An AI-based method and system for intelligent management and control of the whole process of building construction and preposition of operation and maintenance

Abstract

The application relates to the technical field of building construction management, and discloses a method and system for intelligent management and control and pre-operation of building construction full process based on AI, which realizes fine simulation and optimization of the building construction and operation process by constructing a multi-modal digital twin environment and combining reinforcement learning and graph neural network technology. The method can dynamically generate virtual maintenance personnel and tools, and iteratively simulate and maintain the operation in the digital twin environment, real-time evaluates the interaction, optimizes the movement strategy online, and finally generates a structured optimization adjustment suggestion.

Description

Technical Field

[0001] This invention relates to the field of building construction management technology, and more specifically, to an AI-based method and system for intelligent control and pre-operation maintenance of the entire building construction process. Background Technology

[0002] Construction projects often encounter problems such as constantly changing construction environments, inefficient coordination between multiple processes, frequent conflicts in resource allocation, and untimely response to risks and hazards. Traditional management methods often rely on human experience to make decisions, making it difficult to achieve optimal global configuration. To address these challenges, an AI-based intelligent control and operation-prevention system for the entire construction process has been introduced, aiming to improve project management efficiency, optimize resource allocation, and reduce risks.

[0003] However, in practical applications, especially when faced with sudden changes in geological conditions and the resulting design adjustments, how to ensure short-term construction efficiency while also taking into account long-term operation and maintenance convenience has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides an AI-based method for intelligent management and pre-operation and maintenance of the entire construction process, aiming to solve the problem that existing AI-based intelligent management and pre-operation and maintenance systems for the entire construction process introduce high costs and operational difficulties in the future operation and maintenance phases when facing major design changes.

[0005] The technical solution of this application is as follows: This application discloses an AI-based method for intelligent management and pre-operation maintenance of the entire construction process, including: Construct a multimodal digital twin environment, which includes geometric information of the building environment, material properties, light source distribution, and maintainability annotation data of the target components; Using a reinforcement learning framework, a digital model of a virtual maintenance worker and a digital model of a virtual tool that matches the target component are dynamically generated based on the type of the component being maintained. The digital model of the virtual maintenance worker includes musculoskeletal parameters, joint degree-of-freedom constraints, and a biomechanical response model. In a multimodal digital twin environment, a physically based rendering engine and collision detection model are used to drive a virtual maintenance personnel digital model carrying a virtual tool digital model to perform iterative simulated maintenance operations on the target component. During the simulated maintenance operation, a graph neural network is used to evaluate in real time the interaction between the digital model of the virtual maintenance personnel, the digital model of the virtual tool, and the building environment. The interaction includes the physical contact force distribution based on finite element analysis, the operation posture comfort score based on a biomechanical database, and the cumulative joint stress value based on ergonomics. The interaction data is input into the reward function, and the motion strategy of the virtual maintenance personnel digital model is optimized online through a deep Q-network to minimize the maintenance operation risk index identified in the interaction data. The system generates structured optimization and adjustment suggestions, which include quantitative estimates of construction efficiency and operation and maintenance benefits. It also supports adaptive iterative verification of the optimization and adjustment suggestions until the maintenance operation risk index is lower than the Bayesian risk threshold trained from historical accident data.

[0006] Furthermore, the operational posture comfort score based on the biomechanical database was obtained through the following methods: Obtain the joint angle sequence and muscle activation of the digital model of the virtual maintenance personnel; Joint angle sequences and muscle activation are input into a pre-trained long short-term memory network, which is used to learn the non-linear mapping between human posture and subjective comfort. The output includes heatmaps of joint stress distribution and predicted fatigue accumulation values ​​for each maintenance operation stage; Based on the joint stress distribution heatmap and fatigue accumulation prediction, and combined with the operating frequency requirements of the target components, an interpretable long-term occupational health risk warning is generated.

[0007] Furthermore, graph neural networks are used to evaluate interactions in real time, including: The building environment, virtual maintenance personnel digital model and virtual tool digital model are constructed as a heterogeneous graph structure. The nodes of the heterogeneous graph structure include structural entities, human joints and tool contact surfaces. The edges of the heterogeneous graph structure represent spatial constraints, kinematic chain relationships and force transmission paths. A graph attention mechanism is used to calculate the influence weights between nodes, and to dynamically identify critical high-stress paths and operational constraint chains. Based on the changing trends of the influencing weights, predict the risk of operational conflicts within the next 5-10 seconds and adjust the motion planning of the virtual maintenance personnel digital model in advance.

[0008] Furthermore, adaptive iterative verification includes: An external environment feedback loop is set up to receive maintenance operation risk index and actual construction feedback data. The deviation signal output from the external environment feedback loop is input into the value function of reinforcement learning to fine-tune the motion strategy of the virtual maintenance personnel digital model online. Record the risk index decline curve after each iteration. When the slope of the risk index decline curve is lower than the preset threshold, the iteration will automatically stop and the current optimal maintenance plan will be output. At the same time, operation guidance data that can be read by on-site construction personnel wearing devices will be generated.

[0009] Furthermore, it also includes a quantitative assessment of visual accessibility and cognitive load: Acquire dynamic light source distribution and shadow areas in a multimodal digital twin environment; The head pose and gaze trajectory of the virtual maintenance personnel digital model are obtained, and the visible surface coverage of the maintenance target component is extracted through a convolutional neural network. When the visible surface coverage is below a preset threshold, the system automatically generates design modification suggestions for auxiliary lighting or adjustable access ports, and iteratively evaluates the modified visible surface coverage.

[0010] Furthermore, it also includes conflict prediction and avoidance in multi-person collaborative maintenance: Each virtual maintenance worker digital model is assigned a deep reinforcement learning agent, and each agent has an independent operation task priority and motion policy network. Multiple agents share the global digital twin environment state through a framework of centralized training and distributed execution, and learn mutual avoidance and cooperative behavior using counterfactual multi-agent policy gradients; The system predicts the spatial envelope conflict points between agents in real time and adjusts the operation sequence in a bidding manner to generate a conflict-free operation sequence diagram.

[0011] Furthermore, the Bayesian risk threshold is dynamically updated in the following ways: Collect actual safety accident data, operational error reports, and environmental monitoring data from construction sites; Construct a Bayesian hierarchical model, using project type, personnel experience level, time period, and weather conditions as covariates; Each time a simulation maintenance iteration is completed, the posterior distribution of the Bayesian hierarchical model is updated, and the Bayesian risk threshold moves accordingly to the high confidence interval. The updated Bayesian risk threshold is fed back to the reward function of the deep Q-network in real time, forming an online learning loop for risk perception.

[0012] Furthermore, the generation of structured optimization and adjustment suggestions also includes: The output is an enhanced parameter set that can be directly parsed by the building information model. The enhanced parameter set includes modified geometric dimensions, material recommendations, reserved operation space coordinates, and suggested maintenance tool types. Simultaneously, a construction efficiency gain curve and a full lifecycle operation and maintenance cost change graph are generated to support multi-objective optimization decisions; The enhanced parameter set, construction efficiency gain curve, and full life cycle operation and maintenance cost change chart are packaged into an executable decision package, which can be directly called by robots or augmented reality devices on the construction site.

[0013] Furthermore, the executable decision package is transmitted to at least one of the following devices: Path planning module for construction robot controller; Spatial anchor point positioning module for augmented reality headsets; Version update module for the Building Information Modeling (BIM) collaborative platform.

[0014] A system for intelligent management and pre-operation maintenance of the entire construction process based on AI, used to execute the aforementioned method for intelligent management and pre-operation maintenance of the entire construction process based on AI, including: The digital twin building module is used to construct a multimodal digital twin environment, which includes geometric information of the building environment, material properties, light source distribution, and maintainability annotation data of the target components. An artificial intelligence decision engine, comprising a reinforcement learning framework and a deep Q-network, is used to dynamically generate digital models of virtual maintenance personnel and virtual tools, and drive them to perform iterative simulated maintenance operations in a multimodal digital twin environment. The graph neural network evaluation module is used to evaluate the interaction between virtual maintenance personnel and virtual tools and the building environment in real time, and output the maintenance operation risk index. A risk adaptive threshold updater is used to dynamically adjust the risk threshold based on a Bayesian hierarchical model and feed it back to the artificial intelligence decision engine to form an online learning closed loop. The optimization suggestion generation module is used to generate structured optimization and adjustment suggestions and output an executable decision package to an external device.

[0015] The beneficial effects of this invention are: This application discloses an AI-based intelligent management and operation-maintenance pre-process method and system for the entire construction process. It constructs a multimodal digital twin environment to dynamically generate digital models of virtual maintenance personnel and tools, and performs iterative simulation of maintenance operations within this environment. During the simulation, a graph neural network is used to evaluate the interaction between the virtual maintenance personnel, virtual tools, and the building environment in real time, including physical contact force distribution, operational posture comfort scores, and cumulative joint stress values. These interactions are input into a reward function, and a deep Q-network is used to optimize the motion strategy of the virtual maintenance personnel online to minimize the maintenance operation risk index. Finally, structured optimization and adjustment suggestions are generated, and adaptive iterative verification is supported until the risk index is lower than the Bayesian risk threshold. Attached Figure Description

[0016] Figure 1This is a flowchart of an AI-based intelligent management and operation-maintenance pre-process method for the entire construction process, provided by an embodiment of the present invention. Figure 2 This is a structural diagram of an AI-based intelligent management and operation and maintenance system for the entire construction process, provided by an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Reference Figure 1 This invention provides an AI-based method for intelligent management and pre-operation maintenance of the entire construction process, comprising: Step S11: Construct a multimodal digital twin environment. The multimodal digital twin environment includes geometric information of the building environment, material properties, light source distribution, and maintainability annotation data of the target components. Step S12: Using a reinforcement learning framework, a digital model of a virtual maintenance worker and a digital model of a virtual tool that matches the target component are dynamically generated based on the type of the target component being maintained. The digital model of the virtual maintenance worker includes musculoskeletal parameters, joint degree-of-freedom constraints, and a biomechanical response model. Step S13: In a multimodal digital twin environment, a physically based rendering engine and a collision detection model are used to drive the virtual maintenance personnel digital model to carry the virtual tool digital model to perform iterative simulated maintenance operations on the maintenance target component. Step S14: During the simulated maintenance operation, the interaction between the virtual maintenance personnel digital model, the virtual tool digital model and the building environment is evaluated in real time using a graph neural network. The interaction includes the physical contact force distribution based on finite element analysis, the operation posture comfort score based on a biomechanical database, and the cumulative joint stress value based on ergonomics. Step S15: Input the interaction situation into the reward function, and optimize the motion strategy of the virtual maintenance personnel digital model online through a deep Q network to minimize the maintenance operation risk index identified in the interaction situation; Step S16: Generate structured optimization and adjustment suggestions. The optimization and adjustment suggestions include quantitative estimates of construction efficiency and operation and maintenance benefits, and support adaptive iterative verification of the optimization and adjustment suggestions until the maintenance operation risk index is lower than the Bayesian risk threshold trained by historical accident data.

[0019] This application achieves refined simulation and optimization of building construction and operation processes by constructing a multimodal digital twin environment and combining reinforcement learning and graph neural network technologies. This method can dynamically generate virtual maintenance personnel and tools, and perform iterative simulation maintenance operations within the digital twin environment, evaluating interactions in real time to optimize movement strategies online and ultimately generate structured optimization and adjustment suggestions. This approach not only effectively solves problems such as inefficiency, poor resource allocation, and untimely risk response in traditional management methods, but more importantly, it deeply integrates factors such as "future operation and maintenance operability," "long-term maintenance costs," and "accessibility throughout the equipment lifecycle," thereby achieving a balance between short-term construction efficiency and long-term operation and maintenance benefits, avoiding the situation of "efficient construction, inefficient maintenance."

[0020] To better understand the method proposed in this application, the key terms and implementation environment involved will be explained in detail below.

[0021] A "multimodal digital twin environment" refers to a virtual simulation space that integrates multiple data types (such as geometric information, material properties, light source distribution, and maintainability annotation data), capable of reflecting the characteristics of the real building environment in real time and with high fidelity. This environment forms the basis for simulation operations, providing a realistic interactive scenario for virtual maintenance personnel and virtual tools.

[0022] A reinforcement learning framework is a machine learning paradigm that learns optimal policies through the interaction of an agent with its environment. In this application, it is used to dynamically generate digital models of virtual maintenance personnel and virtual tools, and drive them to perform simulated operations.

[0023] The "virtual maintenance personnel digital model" is a virtual entity that includes musculoskeletal parameters, joint degree-of-freedom restrictions, and biomechanical response models, which can simulate the movements and physiological reactions of real maintenance personnel.

[0024] A "virtual tool digital model" is a virtual tool that matches the target component being maintained. Its physical properties and geometry are precisely modeled to simulate the use of a real tool.

[0025] A "physically based rendering engine" is a graphics rendering technology that simulates the propagation and reflection of light in a virtual environment to generate realistic visual effects.

[0026] The "collision detection model" is used to detect physical contact and collisions between virtual maintenance personnel, virtual tools, and the built environment, ensuring the realism of the simulation.

[0027] A graph neural network is a type of neural network that processes graph-structured data, effectively capturing the relationships and information transmission between nodes. In this application, it is used to evaluate the real-time interaction between virtual maintenance personnel and virtual tools and the built environment.

[0028] "Deep Q-Network" is a reinforcement learning algorithm that combines deep learning and Q-learning to optimize the movement strategies of virtual maintenance personnel online.

[0029] The "Bayesian risk threshold" is a risk assessment standard trained based on historical accident data, used to determine whether the risk of maintenance operations is acceptable.

[0030] The method proposed in this application can be implemented in various computing environments, such as high-performance computing clusters, cloud computing platforms, or edge computing devices. These environments require sufficient computing power and storage space to support large-scale digital twin modeling, real-time simulation, and deep learning training.

[0031] The method proposed in this application achieves intelligent control and pre-positioning of operation and maintenance throughout the entire construction process through a series of collaborative technical features.

[0032] First, constructing a multimodal digital twin environment is the foundation of the entire methodology. This environment can be viewed as a highly simulated virtual world, containing the geometric information of the building environment, material properties, light source distribution, and maintainability annotation data for the target components. For example, a precise 3D geometric model of the building can be obtained through technologies such as laser scanning and photogrammetry, and then imported into the digital twin platform. Simultaneously, realistic material properties, such as density, strength, and thermal conductivity, can be assigned to components like walls, pipes, and equipment within the building. Light source distribution information can be obtained by simulating natural and artificial lighting under different times and weather conditions. Maintainability annotation data can be added manually or semi-automatically to the target components, such as marking the location of access panels, bolt types, and disassembly sequences. These data collectively construct a comprehensive and detailed digital twin environment, providing a realistic and reliable background for subsequent simulation operations.

[0033] Secondly, using a reinforcement learning framework, a digital model of a virtual maintenance worker and a digital model of virtual tools matching the target component are dynamically generated based on the type of the component being maintained. For example, when maintaining a high-altitude valve, a reinforcement learning framework can be used to generate a digital model of a virtual maintenance worker with appropriate height, arm span, and joint flexibility, and to equip them with a suitable wrench or power tool. The virtual maintenance worker model includes musculoskeletal parameters, joint freedom constraints, and a biomechanical response model, making its behavior closer to that of a real human. For example, the range of motion of the virtual maintenance worker's joints can be preset, and the muscle stress under different postures can be simulated.

[0034] Next, in the multimodal digital twin environment, a physically based rendering engine and collision detection model are used to drive a virtual maintenance worker digital model carrying virtual tool digital models to perform iterative simulated maintenance operations on the target components. For example, the virtual maintenance worker can be driven to tighten a virtual bolt or disassemble a virtual panel. The physically based rendering engine ensures the visual realism of the simulation process, allowing operators to intuitively observe the actions of the virtual maintenance worker and the use of tools. The collision detection model monitors the physical contact between the virtual maintenance worker, virtual tools, and the building environment in real time, avoiding clipping or unreasonable collisions, thereby ensuring the effectiveness of the simulation operation.

[0035] During simulated maintenance operations, graph neural networks are used to evaluate in real time the interaction between the virtual maintenance personnel's digital model, the virtual tool's digital model, and the built environment. This interaction includes physical contact force distribution based on finite element analysis, comfort scores for operational postures based on a biomechanical database, and cumulative joint stress values ​​based on ergonomics. For example, when a virtual maintenance personnel operates in a confined space, the graph neural network can analyze the contact force distribution between their body and surrounding structures, identifying potential crush or collision risks. Simultaneously, by comparing with a biomechanical database, the comfort of the virtual maintenance personnel's current operational posture can be assessed, such as determining whether they are in a prolonged bent-over or arm-raised position. Furthermore, the cumulative stress values ​​of the virtual maintenance personnel's joints can be calculated to predict potential fatigue or injury after prolonged operation.

[0036] Subsequently, the interaction data is input into a reward function, and a deep Q-network is used to optimize the motion strategy of the virtual maintenance worker's digital model online, minimizing the maintenance operation risk index identified in the interactions. For example, if the graph neural network evaluates a certain operation posture as having a low comfort score or excessive joint stress accumulation, the deep Q-network will treat these unfavorable interactions as negative rewards, prompting the virtual maintenance worker to adjust its motion strategy and find a safer and more comfortable operating method. Through continuous iterative optimization, the virtual maintenance worker's motion strategy will gradually approach the optimal level, thereby significantly reducing the maintenance operation risk.

[0037] Finally, structured optimization suggestions are generated. These suggestions include quantitative estimates of construction efficiency and operational benefits, and support adaptive iterative verification until the maintenance operation risk index falls below a Bayesian risk threshold trained from historical incident data. For example, optimization suggestions might include modifying the installation location of target maintenance components, adjusting the size of access panels, recommending the use of specific tool types, or replanning maintenance routes. These suggestions come with quantitative estimates of improved construction efficiency and reduced operational costs, providing data support for decision-makers. The adaptive iterative verification mechanism ensures the effectiveness of the optimization suggestions; the optimization process only stops when the maintenance operation risk index falls below the preset Bayesian risk threshold, thus guaranteeing the safety and reliability of the final solution.

[0038] The method proposed in this application achieves intelligent control and pre-operation maintenance of the entire construction process by constructing a multimodal digital twin environment and combining advanced AI technologies such as reinforcement learning, graph neural networks, and deep Q-networks.

[0039] This application utilizes digital models of virtual maintenance personnel and tools, along with graph neural networks, to dynamically identify and quantify potential operational risks through real-time evaluation of interactions. For example, when a virtual maintenance personnel discovers that a access port is difficult to reach during simulation, or when joint stress is excessive during operation, the system identifies this as a high-risk scenario and optimizes the movement strategy using a deep Q-network, even generating suggestions to modify the access port's location or size. This proactive, behavior-simulation-based risk assessment and optimization effectively avoids the situation of "efficient construction, inefficient maintenance," achieving a balance between short-term construction efficiency and long-term operational benefits. In this way, this application significantly improves the level of intelligent control and proactive operation and maintenance throughout the entire construction process, enhancing its practical value.

[0040] Specifically, the method for obtaining the operational posture comfort score based on the biomechanical database can be further refined. The acquisition of this score includes the following steps: First, obtain the joint angle sequence and muscle activation of the digital model of the virtual maintenance personnel; Secondly, the joint angle sequence and muscle activation are input into the pre-trained long short-term memory network, which learns the non-linear mapping between human posture and subjective comfort. Next, the output includes a heat map of joint stress distribution and fatigue accumulation prediction values ​​for each maintenance operation stage; Finally, based on the joint stress distribution heatmap and fatigue accumulation prediction, combined with the operating frequency requirements of the target components, an interpretable long-term occupational health risk warning is generated.

[0041] Specifically, when acquiring the joint angle sequence and muscle activation of a virtual maintenance worker's digital model, the preset musculoskeletal parameters, joint degree-of-freedom constraints, and biomechanical response model within the digital model can be used to capture the joint motion trajectory and muscle contraction state of the virtual maintenance worker's digital model in real time during simulated maintenance operations. The joint angle sequence refers to the angle change data of each joint in the virtual maintenance worker's digital model over time, while muscle activation reflects the degree of force exerted by a specific muscle group during the execution of an action.

[0042] Furthermore, the acquired joint angle sequences and muscle activation levels are fed into a pre-trained Long Short-Term Memory (LSTM) network. LSM is a type of recurrent neural network, particularly well-suited for processing and predicting time-series data. This network learns a non-linear mapping between a large amount of real human posture data and corresponding subjective comfort scores, enabling it to accurately assess the comfort of the virtual maintenance personnel digital model under different operating postures. The pre-training process can be based on ergonomic experimental data, professional evaluation data, and historical operation records.

[0043] Therefore, the Long Short-Term Memory (LSTM) network will output a heatmap of joint stress distribution and cumulative fatigue prediction values ​​for each stage of maintenance operations. The joint stress distribution heatmap visually shows the areas and intensity of pressure experienced by each joint in the virtual maintenance personnel's digital model during operation; darker colors indicate greater stress. The cumulative fatigue prediction values ​​quantify the degree of physical fatigue that may result from prolonged or repetitive operations.

[0044] Ultimately, based on the joint stress distribution heatmap and fatigue accumulation prediction values, combined with the actual operating frequency requirements of the target components, the system can generate interpretable long-term occupational health risk warnings. These warnings not only identify potential risks but also provide specific explanations, such as "Maintaining a bent-over posture for extended periods may lead to excessive stress on the lumbar spine; it is recommended to rest or adjust your posture every 30 minutes."

[0045] This application's solution involves a refined analysis of the joint angle sequences and muscle activation levels of a virtual maintenance worker's digital model. By utilizing a long short-term memory network to learn the complex mapping relationship between human posture and subjective comfort, it can output detailed joint stress distribution heatmaps and fatigue accumulation predictions. Through this technical solution, this application provides a more refined and comprehensive mechanism for assessing operational posture comfort. This mechanism not only assesses the immediate comfort of the virtual maintenance worker's digital model during simulated operation, but more importantly, by outputting joint stress distribution heatmaps and fatigue accumulation predictions, and combining them with the operational frequency requirements of the target maintenance components, it can generate interpretable long-term occupational health risk warnings. This allows the system to fully consider the long-term health and occupational safety of operators when optimizing maintenance operation strategies, effectively avoiding the risks of chronic strain and occupational diseases caused by unreasonable operating postures, thereby significantly improving the humanization and sustainability of maintenance solutions.

[0046] In some of the embodiments described above in this application, the specific implementation of the step of using a graph neural network to evaluate the interaction in real time during the simulated maintenance operation is further explained.

[0047] Specifically, the above-mentioned use of graph neural networks to evaluate interactions in real time includes: The building environment, virtual maintenance personnel digital model and virtual tool digital model are constructed as a heterogeneous graph structure. The nodes of the heterogeneous graph structure include structural entities, human joints and tool contact surfaces. The edges of the heterogeneous graph structure represent spatial constraints, kinematic chain relationships and force transmission paths. A graph attention mechanism is used to calculate the influence weights between nodes, and to dynamically identify critical high-stress paths and operational constraint chains. Based on the changing trends of the influencing weights, predict the risk of operational conflicts within the next 5-10 seconds and adjust the motion planning of the virtual maintenance personnel digital model in advance.

[0048] Specifically, when constructing a heterogeneous graph structure, the built environment can be decomposed into multiple structural entities, such as walls, floors, and equipment components, which serve as nodes in the graph. The virtual maintenance personnel digital model is abstracted into a series of human joints, such as shoulders, elbows, wrists, hips, knees, and ankles, which also serve as nodes in the graph. The virtual tool digital model can use its key contact surfaces or operation points as nodes. Each node can contain its own attribute information.

[0049] The use of graph attention mechanism to calculate the influence weights between nodes means that, through the attention mechanism, the graph neural network can dynamically pay attention to the importance of different nodes and their neighboring nodes in the graph according to the task requirements when processing node information.

[0050] In practical applications, based on the changing trends of influencing weights, the system predicts operational conflict risks within the next 5-10 seconds and adjusts the movement planning of the virtual maintenance personnel's digital model in advance. This means that the system can not only identify current risk points but also predict potential risk development trends based on the dynamic changes in weights.

[0051] This application's solution abstracts complex building environments, virtual maintenance personnel digital models, and virtual tool digital models into a heterogeneous graph structure, enabling the representation of various physical and biomechanical interactions in a unified graph data format. Through this technical solution, this application achieves refined and real-time evaluation of interactions during simulated maintenance operations. Compared to traditional rule-based or simple physical model-based evaluation methods, the use of a heterogeneous graph structure and graph attention mechanism allows the system to more comprehensively and accurately capture complex multimodal and multi-scale interaction information and dynamically identify key risk points. More importantly, by analyzing the changing trends of influence weights, this application can predict and proactively avoid future operational conflict risks, elevating risk management from passive response to proactive prevention. This significantly improves the safety, efficiency, and reliability of simulated maintenance operations, providing a more accurate and forward-looking evaluation basis for subsequent motion strategy optimization.

[0052] In some of the embodiments described above in this application, although adaptive iterative verification of optimization and adjustment suggestions is proposed, how to effectively integrate the dynamic feedback from the actual construction site into the simulation verification process, and intelligently determine when the iteration reaches the optimal state to avoid unnecessary computational overhead, remains a problem that needs further improvement. A lack of real-time perception of the external environment and intelligent control of the iterative process may lead to deviations between the simulation results and the actual situation, or inefficiency in the verification process.

[0053] In response, this application further proposes the aforementioned adaptive iterative verification, including: An external environment feedback loop is set up to receive maintenance operation risk index and actual construction feedback data. The deviation signal output from the external environment feedback loop is input into the value function of reinforcement learning to fine-tune the motion strategy of the virtual maintenance personnel digital model online. Record the risk index decline curve after each iteration. When the slope of the risk index decline curve is lower than the preset threshold, the iteration will automatically stop and the current optimal maintenance plan will be output. At the same time, operation guidance data that can be read by on-site construction personnel wearing devices will be generated.

[0054] Specifically, the external environment feedback loop refers to a system capable of continuously monitoring and collecting data on the actual conditions of the construction site. This loop not only receives maintenance operation risk indices generated by simulated maintenance operations but also acquires actual construction feedback data, such as on-site sensor data, construction worker operation logs, and safety incident reports. This real-world data provides the simulated environment with calibration information from the real world, ensuring the accuracy and practicality of the simulation.

[0055] The deviation signal can be understood as the difference between the simulated predicted risk index and the actual construction feedback data. When this deviation signal is input into the value function of reinforcement learning, it acts as a correction factor, guiding the deep Q-network to fine-tune the motion strategy of the virtual maintenance personnel's digital model online. This fine-tuning mechanism enables the virtual maintenance personnel's motion strategy to adaptively adjust according to changes in the actual construction environment, thereby improving its applicability in the real world.

[0056] In practical applications, to ensure the efficiency and convergence of the iterative verification process, the system continuously records the changes in the maintenance operation risk index after each iteration, forming a risk index decline curve. When the slope of this curve, i.e., the rate of decline of the risk index, falls below a preset threshold, it indicates that the optimization of the risk index has plateaued, and the benefits of further iterations will be very limited. At this point, the system automatically stops iterating and outputs the maintenance solution currently considered optimal. Simultaneously, to facilitate actual operation by on-site construction personnel, it also generates operation guidance data that can be read by wearable devices, such as visual instructions on AR glasses or vibration prompts on smartwatches, directly guiding on-site operations.

[0057] This application's solution effectively solves the problem of the disconnect between simulation and reality in the basic scheme by introducing an external environment feedback loop. Through the above technical solution, this application can significantly improve the accuracy and efficiency of adaptive iterative verification. Specifically, the introduction of the external environment feedback loop allows the simulation results to be calibrated with the actual construction situation in real time, greatly narrowing the gap between the virtual environment and the real world, thereby improving the reliability of optimization and adjustment suggestions. The online fine-tuning mechanism ensures that the movement strategy of the virtual maintenance personnel digital model can continuously adapt to the dynamically changing construction environment, maintaining optimal performance even in the face of emergencies or unforeseen circumstances. More importantly, by intelligently determining when to stop the iteration, unnecessary consumption of computational resources is avoided, the generation speed of the optimal maintenance scheme is accelerated, and operational guidance data that can be read by on-site construction personnel's wearable devices is directly generated, greatly improving the feasibility of the scheme and the convenience and safety of on-site operations.

[0058] This application further proposes a quantitative assessment method for visual accessibility and cognitive load, which aims to identify and address potential visual and cognitive obstacles by quantitatively analyzing the visual conditions of the maintenance environment and the cognitive state of virtual maintenance personnel, in order to generate more comprehensive and safer optimization and adjustment recommendations.

[0059] In some embodiments of this application, a quantitative assessment of visual accessibility and cognitive load is also included: Acquire dynamic light source distribution and shadow areas in a multimodal digital twin environment; The head pose and gaze trajectory of the virtual maintenance personnel digital model are obtained, and the visible surface coverage of the maintenance target component is extracted through a convolutional neural network. When the visible surface coverage is below a preset threshold, the system automatically generates design modification suggestions for auxiliary lighting or adjustable access ports, and iteratively evaluates the modified visible surface coverage.

[0060] Specifically, the quantitative assessment of visual accessibility and cognitive load refers to the quantitative analysis of the extent to which virtual maintenance personnel can clearly observe the target components being maintained when performing maintenance operations in a digital twin environment, as well as the information processing load on their brains. The aim is to ensure that maintenance operations are performed under optimal visual and cognitive conditions, thereby improving the accuracy and safety of the operations.

[0061] The acquisition of dynamic light source distribution and shadow areas in a multimodal digital twin environment can be understood as the system monitoring and recording in real time the intensity, direction, color, and other parameters of all light sources (such as natural light and artificial lighting) in the digital twin model, and calculating the real-time illumination intensity and shadow range of each area based on the interaction between these light sources and the geometric information of the building environment. This helps to accurately simulate the lighting conditions of the actual construction site and provides basic data for subsequent visibility assessment.

[0062] In practical applications, obtaining the head pose and gaze trajectory of a virtual maintenance worker's digital model can be achieved, for example, by tracking the 3D rotation and translation data of the virtual maintenance worker's head, as well as the gaze direction and focus position of its virtual eyes. Head pose reflects the virtual maintenance worker's observation direction, while the gaze trajectory more precisely indicates the specific area of ​​focus. These data are key inputs for assessing their visual focus and cognitive load.

[0063] Furthermore, the visible surface coverage of the maintenance target component is extracted using a convolutional neural network. Specifically, this involves taking an image rendered from the perspective of a virtual maintenance personnel in a digital twin environment as input and analyzing the image using a pre-trained convolutional neural network. This network can identify and segment the visible portion of the maintenance target component from the current viewpoint and calculate its proportion in the total surface area, i.e., the visible surface coverage. This coverage directly reflects the clarity of the maintenance personnel's observation of the target component.

[0064] In a preferred implementation, when the visible surface coverage falls below a preset threshold, the system automatically generates design modification suggestions for auxiliary lighting or adjustable access ports. For example, if insufficient light in an area results in low visible surface coverage, the system might suggest adding a virtual auxiliary light source, specifying its location and intensity; if the target component being maintained is obstructed by a structure, the system might suggest adjusting or enlarging the design of the access port. These suggestions aim to directly improve visual conditions. Subsequently, the system updates the digital twin environment based on these modification suggestions and iteratively evaluates the modified visible surface coverage again until a preset visual accessibility standard is reached.

[0065] This application's solution effectively compensates for the shortcomings of basic solutions in considering the comprehensiveness of maintenance operations by introducing a quantitative assessment of visual accessibility and cognitive load. While the basic solution primarily focuses on minimizing physical interaction and biomechanical risks, this solution further incorporates visual and cognitive factors. Through these technical solutions, this application can significantly improve the realism and safety of simulated maintenance operations. First, by quantitatively assessing visual accessibility, the system can proactively identify and resolve potential risks caused by insufficient lighting or obstructed vision, avoiding the "can't see clearly" problem that may occur in actual operations. Second, by analyzing head posture and gaze trajectory, combined with visible surface coverage, the cognitive load of virtual maintenance personnel can be indirectly assessed, thereby optimizing the operating environment and reducing errors caused by excessive cognitive stress. Furthermore, automatically generating design modification suggestions for auxiliary lighting or adjustable access panels and conducting iterative evaluations allows this method to provide forward-looking design optimization solutions, optimizing not only the maintenance operation itself but also the maintainability design of the building structure and facilities. This proactively addresses visual and cognitive obstacles that may be encountered during the operation and maintenance phase during the construction stage, greatly improving overall construction efficiency and operation and maintenance benefits.

[0066] This application further proposes a method for conflict prediction and avoidance in multi-person collaborative maintenance, specifically including: Each virtual maintenance worker digital model is assigned a deep reinforcement learning agent, and each agent has an independent operation task priority and motion policy network. Multiple agents share the global digital twin environment state through a framework of centralized training and distributed execution, and learn mutual avoidance and cooperative behaviors using counterfactual multi-agent policy gradients; Predict the spatial envelope conflict time points between agents in real time, and adjust the operation sequence in a bidding manner to generate a conflict-free operation sequence diagram.

[0067] Specifically, assigning a deep reinforcement learning agent to each virtual maintenance worker digital model means treating each virtual maintenance worker as an independent decision-making unit. This unit is trained using deep reinforcement learning algorithms to learn the optimal movement strategy. Each agent is assigned independent task priorities; for example, some tasks may have higher urgency or criticality, thus influencing the agent's decision-making order and resource allocation. Simultaneously, each agent also possesses an independent movement policy network, responsible for generating specific action instructions, such as movement, grasping, or tool manipulation, based on the current environmental state and task priorities.

[0068] In this framework, multiple agents share the global state of the digital twin environment through centralized training and distributed execution. This means that during the training phase, the policy networks of all agents are jointly trained on a centralized server or computing cluster. During this process, agents can access and share global state information of the entire multimodal digital twin environment, including the positions of other agents, task progress, tool states, and dynamic changes in the building environment. This centralized training helps agents learn globally optimal collaborative behaviors. However, during the actual execution of simulated maintenance operations, each agent independently makes decisions and executes actions based on its trained policy network—that is, distributed execution.

[0069] Furthermore, using counterfactual multi-agent policy gradient learning to learn mutual avoidance and cooperative behavior means that during training, agents not only consider the impact of their own behavior on rewards, but also the different outcomes their own behavior might produce if other agents take different actions. Through this counterfactual reasoning, agents can better understand and predict the behavior of other agents, thereby learning more effective mutual avoidance, resource sharing, and task cooperation strategies to avoid potential conflicts and improve overall collaborative efficiency.

[0070] The proposed solution transforms each virtual maintenance worker's digital model into an entity with autonomous decision-making capabilities by introducing deep reinforcement learning agents. These agents, trained within a centralized framework, can share the global digital twin environment state and learn complex behaviors of mutual avoidance and cooperation in a multi-agent environment using a counterfactual multi-agent policy gradient algorithm.

[0071] In response, this application further proposes a scheme to dynamically update the Bayesian risk threshold in the following manner, in order to achieve more adaptive and robust risk management.

[0072] The Bayesian risk threshold is dynamically updated in the following ways: Collect actual safety accident data, operational error reports, and environmental monitoring data from construction sites; Construct a Bayesian hierarchical model, using project type, personnel experience level, time period, and weather conditions as covariates; Each time a simulation maintenance iteration is completed, the posterior distribution of the Bayesian hierarchical model is updated, and the Bayesian risk threshold moves accordingly to the high confidence interval. The updated Bayesian risk threshold is fed back to the reward function of the deep Q-network in real time, forming an online learning loop for risk perception.

[0073] Specifically, collecting actual safety accident data, operational error reports, and environmental monitoring data from construction sites refers to the system continuously acquiring multi-source, multi-dimensional data from the real construction environment. Actual safety accident data can include detailed records such as accident type, time of occurrence, location, personnel involved, direct causes, and indirect causes; operational error reports cover operational mistakes that did not result in serious consequences but reflect potential risks; environmental monitoring data may include physical environmental parameters affecting construction safety, such as temperature, humidity, wind speed, and light intensity. The purpose of collecting this data is to provide comprehensive, real-time input for the training and updating of the risk model.

[0074] The construction of a Bayesian hierarchical model, incorporating project type, personnel experience level, time period, and weather conditions as covariates, employs a statistical modeling method. This model can handle data structures at different levels and allows information sharing between these levels, thus enabling more accurate parameter estimation. Project type distinguishes the inherent risk characteristics of different construction projects; personnel experience level quantifies the operator's proficiency and potential error rate; time period reflects the risk characteristics of different construction stages or time windows, such as the risks of nighttime construction; and weather conditions directly affect the safety of the working environment and the difficulty of operation. Incorporating these covariates into the model ensures that the calculation of risk thresholds fully considers the differences in specific situations.

[0075] In practical applications, after each simulated maintenance iteration, the posterior distribution of the Bayesian hierarchical model is updated, and the Bayesian risk threshold correspondingly shifts towards a higher confidence interval. This means that with each simulated maintenance operation, the system re-estimates the parameters in the Bayesian hierarchical model based on simulation results, such as the risk index exhibited by the virtual maintenance personnel digital model in a specific operation, and newly collected actual data. The update of the posterior distribution reflects the model's continuous adjustment and refinement of its understanding of risk parameters as new evidence emerges. When the risk assessment results tend to stabilize or the risk index continues to decrease, the Bayesian risk threshold shifts towards a higher confidence interval, indicating that the system's confidence in the current risk assessment has increased, thus enabling the risk threshold to more accurately reflect the true risk level in the current situation.

[0076] Furthermore, the updated Bayesian risk threshold is fed back into the deep Q-network's reward function in real time, forming an online learning loop for risk perception. This means that the dynamically updated, more context-adaptive risk threshold directly affects the deep Q-network's reward calculation when optimizing the movement strategies of the virtual maintenance personnel digital model. When the risk index of the simulated operation is lower than this dynamically updated threshold, the reward is higher; conversely, it is penalized. This feedback mechanism enables the deep Q-network to "perceive" the latest risk context and adjust its learning objectives and strategies accordingly. This prompts the virtual maintenance personnel digital model to learn safer operational behaviors that better match the current actual risk level, forming a continuously optimizing, self-adaptive risk management and strategy learning process.

[0077] This application's solution effectively addresses the limitations of traditional static thresholds in handling complex and ever-changing construction environments by introducing a dynamic update mechanism for Bayesian risk thresholds. Through this technical solution, this application enables dynamic and contextualized management of the Bayesian risk threshold for maintenance operation risk indices. Compared to static thresholds trained on single historical data, this application's solution significantly improves the accuracy and real-time performance of risk assessment, allowing it to better adapt to the complex and ever-changing environment, personnel, and project conditions at construction sites. This dynamic update mechanism enables the system to identify potential risks earlier and more accurately, and prompts the virtual maintenance personnel digital model to learn safer operational strategies that better reflect the current actual risk level. Therefore, it not only effectively reduces the accident rate and operational errors in actual construction but also improves the overall safety and efficiency of the construction process, providing a more robust and reliable risk warning and avoidance capability for intelligent control and pre-maintenance management throughout the entire construction process.

[0078] While some embodiments of this application propose generating structured optimization and adjustment suggestions that include quantitative estimates of construction efficiency and operation and maintenance benefits, in practical applications, these quantitative estimates may remain at a high level, lacking sufficient detail and operability. They are difficult to directly parse and utilize by advanced technology platforms such as Building Information Modeling (BIM) systems, construction robots, or augmented reality devices, and also struggle to fully support complex decisions involving multi-objective trade-offs. Without addressing these issues, the generated optimization suggestions may fail to fully realize their potential in intelligent construction and operation and maintenance, resulting in inefficient information transmission and difficulty in achieving seamless integration from simulation to actual execution.

[0079] In this regard, this application further proposes steps for generating the above-mentioned structured optimization and adjustment suggestions, including: The output is an enhanced parameter set that can be directly parsed by the building information model. The enhanced parameter set includes modified geometric dimensions, material recommendations, reserved operation space coordinates, and suggested maintenance tool types. Simultaneously, a construction efficiency gain curve and a full lifecycle operation and maintenance cost change graph are generated to support multi-objective optimization decisions; The enhanced parameter set, construction efficiency gain curve, and full life cycle operation and maintenance cost change chart are packaged into an executable decision package, which can be directly called by robots or augmented reality devices on the construction site.

[0080] Specifically, the enhanced parameter set refers to a set of optimized and adjusted data that can be directly used to update the Building Information Model (BIM). Modified geometric dimensions can be understood as dimensional adjustment suggestions for the target component or its surrounding structure, such as enlarging access panels or adjusting component installation positions to facilitate maintenance operations. Material recommendations refer to optimal material selection suggestions based on the maintenance environment, component characteristics, and optimization objectives (such as durability and cost). Reserved operating space coordinates refer to the minimum space range and precise location explicitly defined and suggested for operation by maintenance personnel or robots in the 3D digital twin environment. Recommended maintenance tool types refer to the specific types or models of tools recommended based on the nature of the maintenance task and the characteristics of the target component. These parameters are designed to be directly parsed by the BIM, meaning they conform to BIM data standards and interface specifications, enabling automated import and updates, thereby avoiding errors and efficiency losses caused by manual intervention.

[0081] The construction efficiency gain curve can be understood as a trend chart showing the improvement in indicators such as expected construction or maintenance task completion time and resource consumption relative to the baseline solution over time or iterations after adopting optimization suggestions. The life-cycle operation and maintenance cost change chart quantitatively displays the impact of the optimization solution on the total cost of the entire building or component's lifecycle, considering costs at all stages including initial investment, operating energy consumption, periodic maintenance, component replacement, and disposal. These two types of charts aim to provide intuitive and comprehensive economic and efficiency assessments, thereby supporting decision-makers in balancing multiple conflicting objectives, such as improving efficiency while simultaneously reducing long-term operation and maintenance costs.

[0082] In practical applications, the executable decision package is a data set that integrates and encapsulates the aforementioned set of enhanced parameters, construction efficiency gain curves, and lifecycle maintenance cost variation diagrams. This decision package is designed in a standardized format that can be directly invoked by external devices (such as robots or augmented reality devices on the construction site), such as JSON, XML, or other specific API interface data. Its purpose is to achieve seamless transmission and execution of optimization suggestions from the digital twin environment to the physical construction site, greatly simplifying the information conversion and command issuance process.

[0083] This application addresses the limitations of basic solutions where optimization suggestions are often too abstract and difficult to apply directly to actual construction and operation by refining optimization recommendations into a set of enhanced parameters that can be directly parsed by the Building Information Model (BIM). This is supplemented by quantified construction efficiency gain curves and full lifecycle operation and maintenance cost change graphs. Through this technical solution, this application provides more specific, actionable, and easily integrated optimization recommendations. This significantly improves the practicality and implementability of the optimization recommendations, enabling them to directly drive BIM updates and automated operation of on-site equipment. Furthermore, by providing multi-dimensional quantitative assessments (such as construction efficiency gain curves and full lifecycle operation and maintenance cost change graphs), it greatly enhances the comprehensiveness and scientific rigor of decision support, allowing decision-makers to make better choices in complex trade-offs. Therefore, this application effectively compensates for the shortcomings of basic solutions in terms of the refinement and executability of optimization recommendations, achieving a seamless transition from intelligent analysis to intelligent execution, thereby significantly improving the overall efficiency, safety, and economic benefits of building construction and operation and maintenance.

[0084] Specifically, after generating structured optimization and adjustment suggestions, the above method packages these suggestions into an executable decision package. To ensure that this executable decision package can be effectively applied to actual construction and operation and maintenance scenarios, this application further clarifies its specific transmission path and receiving equipment.

[0085] The executable decision package is transmitted to at least one of the following devices: Path planning module for construction robot controller; Spatial anchor point positioning module for augmented reality headsets; Version update module for the Building Information Modeling (BIM) collaborative platform.

[0086] The path planning module of a construction robot controller refers to a control system integrated inside or outside the construction robot. Its main function is to receive and parse the optimized operation instructions, path information, and tool usage strategies contained in the executable decision package, and transform them into motion planning instructions that the robot can recognize and execute, so as to guide the construction robot to perform precise and efficient operations on site. The path planning module can generate the optimal collision-free path based on the three-dimensional spatial coordinates and time series data in the decision package.

[0087] Furthermore, the spatial anchor point positioning module of the augmented reality (AR) headset can be understood as a high-precision positioning and tracking system integrated into the AR headset. Its purpose is to receive data from the executable decision package regarding virtual maintenance personnel operation instructions, the precise location of maintenance target components, and risk area warnings, and to accurately overlay this virtual information onto the real construction environment using spatial anchor point technology. In this way, on-site operators can obtain intuitive, real-time visual guidance, thereby improving the accuracy and safety of operations.

[0088] In practical applications, the version update module of a Building Information Modeling (BIM) collaboration platform is a core functional component of the BIM management system. It can be an API interface or a data synchronization service, designed to receive enhanced parameter sets from executable decision packages, such as modified geometric dimensions, material recommendations, reserved operating space coordinates, and suggested maintenance tool types, and seamlessly integrate them into the existing BIM model. This allows the BIM model to reflect the latest optimized design and maintenance plans in real time, ensuring that all project stakeholders always make collaborative decisions and manage based on the latest and most accurate information, avoiding design conflicts or construction errors caused by information lag.

[0089] This application's solution achieves seamless integration between simulation optimization results and actual construction site conditions by transmitting the executable decision package to the aforementioned specific equipment. Through this technical solution, this application ensures that maintenance plans optimized through simulation and risk assessment can be efficiently and accurately applied to actual construction and operation scenarios. This not only significantly improves the practicality and operability of optimization and adjustment suggestions but also significantly reduces the complexity and potential risks of on-site operations through automated equipment and augmented reality technology. Simultaneously, it promotes dynamic updates of the BIM model and multi-party collaboration, thereby comprehensively improving the intelligent management and control level of the entire construction process and the efficiency of proactive operation and maintenance.

[0090] Construction projects often encounter problems such as constantly changing construction environments, low efficiency in coordination between multiple processes, frequent conflicts in resource allocation, and untimely response to risks and hazards. Traditional management methods often rely on human experience to make decisions, making it difficult to achieve globally optimal configuration. To address these challenges, an AI-based intelligent control and operation-maintenance system for the entire construction process has been introduced, aiming to improve project management efficiency, optimize resource allocation, and reduce risks. However, in practical applications, especially when facing sudden changes in geological conditions and subsequent design adjustments, how to ensure both short-term construction efficiency and long-term operational convenience has become a pressing problem to be solved.

[0091] refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of an AI-based intelligent management and operation-maintenance system for the entire construction process, provided in an embodiment of the present invention, including: The digital twin building module is used to construct a multimodal digital twin environment, which includes geometric information of the building environment, material properties, light source distribution, and maintainability annotation data of the target components. An artificial intelligence decision engine, comprising a reinforcement learning framework and a deep Q-network, is used to dynamically generate digital models of virtual maintenance personnel and virtual tools, and drive them to perform iterative simulated maintenance operations in a multimodal digital twin environment. The graph neural network evaluation module is used to evaluate the interaction between virtual maintenance personnel and virtual tools and the building environment in real time, and output the maintenance operation risk index. A risk adaptive threshold updater is used to dynamically adjust the risk threshold based on a Bayesian hierarchical model and feed it back to the artificial intelligence decision engine to form an online learning closed loop. The optimization suggestion generation module is used to generate structured optimization and adjustment suggestions and output an executable decision package to an external device.

[0092] This system, through its modular design, organically combines the construction of a digital twin environment, the execution of AI decisions, the evaluation of interactions, the dynamic updating of risk thresholds, and the generation and output of optimization suggestions. Its core lies in achieving intelligent control and proactive optimization of the entire construction and operation process through the collaborative work of its modules. The digital twin construction module provides a high-fidelity virtual operating environment for the entire system; the AI ​​decision engine drives virtual entities to simulate operations and optimize their strategies within this environment; the graph neural network evaluation module quantifies operational risks in real time; the risk adaptive threshold updater ensures the dynamism and accuracy of risk assessment; and finally, the optimization suggestion generation module transforms the optimization results into actionable suggestions and outputs them directly to on-site equipment. This effectively solves problems such as inefficiency, poor resource allocation, and untimely risk response in traditional management methods. It also deeply integrates factors such as "future operational feasibility," "long-term maintenance costs," and "accessibility throughout the equipment's lifecycle," achieving a balance between short-term construction efficiency and long-term operational benefits, avoiding a situation of "efficient construction, inefficient maintenance."

[0093] Compared with existing technologies, the core innovation of the AI-based intelligent management and operation-maintenance system for the entire construction process proposed in this application lies in its modular system architecture. This architecture deeply integrates factors such as "future operational feasibility," "long-term maintenance costs," and "accessibility throughout the equipment's lifecycle" into the scheme optimization decision-making process during the construction phase through dynamic simulation and AI optimization. Traditional AI management systems primarily focus on static geometric conflict detection and construction efficiency optimization, failing to fully assess the impact of construction schemes on the actual operating space of future maintenance personnel, the ease of tool use, and the obstruction of maintenance access routes by large equipment installations.

[0094] This system provides a high-fidelity environment through a digital twin construction module, an artificial intelligence decision engine drives virtual maintenance personnel and tools to simulate behavior, a graph neural network assessment module quantifies interactive risks in real time, and a risk adaptive threshold updater dynamically adjusts risk thresholds, forming a complete, closed-loop intelligent decision-making system. For example, in pipeline rerouting construction, this system can dynamically identify and quantify potential maintenance risks, such as whether a certain inspection port is difficult to reach or whether joint stress is excessive during operation, by simulating the actual operation of virtual maintenance personnel. The optimization suggestion generation module can then transform these risks into specific, actionable optimization suggestions and output them directly to the field equipment. This proactive, behavior simulation-based risk assessment and optimization enables the system to effectively avoid the situation of "efficient construction, inefficient maintenance," achieving a balance between short-term construction efficiency and long-term operation and maintenance benefits, and significantly improving the intelligent management and pre-maintenance capabilities and practical value of the entire construction process.

[0095] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent management and pre-operation maintenance of the entire construction process based on AI, characterized in that, include: Construct a multimodal digital twin environment, which includes geometric information of the building environment, material properties, light source distribution, and maintainability annotation data of the target components; Using a reinforcement learning framework, a digital model of a virtual maintenance worker and a digital model of a virtual tool that matches the target component are dynamically generated based on the type of the component being maintained. The digital model of the virtual maintenance worker includes musculoskeletal parameters, joint degree-of-freedom constraints, and a biomechanical response model. In a multimodal digital twin environment, a physically based rendering engine and collision detection model are used to drive a virtual maintenance personnel digital model carrying a virtual tool digital model to perform iterative simulated maintenance operations on the target component. During the simulated maintenance operation, a graph neural network is used to evaluate in real time the interaction between the digital model of the virtual maintenance personnel, the digital model of the virtual tool, and the building environment. The interaction includes the physical contact force distribution based on finite element analysis, the operation posture comfort score based on a biomechanical database, and the cumulative joint stress value based on ergonomics. The interaction data is input into the reward function, and the motion strategy of the virtual maintenance personnel digital model is optimized online through a deep Q-network to minimize the maintenance operation risk index identified in the interaction data. The system generates structured optimization and adjustment suggestions, which include quantitative estimates of construction efficiency and operation and maintenance benefits. It also supports adaptive iterative verification of the optimization and adjustment suggestions until the maintenance operation risk index is lower than the Bayesian risk threshold trained from historical accident data.

2. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 1, is characterized in that... The operational posture comfort score based on the biomechanical database was obtained through the following methods: Obtain the joint angle sequence and muscle activation of the digital model of the virtual maintenance personnel; Joint angle sequences and muscle activation are input into a pre-trained long short-term memory network, which is used to learn the non-linear mapping between human posture and subjective comfort. The output includes heat maps of joint stress distribution and predicted fatigue accumulation values ​​for each maintenance operation stage; Based on the joint stress distribution heatmap and fatigue accumulation prediction, and combined with the operating frequency requirements of the target components, an interpretable long-term occupational health risk warning is generated.

3. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 1, is characterized in that... Real-time evaluation of interactions using graph neural networks, including: The building environment, virtual maintenance personnel digital model and virtual tool digital model are constructed as a heterogeneous graph structure. The nodes of the heterogeneous graph structure include structural entities, human joints and tool contact surfaces. The edges of the heterogeneous graph structure represent spatial constraints, kinematic chain relationships and force transmission paths. A graph attention mechanism is used to calculate the influence weights between nodes, and to dynamically identify critical high-stress paths and operational constraint chains. Based on the changing trends of the influencing weights, predict the risk of operational conflicts within the next 5-10 seconds and adjust the motion planning of the virtual maintenance personnel digital model in advance.

4. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 1, is characterized in that... Adaptive iterative verification includes: An external environment feedback loop is set up to receive maintenance operation risk index and actual construction feedback data. The deviation signal output from the external environment feedback loop is input into the value function of reinforcement learning to fine-tune the motion strategy of the virtual maintenance personnel digital model online. Record the risk index decline curve after each iteration. When the slope of the risk index decline curve is lower than the preset threshold, the iteration will automatically stop and the current optimal maintenance plan will be output. At the same time, operation guidance data that can be read by on-site construction personnel wearing devices will be generated.

5. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 1, is characterized in that... It also includes a quantitative assessment of visual accessibility and cognitive load: Acquire dynamic light source distribution and shadow regions in a multimodal digital twin environment; The head pose and gaze trajectory of the virtual maintenance personnel digital model are obtained, and the visible surface coverage of the maintenance target component is extracted through a convolutional neural network. When the visible surface coverage is below a preset threshold, the system automatically generates design modification suggestions for auxiliary lighting or adjustable access ports, and iteratively evaluates the modified visible surface coverage.

6. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 1, is characterized in that... It also includes conflict prediction and avoidance in multi-person collaborative maintenance: Each virtual maintenance worker digital model is assigned a deep reinforcement learning agent, and each agent has an independent operation task priority and motion policy network. Multiple agents share the global digital twin environment state through a framework of centralized training and distributed execution, and learn mutual avoidance and cooperative behavior using counterfactual multi-agent policy gradients; Predict the spatial envelope conflict time points between agents in real time, and adjust the operation sequence in a bidding manner to generate a conflict-free operation sequence diagram.

7. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 1, is characterized in that... The Bayesian risk threshold is dynamically updated in the following ways: Collect actual safety accident data, operational error reports, and environmental monitoring data from construction sites; Construct a Bayesian hierarchical model, using project type, personnel experience level, time period, and weather conditions as covariates; Each time a simulation maintenance iteration is completed, the posterior distribution of the Bayesian hierarchical model is updated, and the Bayesian risk threshold moves accordingly to the high confidence interval. The updated Bayesian risk threshold is fed back to the reward function of the deep Q-network in real time, forming an online learning loop for risk perception.

8. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 1, is characterized in that... The generation of structured optimization and adjustment suggestions also includes: The output is an enhanced parameter set that can be directly parsed by the building information model. The enhanced parameter set includes modified geometric dimensions, material recommendations, reserved operation space coordinates, and suggested maintenance tool types. Simultaneously, a construction efficiency gain curve and a full life-cycle operation and maintenance cost change graph are generated to support multi-objective optimization decisions; The enhanced parameter set, construction efficiency gain curve, and full life cycle operation and maintenance cost change chart are packaged into an executable decision package, which can be directly called by robots or augmented reality devices on the construction site.

9. The method for intelligent control and pre-operation maintenance of the entire construction process based on AI, as described in claim 8, is characterized in that... The executable decision package is transmitted to at least one of the following devices: Path planning module for construction robot controller; Spatial anchor point positioning module for augmented reality headsets; Version update module for the Building Information Modeling (BIM) collaborative platform.

10. A system for intelligent management and pre-operation maintenance of the entire construction process based on AI, used to execute the method for intelligent management and pre-operation maintenance of the entire construction process based on AI as described in claim 1, characterized in that, include: The digital twin building module is used to construct a multimodal digital twin environment, which includes geometric information of the building environment, material properties, light source distribution, and maintainability annotation data of the target components. An artificial intelligence decision engine, comprising a reinforcement learning framework and a deep Q-network, is used to dynamically generate digital models of virtual maintenance personnel and virtual tools, and drive them to perform iterative simulated maintenance operations in a multimodal digital twin environment. The graph neural network evaluation module is used to evaluate the interaction between virtual maintenance personnel and virtual tools and the building environment in real time, and output the maintenance operation risk index. A risk adaptive threshold updater is used to dynamically adjust the risk threshold based on a Bayesian hierarchical model and feed it back to the artificial intelligence decision engine to form an online learning closed loop. The optimization suggestion generation module is used to generate structured optimization and adjustment suggestions and output an executable decision package to an external device.

Images