A system of building vertical domain large model and digital twin intelligent agent

Abstract

The application provides a kind of building vertical domain large model and digital twin intelligent agent system, it is related to building industrialization and intelligent technology field. Including perception module, digital twin module, building vertical domain large model module and execution control module. Perception module collects the multi-source heterogeneous data of building entity;Digital twin module is based on unified space-time datum fusion data, constructs synchronous mapping dynamic digital twin and outputs multi-modal state information;Building vertical domain large model module is based on building field pre-training, carries out multi-modal understanding and expert reasoning to state information, generates optimization decision instruction;Execution control module converts instruction into control signal and issues execution, simultaneously collects feedback data for updating digital twin state and large model parameter. The application realizes the fusion of building full life cycle data and intelligent decision-making closed loop, and improves the overall intelligent level and self-adaptive optimization ability of building industry.

Description

Technical Field

[0001] This invention relates to the field of building industrialization and intelligent technology, specifically a system for a large-scale building vertical model and a digital twin intelligent agent. Background Technology

[0002] Currently, various information technologies have been partially applied in the digital transformation of the construction industry. For example, building information modeling (BIM) technology provides 3D visualization and information management tools for the design phase; Internet of Things (IoT) sensor networks enable real-time monitoring of building structures, equipment, and the environment; these technologies together constitute a digital infrastructure platform for certain stages of the building's entire lifecycle.

[0003] However, existing technologies still have significant limitations in pursuing higher levels of intelligence and global optimization. First, because data generated at different stages differs in format, standards, and spatiotemporal references, and is managed independently by different systems, the complete building lifecycle data remains fragmented, making effective integration and interoperability within a unified framework difficult. This hinders cross-stage and cross-professional collaborative analysis and decision-making. Second, precisely because of these data-level barriers, the intelligent capabilities of existing systems are mostly limited to processing limited structured data using pre-defined rule bases. They lack the ability to deeply understand and integrate massive amounts of multi-source, heterogeneous, and unstructured information such as drawings, text logs, and point clouds. Consequently, they struggle to handle decision-making tasks requiring creative optimization and dynamic adjustments based on complex domain knowledge, thus limiting both the depth and breadth of their intelligence. Summary of the Invention

[0004] To achieve the above objectives, this invention proposes a system of a large building vertical model and a digital twin intelligent agent, including a perception module, a digital twin module, a large building vertical model module, and an execution control module; The perception module acquires real-time operational data through sensors deployed on the physical building, obtains geometric and attribute data through BIM modeling, and acquires point cloud data through 3D laser scanning, outputting multi-source heterogeneous building lifecycle data; The digital twin module receives building lifecycle data and fuses it based on a unified spatiotemporal reference to construct a dynamic digital twin that is synchronously mapped to the building's physical entity. The digital twin integrates a visualization engine and a simulation engine, which are used to perform simulation and deduction based on candidate states generated from the bound data and determined to be valid states by the event consistency rules. It provides output in the form of multimodal state information with time-series labels and component semantic labels, based on the valid states. Candidate states that fail the determination are not used as the source of the multimodal state information output until they are confirmed as valid states. The building vertical domain large model module receives multimodal state information, is based on the Transformer architecture, and is pre-trained using building domain corpus. It has the ability to understand and reason about BIM views, construction log text, and point cloud data in a multimodal manner. By parsing multimodal state information and combining it with the building knowledge base for expert-level reasoning, it generates decision instructions that include production scheduling, construction optimization, and energy consumption control. The execution control module receives decision instructions, converts them into driveable control signals, and issues them for execution. After execution, the execution control module collects execution result data according to the evidence item format defined in the specification and sends it back to the digital twin module and the building vertical domain large model module. The digital twin module constructs candidate states based on the returned execution result data and submits them to the event consistency rules for adjudication. Only candidate states that are adjudicated as valid states are used to replace the current valid state of the digital twin. The building vertical domain large model module only uses the valid state as the input evidence set for near-end strategy optimization or other parameter updates when it receives a valid state confirmed by the digital twin module.

[0005] As a further technical solution, the building lifecycle data acquired by the sensing module undergoes data standardization processing before being input into the digital twin module; this processing includes applying formulas to real-time operational data with different physical dimensions collected by sensors, such as temperature, pressure, and displacement. Normalization calculations are performed to transform the data to a uniform numerical range of 0 to 1 or -1 to 1, in order to eliminate the influence of dimensions and accelerate the convergence of subsequent models; for energy consumption monitoring data, the formula is used. Standardization calculations are performed, where μ is the mean of historical energy consumption data and σ is its standard deviation, so that the processed data has a mean of 0 and a standard deviation of 1. Next, feature extraction and dimensionality reduction are performed. For high-dimensional data, including point clouds and images, pre-trained convolutional neural networks or point cloud feature networks are used to extract features and obtain their high-level semantic feature vectors. Then, principal component analysis is used for dimensionality reduction. During the dimensionality reduction process, a variance contribution rate threshold is set to retain the feature dimensions that can explain more than half of the variance information. This reduces the subsequent computational complexity and effectively improves the training efficiency and generalization ability of the building vertical domain large model module.

[0006] As a further technical solution, the process of constructing a dynamic digital twin module includes: establishing a unified spatiotemporal coordinate system based on the origin of the building BIM model or geodetic coordinates; using the geometric and attribute data obtained by the perception module through BIM modeling as basic static data, determining the spatial position, geometric dimensions, material properties, and subordinate relationships of each building component in the coordinate system to form a static skeleton; uniquely binding the operational data obtained by sensors to the corresponding building components in the basic static data according to the sensor's installation position information on the physical entity; spatially aligning the acquired point cloud data with the basic static data and associating the aligned point cloud clusters with the corresponding building components; uniformly receiving and caching the bound data, and sorting and organizing the data according to a preset receiving order rule; For operational data with a sampling frequency lower than that required by the receiving order rule, equally spaced data points are generated to construct candidate states and supplement the evidence items required for the candidate states. For data packets with disordered transmission order, their order is rearranged using the precise timestamps they carry as evidence items and included in the evidence set of the candidate states. The system forms candidate states using the evidence set and submits them to the event consistency rule for adjudication. Only when the candidate state is adjudicated according to the event consistency rule is the candidate state confirmed as the current established state of the digital twin and used to drive the visualization representation and physical attribute parameter update of the corresponding component in the basic static data. Candidate states that fail the adjudication are not used to drive the visualization representation and physical attribute parameter update and are recorded in the event audit record in an immutable manner for tracing and handling.

[0007] As a further technical solution, when any candidate state fails the event consistency rule, the digital twin automatically enters a suspended existence state. In this suspended existence state, the digital twin stops outputting any information deemed as currently valid, ceases to serve as input for simulation, decision generation, model parameter updates, and execution control, and writes the failed candidate states and all their original event evidence into the event audit log in an immutable manner. The digital twin exits the suspended existence state and resumes valid existence only when any of the following conditions are met: subsequently generated candidate states pass the event consistency rule; or the candidate states reach a pre-set evidence convergence condition; and when N consecutive candidate state failures reach a preset threshold, the system triggers an abnormal handling process, including implementing isolation protection measures for affected components or subsystems and recording unrecoverable event logs to prevent components or subsystems from misleading the overall system's decision-making. As a further technical solution, the candidate state refers to a component or system state description formed by combining basic static data and bound operational data in a predetermined receiving order. The candidate state is generated based on the currently received multi-source dataset, but in this invention, it does not automatically become the current state of the digital twin. The event consistency rules and state adjudication mechanism refer to a set of rules used to determine whether a candidate state exists. This set of rules includes at least: continuity judgment (i.e., key events / data logically form a non-conflicting sequence), integrity judgment (i.e., the candidate state contains the minimum set of event evidence required by the component), and component constraint consistency judgment (i.e., the component attributes in the candidate state do not conflict with the state skeleton or engineering constraints of the BIM static digital twin module updated based on feedback data). This invention determines whether a candidate state is qualified to be established by comparing it item by item with the above rules and making a pass / fail decision. The established state, i.e., the current established state, refers to a candidate state that has passed the event consistency rule adjudication. Only the established state is used as the "current state" of the digital twin to drive simulation, deduction, decision input, and model parameter updates. The determination result of the established state is the sole determining factor for the existence of the digital twin. The "pause" state occurs when a candidate state fails to pass the state adjudication mechanism, at which point the digital twin enters a "pause" state. During this pause, the digital twin ceases to output any information deemed valid as the current state and stops serving as an input source for simulation and decision-making. All original event evidence and adjudication results related to the candidate state are written into the event audit log for tracing and handling. The event audit log and evidence aggregation conditions are defined as follows: the event audit log is an append-only record of the adjudication process and its original evidence, an immutable append-only log structure used for subsequent manual or procedural verification. The evidence aggregation conditions are pre-set conditions that allow a previously unadjudicated candidate state to gain validity when sufficient evidence is accumulated or key event evidence is supplemented. The "exit from pause" and "abnormal handling" state occurs when the digital twin exits the pause and resumes valid existence if either of the following conditions is met: subsequently generated candidate states pass the event consistency rule adjudication; or the candidate state or its evidence meets the pre-set evidence aggregation conditions. If the candidate state determination fails N times consecutively and reaches a preset threshold, the system can trigger an exception handling process, such as implementing isolation protection for relevant components or subsystems and recording unrecoverable event logs.

[0008] As a further technical solution, the large-scale building domain model module, after pre-training based on a corpus of the building domain, is optimized through a supervised fine-tuning stage to adapt to specific building tasks. In this stage, a large amount of manually labeled building task data is used for supervised training of the large-scale building domain model. This data includes paired multimodal state information as input and expert decision labels as expected output. The large-scale domain model calculates the predicted probability distribution for various decision instructions using forward computation and employs the cross-entropy loss function. Calculate the error between the output probability distribution of the large-scale building vertical model and the true label, where y i p represents the 0 or 1 value of the i-th dimension in the one-hot encoded vector of the true label. i The probability value of the decision instruction belonging to the i-th class is predicted for the large-scale building vertical model. The gradient of the loss function with respect to the parameters of each layer of the large-scale vertical model is calculated through the backpropagation algorithm, and the cross-entropy loss function is minimized using stochastic gradient descent, thereby adjusting the model parameters so that the large-scale vertical model learns to accurately map from multimodal state information to expert-level decision instructions.

[0009] As a further technical solution, after supervised fine-tuning, the building vertical domain large model module further optimizes its policies through reinforcement learning based on human feedback to improve its decision quality and acceptability. This process introduces a reward model, which is trained based on human experts' ratings of a large number of decision instruction samples in dimensions such as safety, economy, and feasibility. During the reinforcement learning phase, the decision instructions generated by the vertical domain large model module are submitted to the reward model for quality scoring. This quality score is used as the reward signal in reinforcement learning, and a proximal policy optimization algorithm is adopted to maximize the expected reward function. Multiple rounds of iterative training are performed to target the objective; where π θ The strategy of the vertical domain large model module under the current parameter θ represents the probability distribution of the generated action, i.e., decision instruction a, given state s; r(s,a) represents the scalar reward value given by the reward model for state s and the generated decision instruction a; the strategy parameter θ is updated by the proximal policy optimization algorithm, and the step size of each update is constrained to ensure training stability, so that the decision instruction finally generated by the vertical domain large model is more in line with the value judgment and domain experience of human experts.

[0010] As a further technical solution, the building vertical domain large model module performs multi-objective optimization calculations to balance multiple conflicting objectives when generating the final decision instruction; it constructs an objective function with the objective of minimizing the comprehensive cost consisting of energy consumption, time, and economic costs. Where E represents the predicted energy consumption based on digital twin simulation, T represents the time cost of the task, C represents the economic cost involving materials, manpower, and equipment, and w1, w2, and w3 are weight coefficients dynamically configured according to different project stages such as the initial construction period, peak period, and final period, and different management strategies such as cost priority, schedule priority, and green energy-saving priority, with the sum of the weights being 1; the vertical domain large model module solves the Pareto optimal solution set of the objective function under multiple constraints such as resource constraints, physical law constraints, and safety specification constraints, that is, it finds a set of solutions in which the improvement of any objective will inevitably lead to the deterioration of at least one other objective; finally, the vertical domain large module selects a final decision instruction output from the Pareto optimal solution set according to preset screening rules, such as maximum comprehensive utility and minimum maximum regret value, thereby realizing intelligent trade-offs of multiple objectives under complex constraints.

[0011] As a further technical solution, when the execution control module converts the decision commands generated by the building vertical domain large model module into specific control signals, a PID control algorithm is used for actuators requiring continuous adjustment, such as variable frequency water pumps, temperature control valves, and lighting dimmers; the PID control algorithm calculates the control quantity in real time. Where e(t) is the real-time error signal between the setpoint and the actual feedback value from the device, derived from the sensor; K p K i K d The proportional, integral, and derivative parameters of the controller, pre-tuned through engineering, are used to reflect the response strength to current error, historical error accumulation, and error change trends, respectively. The execution control module dynamically calculates and outputs control signals u(t) to drive the corresponding actuators to produce actions, thereby accurately tracking and stably controlling physical quantities such as temperature, flow rate, and illuminance, ensuring the accurate realization of decision-making intentions.

[0012] As a further technical solution, the building vertical domain large model module updates parameters online and offline based on the data fed back by the execution control module, including experience playback and a dual-thread update mechanism. The building vertical domain large model module constructs an experience tuple (s, a, r, s') together with the execution results of each decision instruction fed back by the execution control module, including the final state and various cost indicators, the multimodal state information on which the decision instruction was generated, and the decision instruction itself, and stores them in a first-in-first-out experience playback buffer. Here, s represents the state information before the decision, a represents the decision instruction generated by the building vertical domain large model module, r represents the immediate reward value calculated by the reward model based on the execution results, such as actual energy savings and shortened construction period, and s' represents the reward value after the instruction is executed. The new state is updated by the digital twin module; the parameter update process includes asynchronous online update threads and offline update threads; the online update thread periodically samples a batch of experience tuples from the buffer, for example, after collecting N new experiences, and uses the near-end policy optimization algorithm to calculate the policy gradient based on these samples, and incrementally updates the parameters of the large vertical model with a small learning rate using stochastic gradient descent to achieve rapid online adaptation; the offline update thread is automatically started when the system detects low computational load, performs batch sampling of all historical experience data accumulated in the buffer, calculates a more stable policy gradient estimate, and performs a complete and in-depth parameter retraining of the large vertical model to fully explore historical experience and optimize long-term policy performance.

[0013] This invention provides a system for a large-scale building vertical model and a digital twin intelligent agent, which has the following beneficial effects: 1. This invention solves the problem of information incompatibility and collaboration caused by inconsistent data formats and fragmented systems by constructing a dynamic digital twin that is synchronously mapped to physical entities and integrates multi-source heterogeneous building data. This enables the effective integration and interoperability of building lifecycle data under a unified spatiotemporal benchmark.

[0014] 2. This invention addresses the limitations of traditional systems that rely on fixed rule bases, making it difficult to deeply process unstructured data such as drawings, text, and point clouds, and hindering complex correlation analysis and creative optimization decisions. By using a large vertical domain model, this invention enables the automatic generation of optimization instructions covering multiple objectives such as production, construction, and energy consumption based on unified and integrated full-dimensional state information, significantly improving the system's intelligent decision-making level in complex scenarios.

[0015] 3. By establishing an optimization mechanism that includes execution feedback and online updating of model parameters, this invention solves the defects of static intelligent systems in the face of dynamically changing environments and tasks, such as rigid decision-making strategies and inability to continuously improve themselves. Ultimately, it realizes a fundamental improvement in building life cycle management from single static optimization to continuous dynamic self-adaptation, enhancing the long-term adaptability and overall performance of the system. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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] like Figure 1-2 As shown, taking a specific smart factory production scheduling scenario as an example, the implementation of the present invention will be described in further detail.

[0019] When implementing this invention system, the first step is system initialization and data foundation construction. The perception module is deployed based on the digital design requirements of the target building, such as a three-story office building. During the design phase, designers use Autodesk Revit software to create a complete Building Information Model (BIM) including beams, slabs, columns, walls, and electromechanical pipelines. The BIM strictly adheres to the IFC standard and is exported as an IFC XML file, containing not only the geometric dimensions and spatial location of components but also attribute information such as material strength grade, concrete grade, and steel reinforcement specifications. This data constitutes the basic static data. Simultaneously, to construct a high-precision digital twin environment for the factory, a terrestrial 3D laser scanner is used to scan the factory workshop where prefabrication is planned. The acquired point cloud data is stored in LAS format with a point cloud density of 100,000 points per square meter, used to accurately recreate the physical space of facilities such as machine tools, AGV channels, and material storage areas within the workshop. At the physical entity level, the system includes CNC machine tool spindles, material handling AGVs, workshop environmental monitoring points, and key power supply points within the factory workshop. An IoT sensor network is deployed in locations such as loops; these sensors include a vibration-temperature composite sensor for monitoring machine tool spindle vibration and temperature, with a sampling frequency of 100 times per second; UWB positioning tags for tracking the real-time position of AGVs, with an update frequency of 10 times per second; a temperature and humidity sensor for monitoring ambient temperature and humidity, with a sampling frequency of once per minute; and smart meters for monitoring the total energy consumption of each production line, with a data recording interval of once per minute; all sensor data is transmitted to the edge computing gateway through a 5G industrial private network deployed within the factory, initially aggregated, encapsulated in JSON format, and uploaded to the system's sensing module data access terminal via the MQTT protocol.

[0020] Upon receiving the aforementioned multi-source heterogeneous data, the sensing module immediately initiates a data standardization preprocessing procedure. For real-time operational data collected by sensors, due to the differences in their physical dimensions and numerical ranges, normalization processing is required to eliminate the influence of dimensions. For example, for spindle temperature data, assuming that its reasonable long-term monitored range is between 20 degrees Celsius and 80 degrees Celsius, the system will set a minimum threshold X. min The maximum threshold X is 20. max The value is 80. For any real-time acquired temperature reading X, it is calculated using the formula... The data is converted to a uniform numerical range of 0 to 1. If a sensor reading of 100 degrees Celsius is detected due to interference at a certain moment, exceeding the maximum threshold, it is marked as an outlier and removed. The data is then filled in using linear interpolation of valid data from the previous and next time moments. For energy consumption monitoring data, the Z-score standardization method is used. The system first calculates the mean μ and standard deviation σ of energy consumption data for a specific production line over the past 24 hours. For a new energy consumption reading χ, the formula is used... The system performs calculations to obtain data that follows a standard normal distribution. For high-dimensional point cloud data from 3D laser scanning and 2D view images rendered from BIM models, the system uses Principal Component Analysis (PCA) for feature extraction and dimensionality reduction. Specifically, for point cloud data, the system first calculates the eigenvalues ​​and eigenvectors of its covariance matrix, and then selects the top k principal components with a cumulative contribution rate of 95% as the feature representation after dimensionality reduction. This approach retains most of the geometric information while reducing the data dimension from hundreds of thousands of points to hundreds of feature dimensions, thus reducing the computational complexity of subsequent calculations. After cleaning, standardization, and dimensionality reduction, the multi-source data is packaged into a structured data package with a unified timestamp and data source identifier, ready to be output to the digital twin module.

[0021] The digital twin module is the core of constructing a synchronized virtual-real mapping. First, it establishes a unified three-dimensional spatiotemporal coordinate system in virtual space, with a fixed corner of the factory workshop as the origin. Next, it parses the BIM model IFC data provided by the sensing module into basic static data, precisely determining the spatial position, geometric dimensions, and assembly relationships of each prefabricated component, such as a prestressed beam numbered PCL-203, within this coordinate system. Simultaneously, the point cloud data of the factory workshop is spatially aligned with the basic static data using an iterative nearest-point (ICP) registration algorithm, ensuring that the physical position deviation between the virtual machine tool model and the actual machine tool is less than two millimeters. Then, the digital twin module performs data stream binding and time synchronization. It dynamically binds the vibration and temperature sensor data stream to the corresponding "CNC machine tool spindle number one" model component in the digital twin based on its installation location information; it binds the AGV's UWB positioning data stream to the "AGV model." To achieve synchronization of all data on a unified time axis, the system sets a reference timing frequency, such as 100 Hz; for the sampling frequency... For data with frequencies below 100 Hz, such as energy consumption data once per minute, the system uses a linear interpolation algorithm to generate equally spaced simulated data points between two actual sampling points per minute, adapting it to the 100 Hz update rhythm. For data packets with delayed timestamps that may exist during network transmission, the system caches and reorders them based on their timestamps to ensure the causal correctness of state updates. Finally, the digital twin module submits the candidate states formed by the bound data to the state adjudication mechanism. Only when the candidate state is adjudicated as an established state by the event consistency rule will the digital twin update the visualization and physical attribute parameters corresponding to the established state to the virtual model. For example, if a candidate state is adjudicated as an established state and contains evidence that the spindle temperature is standardized to 0.6, the color of the machine tool spindle model will be updated to orange according to the temperature-color mapping table defined in the specification. During the period when it is not adjudicated as an established state or when the candidate state is in the stage of failing adjudication, the candidate state shall not be used to drive the update of the position or attributes of the virtual model. The multimodal state information, which takes only the established state as input, is continuously provided to the building vertical domain large model module as simulation and decision input.

[0022] The building vertical domain large model module serves as the intelligent decision-making center of the system. Its internal pre-trained model is trained on a massive corpus of building domain data, including building design codes, construction process manuals, material performance databases, and historical project management logs. The model has 300 million parameters and adopts a 32-layer Transformer decoder architecture. Upon receiving multimodal state information streamed from the digital twin, the vertical domain large model first performs multimodal understanding and state analysis. It can simultaneously understand the geometric relationships of components in the BIM view, parse the "rebar binding completed" process status from the text log, and identify changes in inventory levels in the material storage area from the point cloud data. Combined with the built-in building knowledge base, such as concrete curing time requirements and crane hoisting safety regulations, the vertical domain large model performs expert-level reasoning on the current overall state. In this embodiment, the system objective is to process a batch of new personalized staircase component orders requiring delivery within five days. The vertical domain large model needs to generate optimal production scheduling and resource allocation decision instructions. To this end, it initiates multi-objective optimization calculations. The decision objective function is defined as follows: Where E represents predicted energy consumption, which is the predicted energy consumption value obtained by simulating candidate scheduling schemes that have been determined to be valid in the digital twin by the event consistency rule based on the digital twin simulation engine; the simulation operation only uses the valid state as the input evidence set, and does not use undecided candidate states as simulation input; T represents total time cost, i.e., order delivery cycle; C represents economic cost, including equipment depreciation, labor and material costs; weighting coefficients w1, w2, w3 is dynamically configured by the project administrator to 0.2, 0.5, and 0.3 based on the current peak and valley electricity price periods and the urgency of delivery; Under constraints, including maximum machine tool workload, number of AGVs, and material supply limitations, the vertical domain large model uses a multi-objective evolutionary algorithm to solve for the Pareto optimal solution set, and finally selects a balanced solution: split the new order into two batches, insert them into specific gaps in the production queues of Line 1 and Line 3 respectively, and replan the AGV material delivery route to reduce empty running distance; The decision instructions are specifically expressed as a series of structured operations: "At 2:30 PM, issue the processing program for stair component ST-01 to CNC machine tool No. 1; at 2:45 PM, dispatch AGV No. 3 to the raw material warehouse A area to load the corresponding type of steel; adjust the power of the laser cutting machine on Line 3 to 85% of the rated value to reduce instantaneous energy consumption."

[0023] After receiving the aforementioned decision instructions, the execution control module is responsible for converting them into control signals recognizable by the underlying devices. For scenarios requiring continuous and precise control, such as maintaining the temperature of a certain area in the workshop at 24 degrees Celsius, the execution control module employs a PID control algorithm. This algorithm calculates the control quantity u(t) in real time, and its formula is: Where e(t) is the error signal between the set temperature of 24 degrees and the actual feedback value of the temperature sensor; the proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd are set to 2.0, 0.5, and 0.1 respectively by the system in the early stage; the execution control module dynamically calculates and outputs a 4 to 20 mA analog current signal to the air conditioning unit actuator of the building automation system to drive the valve opening change, so as to achieve accurate tracking and stable control of temperature; for discrete instructions, such as AGV scheduling, the execution control module sends instruction messages to the AGV central scheduling system in JSON format through the standard RESTful API interface, the content of which includes the target point coordinates, path priority and task ID; after the AGV performs the transportation task, its status update data and task completion confirmation signal are collected and fed back to the system as execution result data.

[0024] Feedback and learning are key to the system's continuous optimization. After each decision command is executed, the execution control module collects data such as "the actual processing time for stair component ST-01 was 38 minutes, 3 minutes longer than expected" and "the actual empty travel distance of AGV No. 3 was 15 meters, 2 meters longer than planned," which are fed back to the digital twin module and the building vertical model module. The digital twin module immediately uses this feedback data to update the state of the dynamic digital twin, corrects the simulation model parameters, and enables the virtual model and the digital twin module to construct candidate states for subsequent judgment based on the execution result data returned by the execution control module. The candidate states are then submitted to the event consensus mechanism. The system uses consistency rules to determine whether a candidate state is valid. Only candidate states that are deemed valid by the event consistency rules are confirmed and updated as the current valid state of the digital twin module. Each decision and execution process is constructed as an experience tuple (s, a, r, s'), where: s represents the state description that was determined to be valid before the decision; a represents the decision instruction generated and issued by the building vertical domain large model module; r represents the immediate reward value calculated and output by the reward model based on the execution result; and s represents the candidate state formed after execution, which is only confirmed and replaced as the new current valid state after being determined by the event consistency rules. Candidate states that have not been determined by the event consistency rules cannot be directly used as s' in the experience tuple for subsequent simulations, deductions, or model parameter updates. Only when s' is determined to be valid can the experience tuple containing s' be used for the experience replay buffer and subsequent incremental updates or offline retraining. The experience replay buffer stores experience tuples that have been ruled as valid and their corresponding ruling records in an append-only structure. To avoid misleading training due to unresolved data, the system must not write unresolved candidate states or their derived experience tuples into the experience replay buffer.

[0025] The system's parameter update process is completed collaboratively by asynchronous online and offline update threads. For every 1,000 new experiences accumulated, the online update thread randomly samples a small batch of data containing 256 experiences from the buffer, calculates the policy gradient using the Proximal Policy Optimization (PPO) algorithm, and incrementally updates the policy network parameters of the large-scale vertical model using stochastic gradient descent with a learning rate of 0.0001, aiming to quickly adapt to recent task patterns. The offline update thread starts every day in the early morning when the system load is low, shuffles all historical experience data in the buffer, samples it in batches, and performs a complete retraining of the model parameters. This process aims to improve the model's long-term generalization ability and policy stability. When faced with new order fluctuations, equipment malfunctions, or process changes, the system can continuously evolve, thereby continuously improving the overall efficiency, flexibility, and intelligence level of industrialized building production.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for a large-scale building vertical model and a digital twin intelligent agent, characterized in that, It includes a perception module, a digital twin module, a large-scale building vertical model module, and an execution control module; The perception module acquires real-time operational data through sensors deployed on the physical building, obtains geometric and attribute data through BIM modeling, and acquires point cloud data through 3D laser scanning, outputting multi-source heterogeneous building lifecycle data; The digital twin module receives building lifecycle data and fuses it based on a unified spatiotemporal benchmark to construct a dynamic digital twin that is synchronously mapped to the building's physical entity. The digital twin integrates a visualization engine and a simulation engine, which are used to perform simulation and deduction based on candidate states generated from the bound data and determined to be valid states by the event consistency rules. The digital twin provides output in multimodal state information with time-series labels and component semantic labels, based on the valid states. Before a candidate state that has not passed the adjudication is confirmed as an established state, it will not be used as the output source of the multimodal state information; The building vertical domain large model module receives multimodal state information, is based on the Transformer architecture, and is pre-trained using building domain corpus. It has the ability to understand and reason about BIM views, construction log text, and point cloud data in a multimodal manner. By parsing multimodal state information and combining it with the building knowledge base for expert-level reasoning, it generates decision instructions that include production scheduling, construction optimization, and energy consumption control. The execution control module receives decision instructions, converts them into driveable control signals, and issues them for execution. After execution, the execution control module collects execution result data according to the evidence item format defined in the manual and sends it back to the digital twin module and the building vertical domain large model module. The digital twin module constructs candidate states based on the returned execution result data and submits them to the event consistency rules for adjudication. Only candidate states that are adjudicated as valid states are used to replace the current valid state of the digital twin. The building vertical domain large model module only uses the established status as the input evidence set for near-end strategy optimization or parameter update when it receives the established status confirmed by the digital twin module.

2. The system for a large-scale building vertical domain model and a digital twin intelligent agent according to claim 1, characterized in that: Before being input into the digital twin module, the building lifecycle data acquired by the sensing module undergoes data standardization processing. This processing includes: for real-time operational data collected by sensors and possessing different physical dimensions, using formulas... Normalization calculations are performed to convert the data to a uniform numerical range; for energy consumption monitoring data, the formula is used. Standardization calculations are performed, where μ is the mean and σ is the standard deviation. Then, feature extraction and dimensionality reduction are performed. Features are extracted from high-dimensional data, including point clouds and images, and dimensionality reduction is performed while retaining some variance information to reduce computational complexity and improve model training efficiency.

3. The system of a large-scale building vertical domain model and digital twin intelligent agent according to claim 1, characterized in that: The dynamic digital twin module includes: establishing a unified spatiotemporal coordinate system; using acquired geometric and attribute data as basic static data, determining the spatial position and subordinate relationship of building components in the coordinate system; binding the operational data acquired by sensors to the corresponding building components in the basic static data based on the sensor's installation position on the physical entity; associating the acquired point cloud data with the basic static data through spatial alignment; receiving and caching the bound data in a unified order, and sorting and organizing the data according to a preset receiving order rule; characterized in that the digital twin does not directly use the latest received data as the current state, but introduces a state adjudication mechanism for candidate... The existence of states is determined; the state determination mechanism includes: defining the state formed by the bound data at any given time as a candidate state; determining whether the candidate state satisfies continuity, integrity, and component constraint consistency according to a pre-set event consistency rule; and only when a candidate state passes the event consistency rule determination is it considered the current established state of the digital twin; wherein, the current state of the digital twin is not determined by time sequence or data age, but solely by whether it passes the event consistency rule determination; candidate states that fail the determination do not participate in the state progression of the digital twin and are not used for subsequent simulation, deduction, or decision input; the digital twin is considered to exist effectively only when at least one state is determined to be established.

4. The system of a large-scale building vertical model and digital twin intelligent agent according to claim 1, characterized in that: The aforementioned large-scale architecture model module is pre-trained on an architecture domain corpus and then optimized through a supervised fine-tuning stage. In this stage, the large-scale architecture model is trained using labeled architecture task data. The predicted probability distribution is obtained through forward computation, and a cross-entropy loss function is employed. Calculate the error between the output of the large-scale building vertical model and the true label, where y i For real labels, p i The probability values ​​predicted by the large model of the building vertical domain are used to adjust the model parameters by minimizing the cross-entropy loss function through backpropagation.

5. The system of a large-scale building vertical model and digital twin intelligent agent according to claim 4, characterized in that: Following supervised fine-tuning, the building vertical domain large model module further optimizes its policies through reinforcement learning based on human feedback. During this process, a reward model is introduced to score the quality of the decision instructions generated by the vertical domain large model module. Using the quality score as the reward signal, a proximal policy optimization algorithm is employed to maximize the expected reward function. Multiple rounds of iterative training are performed for the target, where πθ represents the policy of the vertical domain large model module, and r(s,a) represents the reward value given by the reward model.

6. The system for a large-scale building vertical model and a digital twin intelligent agent according to claim 1, characterized in that: When generating decision instructions, the large-scale building vertical model module performs multi-objective optimization calculations; it constructs an objective function with the goal of minimizing the overall cost. Where E represents predicted energy consumption, T represents time cost, C represents economic cost, and w1, w2, and w3 are weight coefficients dynamically configured according to project stage and strategy; the vertical domain large model module solves the Pareto optimal solution set of the objective function under multiple constraints and selects the final decision instruction from it.

7. The system of a large-scale building vertical domain model and digital twin intelligent agent according to claim 1, characterized in that: When the execution control module converts the decision command into a control signal, it employs a PID control algorithm for actuators requiring continuous adjustment; the PID control algorithm calculates the control quantity in real time. Where e(t) is the error signal between the set value and the actual feedback value of the device, K p ,K i ,K d The controller parameters are pre-tuned; by dynamically calculating and outputting control signals, the actuator is driven to perform actions, thereby tracking and stabilizing the physical quantity.

8. The system of a large-scale building vertical model and digital twin intelligent agent according to claim 1, characterized in that: The process of updating parameters based on feedback data in the building vertical domain large model module includes the module taking the execution result of the decision instruction fed back by the control module each time, the multimodal state information on which the decision instruction was generated, and the decision instruction itself, and storing them in the experience replay buffer. Where s represents state information, a represents the generated decision instruction, r represents the immediate reward value calculated by the reward model based on the execution result, and s' represents the new state after execution; The parameter update process includes asynchronous online update threads and offline update threads. The online update thread periodically samples a portion of the batch experience tuples from the buffer, calculates the policy gradient using the near-end policy optimization algorithm, and incrementally updates the large vertical model using stochastic gradient descent. The offline update thread, when the system load is low, starts batch sampling of all historical experience data in the buffer, calculates the policy gradient of the reward function, and performs a complete retraining of the model parameters.

Images