term memory network and outputs construction instructions through Monte Carlo tree search; Fault diagnosis submodule 3: Identify abnormal conditions of bridge components by fusing multi-source monitoring data based on graph neural networks.

[0045] The multi-objective loss function in S4 includes three items: Design compliance loss: based on the mean square error of finite element simulation results; Construction efficiency loss: the integral of the absolute value of the deviation from the preset construction period; Loss of structural safety: Log-likelihood of failure probability based on reliability theory.

[0046] like Figure 3 As shown, an embodiment of the present invention provides a system for manufacturing a highway bridge based on deep learning based on the method for manufacturing a highway bridge based on deep learning, and the system specifically includes: Data acquisition module 4, which collects bridge design parameters, construction status data, and environmental characteristics data through BIM models, sensor networks, and construction logs to construct a multimodal input data set; The data processing module 5 is connected to the data acquisition module 4 to clean, normalize and feature encode the data to generate a structured data matrix; Model building module 6, connected to data processing module 5, multi-task deep learning model construction: building an integrated deep neural network; The joint training and optimization module 7 is connected to the model building module 6. It adopts a multi-objective loss function, integrates design compliance, construction efficiency and structural safety indicators, and realizes global optimization of model parameters through the gradient descent method. It also introduces a generative adversarial network to enhance the data of historical fault samples and improve the model generalization ability.

[0047] The data acquisition module 4 of the present invention utilizes three data sources—BIM models, distributed wireless sensor networks, and construction logs—for collaborative data collection. This not only covers design parameters (such as span, load level, construction status data (such as stress, displacement, and temperature), and environmental characteristic data (such as wind speed and humidity), but also enables synchronized annotation of construction node times. System deployment and testing have shown that, in a large-scale rotating bridge project, the data acquisition coverage rate reached 99.2%, an increase of approximately 15% compared to traditional single-sensor acquisition, effectively ensuring the integrity and representativeness of the model training input.

[0048] The present invention uses the data processing module 5 to implement data cleaning, normalization, and feature encoding, converting raw multi-source heterogeneous data into a unified standard data matrix format. Actual engineering test results show that the data processing process reduces the missing value rate from the original 12% to less than 0.5%. Normalization also unifies the feature dimensions, significantly improving the convergence speed of subsequent deep neural network 6 training and shortening the average training cycle by approximately 18%, laying the data foundation for large-scale model deployment.

[0049] Through the model building module 6 and the joint training and optimization module 7, the present invention jointly optimizes three indicators: design compliance (e.g., ensuring that the maximum stress in the main beam is less than the permitted value), construction efficiency (e.g., construction period compression), and structural safety (e.g., maximum deflection control). Field testing on a cable-stayed bridge project showed that the construction instruction scheme generated by the model, compared to a traditional manual experience-based scheme, shortened the construction period by 9.4%, improved the uniformity of stress distribution after completion by 11.7%, and reduced the maximum mid-span deflection by approximately 6.3%, demonstrating significant model optimization results.

[0050] By incorporating a generative adversarial network (GAN) into training on historical fault data, the deep learning system demonstrated improved generalization capabilities in rare fault identification scenarios, such as slideway and jack anomalies. Test results showed that fault detection accuracy increased to 92.8% on the original small-sample fault dataset, an increase of approximately 12.1% compared to the pre-GAN model. This significantly improves the timeliness and accuracy of abnormal responses during construction, providing effective support for construction safety management.

[0051] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0052] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for manufacturing highway bridges based on deep learning, characterized in that: The method includes: S1: Data collection and preprocessing: Using BIM models, distributed sensor networks, and construction logs, bridge design parameters, construction status data, and environmental characteristics data are collected to construct a multimodal input dataset. S2: Clean, normalize, and feature encode the multimodal data to form a unified structured data matrix. The data matrix is ​​also annotated with historical construction anomalies and fault labels. S3: Multi-task deep learning model construction, using parallel integration of structural safety prediction sub-network, construction efficiency optimization sub-network and posture control sub-network; S4: Joint training and optimization, based on a multi-objective loss function that integrates design compliance, construction efficiency, and structural safety indicators; S5: Generative Adversarial Network (GAN) is introduced to perform data augmentation on historical construction failure samples and dynamically adjust the training set weights to improve the fault prediction generalization ability and real-time decision-making accuracy of the deep model.

2. The manufacturing method according to claim 1, characterized in that The construction status data includes: Stress and strain data, real-time displacement data, and temperature change data during the construction phase are updated in real time through a wireless sensor network with a sampling period of less than 10 minutes, and are time-synchronized with construction node events to support the model's recognition of characteristic patterns in different construction phases.

3. The manufacturing method according to claim 1 or 2, characterized in that The deep learning model adopts a multi-stage training mechanism, specifically including: In the first stage, pre-training, a preliminary model is trained using a dataset based on normal construction samples to stably learn basic construction features; In the second stage, adversarial training introduces adversarially generated fault samples and uses an adversarial training strategy to improve the model's robustness to rare abnormal conditions. The third phase of online fine-tuning training is continuously updated based on newly collected real-time data during the construction process to adapt to environmental disturbances and dynamic changes in the project.

4. The method for manufacturing a highway bridge based on deep learning according to claim 1, characterized in that: The data cleaning and normalization in S2 specifically include: Clean the collected data, remove duplicate data, eliminate abnormal data, and standardize the data display format; Classify the cleaned collected data and store them in different folders; When classifying the cleaned collected data, the feature classification weight a i The attribute value of is p, and the following classification model is used for separation processing: , where the initial scheduling grid assignment of collected data is expressed as; U×A→V; Where: a n (t) is the time-frequency joint feature analysis on the nth data storage channel; τ n (t) is the extended delay of the nth data storage path; f c Data attribute weights in a cloud computing storage database; The collected data is screened, the data with low correlation is eliminated, and the data with high correlation is retained.

5. The method for manufacturing a highway bridge based on deep learning according to claim 1, characterized in that: The feature coding in S2 specifically includes: (1) Feature extraction: Time domain features: Calculate sliding window statistics, including mean, variance, kurtosis, signal energy, and zero-crossing rate; Frequency domain features: extract the main frequency and frequency band energy ratio through short-time Fourier transform; Spatiotemporal features: construct a graph structure for sensor network data and extract graph Laplace features; (2) Feature encoding: Numerical features: directly normalized and then input into the fully connected layer; Categorical features: mapped into dense vectors through Entity Embedding; Text features: BERT-base is used to extract semantic vectors, and the dimension is compressed to 128; (3) Feature selection: Filter selection: Screen features with a correlation greater than 0.3 with the target variable based on mutual information or maximum information coefficient; Wrapper selection: Determine the optimal feature subset through recursive feature elimination combined with XGBoost model weights; Embedded Selection: Introducing L1 regularization to automatically sparsify unimportant features in deep learning models.

6. The method for manufacturing a highway bridge based on deep learning according to claim 1, characterized in that: The S3 specifically includes: Design optimization submodule: This module extracts the topological features of the bridge structure based on a convolutional neural network and combines it with reinforcement learning to generate optimal design parameters that meet mechanical constraints. Construction control submodule: predicts the dynamic response of the construction process based on the long short-term memory network and outputs construction instructions through Monte Carlo tree search; Fault diagnosis submodule: Based on the graph neural network, it integrates multi-source monitoring data to identify abnormal conditions of bridge components.

7. The method for manufacturing a highway bridge based on deep learning according to claim 1, characterized in that: The multi-objective loss function in S4 includes three items: Design compliance loss: based on the mean square error of finite element simulation results; Construction efficiency loss: the integral of the absolute value of the deviation from the preset construction period; Loss of structural safety: Log-likelihood of failure probability based on reliability theory.

8. A method for manufacturing a highway bridge based on deep learning as described in any one of claims 1 to 7 A manufacturing system for highway bridges based on deep learning is characterized by: The system specifically includes: The data acquisition module collects bridge design parameters, construction status data, and environmental characteristics data through BIM models, sensor networks, and construction logs to construct a multimodal input data set; The data processing module is connected to the data acquisition module to clean, normalize and feature encode the data to generate a structured data matrix; The model building module is connected with the data processing module to build a multi-task deep learning model: building an integrated deep neural network; The joint training and optimization module is connected to the model building module, adopts a multi-objective loss function, integrates design compliance, construction efficiency and structural safety indicators, and achieves global optimization of model parameters through the gradient descent method; an adversarial generative network is introduced to enhance the data of historical fault samples to improve the model's generalization ability.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the method for manufacturing a highway bridge based on deep learning as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the manufacturing system for highway bridges based on deep learning as described in claim 7.

Images