Digital twin model construction method and system for multi-party collaboration in house construction

Abstract

This invention relates to the field of construction monitoring technology and discloses a method and system for constructing a digital twin model for multi-party collaboration in building construction. The method includes: design data acquisition; design scheme and interface standard formulation; adaptive adjustment of interface parameters; and digital twin model construction. The method acquires standard design data for building construction, generates parameterized design schemes through parametric modeling, and formulates dynamically reconfigurable interface standards. Then, it uses a dynamic interface generator to adaptively adjust interface parameters. Finally, based on standard point cloud data and construction site point cloud data, it employs LOD hierarchical technology combined with a lightweight adaptive point cloud registration algorithm to construct a digital twin model synchronized with the physical entity of the building construction. This effectively improves the matching accuracy and dynamic adaptability between the digital twin model and the physical entity of the building construction, significantly increases modeling efficiency, reduces computational consumption, and solves the problems of insufficient model accuracy and poor dynamic adaptability in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of construction monitoring technology, specifically to a method and system for constructing a digital twin model for multi-party collaboration in building construction. Background Technology

[0002] With the acceleration of urbanization and the diversification of housing demands, modular construction has become a core direction for the transformation and upgrading of the construction industry. The industrialization and large-scale development of modular housing places extremely high demands on the collaboration of the industrial chain. The modular housing industrial chain involves four main participants: design institutes, component factories, logistics companies, and construction companies. Efficient collaboration among these entities is key to achieving seamless integration of the entire process from design, production, logistics, to assembly. Therefore, building a multi-party collaborative management system adapted to modular housing construction has become an inevitable requirement for the industry's development.

[0003] Digital twin technology, as a core technology for achieving precise mapping and real-time linkage between physical entities and virtual models, provides crucial technical support for multi-party collaboration in modular housing construction. High-precision digital twin models can accurately reproduce the physical structure and interface characteristics of modular houses, providing a reliable basis for design optimization, production parameter matching, on-site assembly guidance, and multi-party collaborative scheduling.

[0004] In existing technologies, for example, the enterprise collaboration system based on the industrial chain disclosed in Chinese patent application CN117311789 constructs a collaborative architecture for a group of enterprises in the industrial chain, which solves the problem that existing industrial chain-related software and industrial internet platforms cannot meet the service needs of large-scale, complex, and dynamically changing enterprise groups. Another example is the model management and collaborative analysis method and system based on digital twins announced in Chinese invention patent application CN119359477B, which studies the application of digital twin technology in collaborative analysis.

[0005] However, existing technologies still have the following problems in constructing and applying digital twin models for multi-party collaboration in modular housing construction, making it difficult to meet actual collaboration needs. Specifically, these problems are as follows: The modular houses designed using static standards have insufficient standardization of module interfaces and lack dynamic adjustment capabilities. Design change information cannot be transmitted to the digital twin modeling stage in real time, resulting in the model being unable to be updated in a timely manner according to the dynamic changes in design and construction, and the model has poor timeliness and adaptability.

[0006] The core algorithms used in model building, such as point cloud registration, are mostly general-purpose algorithms that do not take into account the special characteristics of modular houses. They cannot accurately capture the interface features, key structural details, and actual conditions of the construction site of modular houses, resulting in a low degree of matching between the constructed digital twin model and the physical entity.

[0007] In summary, there is an urgent need for a digital twin model construction method for multi-party collaboration in housing construction to solve the problems of insufficient model accuracy and poor dynamic adaptation capability in existing technologies. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for constructing digital twin models for multi-party collaboration in housing construction. This method achieves the construction of high-precision and highly adaptable digital twin models for multi-party collaboration in housing construction, solving the problems of insufficient model accuracy and poor dynamic adaptability in existing technologies.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a digital twin model for multi-party collaboration in housing construction, comprising the following steps: Obtain standard design data for housing construction. The standard design data includes standardized component library data and external input requirement data. The standardized component library data consists of predefined building modules and their parameterized attribute data. The external input requirement data reflects the needs of the housing construction participants. Based on the standard design data, a parametric design scheme is generated through a parametric modeling algorithm. At the same time, a dynamic reconfigurable interface standard for preset interfaces supported by a dynamic interface generator is established. The dynamic interface generator is used to adjust the connection method of the building module according to the external input requirement data. Based on the parametric design scheme and the dynamic reconfigurable interface standard, the interface parameters are adaptively adjusted by a dynamic interface generator that integrates a dynamic graph neural network interface adaptation algorithm to obtain the interface adaptation parameters. Based on the standard point cloud data extracted from the BIM model and the collected point cloud data from the construction site, the interface adaptation parameters are integrated, and a lightweight modeling process is performed using the LOD hierarchical method combined with a lightweight adaptive point cloud registration algorithm to obtain a digital twin model synchronized with the physical entity of the building construction.

[0010] This application provides a digital twin model construction system for collaborative management of housing construction, and the method for constructing the digital twin model for collaborative management of housing construction includes: The design data acquisition module is used to acquire standard design data for building construction. The standard design data includes standardized component library data and external input requirement data. The standardized component library data consists of predefined building modules and their parameterized attribute data. The external input requirement data reflects the construction requirements of the building construction participants. The design scheme and interface standard formulation module is used to generate a parametric design scheme based on the standard design data through a parametric modeling algorithm, and at the same time formulate a dynamic reconfigurable interface standard for the preset interface supported by the dynamic interface generator. The interface parameter adaptive adjustment module is used to adaptively adjust the interface parameters based on the parameterized design scheme and the dynamic reconfigurable interface standard by using a dynamic interface generator that integrates a dynamic graph neural network interface adaptation algorithm to obtain interface adaptation parameters. The digital twin model construction module is used to integrate standard point cloud data extracted from the BIM model with the collected point cloud data from the construction site, and to perform lightweight modeling processing by combining the interface adaptation parameters with the LOD hierarchical method and the lightweight adaptive point cloud registration algorithm to obtain a digital twin model that is synchronized with the physical entity of the building construction.

[0011] Compared with the prior art, the technical solution of this application has the following beneficial effects: 1. This application generates an adaptable parametric design scheme based on external input demand data reflecting the construction needs and environment of the building, and formulates a dynamic reconfigurable interface standard. At the same time, it uses a dynamic interface generator to complete the adaptive adjustment of interface parameters. The generated interface adaptation parameters can accurately match various constraints of the construction environment, and the parameters are integrated into the modeling process. This allows the digital twin model to not only adapt to conventional construction scenarios, but also to dynamically adapt to the actual conditions of the construction site, solving the problem of traditional digital twin models being out of touch with the actual construction scenario and having poor adaptability.

[0012] 2. This application achieves closed-loop optimization of the interface adaptation algorithm based on a reinforcement learning reward mechanism. It generates reward values ​​with interface installation success rate and module connection stability as core indicators to drive algorithm parameter iteration, ensuring the rationality and feasibility of interface adaptation parameters, effectively improving the accuracy of interface installation and the long-term reliability of module connection, reducing interface connection deviations and failures during construction, reducing construction rework rate, and ensuring the connection quality of modular housing construction.

[0013] 3. This application extracts features through a lightweight convolutional neural network, optimizes the point cloud registration process for jittery scenes, occluded scenes, and complex terrains, significantly reduces the computational power consumption of point cloud registration, adapts to the computational power limitations of mobile devices and edge nodes, and significantly shortens the model building and point cloud processing time compared with traditional full-dimensional high-detail modeling and conventional point cloud registration algorithms, improves the overall modeling efficiency, and realizes rapid construction and smooth interaction of digital twin models. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method for constructing a digital twin model for multi-party collaboration in housing construction according to the present invention; Figure 2 This is a structural diagram of the digital twin model construction system for multi-party collaboration in housing construction according to the present invention; Figure 3This is a structural diagram of a modular housing industry chain collaborative system provided in a specific embodiment of the present invention. Detailed Implementation

[0015] Please see Figure 1 The method for constructing a digital twin model for multi-party collaboration in house construction in this embodiment includes the following steps: Obtain standard design data for housing construction. The standard design data includes standardized component library data and external input requirement data. The standardized component library data consists of predefined building modules and their parameterized attribute data. The external input requirement data reflects the needs of the housing construction participants. Based on the standard design data, a parametric design scheme is generated through a parametric modeling algorithm. At the same time, a dynamic reconfigurable interface standard for preset interfaces supported by a dynamic interface generator is established. The dynamic interface generator is used to adjust the connection method of the building module according to the external input requirement data. Based on the parametric design scheme and the dynamic reconfigurable interface standard, the interface parameters are adaptively adjusted by a dynamic interface generator that integrates a dynamic graph neural network interface adaptation algorithm to obtain the interface adaptation parameters. Based on the standard point cloud data extracted from the BIM model and the collected point cloud data from the construction site, the interface adaptation parameters are integrated, and a lightweight modeling process is performed using the LOD (Level of Detail) hierarchical method combined with a lightweight adaptive point cloud registration algorithm to obtain a digital twin model synchronized with the physical entity of the building construction.

[0016] In a specific embodiment, the generation of parametric design schemes involves: receiving external input requirement data and site conditions and environmental parameters from the construction site, based on standardized component library data; generating parametric design schemes adapted to the requirements through parametric modeling algorithms; and outputting core design parameters such as module type, geometric dimensions, and material properties. After generating the parametric design schemes, the corresponding manufacturability and assemblability must be verified in real time to avoid assemblability conflicts during interface adaptation and ensure the feasibility of the interface adaptation process. The dynamic reconfigurable interface standard, for the mechanical, electrical, and spatial interfaces supported by the dynamic generator, clarifies the tolerance range, angle adjustment interval, and functional requirements for interface adaptation, and presets basic rules for interface adaptive adjustment to facilitate the subsequent implementation of dynamic interface adaptation algorithms.

[0017] In this embodiment, the constructed digital twin model achieves precise mapping of the geometric shape and functional parameters of the three core interfaces of the modular house: mechanical, electrical, and spatial. Furthermore, a lightweight adaptive point cloud registration algorithm is used to achieve high-precision registration between the construction site point cloud data and standard point cloud data, ensuring a high degree of consistency between the model's spatial location and geometric shape and the physical entities at the construction site. Simultaneously, LOD layered modeling ensures dual fidelity in both the model's geometric shape and attribute parameters, significantly improving the overall restoration accuracy of the digital twin model and solving the problems of insufficient model accuracy and poor dynamic adaptability in existing technologies.

[0018] Preferably, the preset interfaces supported by the dynamic interface generator include mechanical interfaces, electrical interfaces, and spatial interfaces; The dynamic interface generator supports preset interfaces that integrate multi-dimensional micro sensors. The multi-dimensional micro-sensors include a pressure sensor, a displacement sensor, an angle sensor, and a temperature sensor mounted on the mechanical interface; a signal strength sensor and a voltage and current sensor mounted on the electrical interface; and a laser rangefinder and a contour sensor mounted on the spatial interface.

[0019] Specifically, the sensors installed on the mechanical interface are used to collect locking stress, interface gap displacement, docking angle, and ambient temperature. The sampling frequency is generally set to 10Hz to ensure real-time capture of stress changes, gap deviations, angle shifts, and ambient temperature changes. Ambient temperature monitoring is often used in extreme building scenarios. The sensors integrated into the electrical interface are used to collect signal strength reflecting the stability of data communication and voltage and current reflecting wireless charging efficiency. This is to monitor the risk of communication interruption and charging loss during interface insertion and removal in real time, avoiding electrical faults caused by poor interface contact. The sensors on the spatial interface are used to collect spatial gaps and interface morphological deviations at the junctions of building modules, thereby capturing the dynamic morphological changes of the spatial interface and providing accurate data for parametric fuzzy boundary adjustment.

[0020] In one specific embodiment, the mechanical interface includes an intelligent locking device to achieve tolerance within a preset tolerance range (e.g., ±5mm) and adaptive adjustment within a preset angle range (e.g., 0°-90°), ensuring the stability and adaptability of the mechanical connection. For the electrical interface, it integrates wireless charging and data communication functions to achieve plug-and-play functionality, ensuring the stability of the electrical connection and smooth communication. For the spatial interface, through parametric fuzzy boundary design, it supports dynamic adjustment of the shape of the junction of building modules, adapting to spatial deviations between modules and improving the fit of the spatial connection.

[0021] In this embodiment, specific sensor types are matched to the functional characteristics of different interfaces. Mechanical interfaces focus on the physical state of connection, electrical interfaces emphasize communication and power supply status, and spatial interfaces accurately capture spatial morphological deviations, providing comprehensive and accurate raw data support for adaptive adjustment of interface parameters. Furthermore, the high compatibility between sensor and interface types ensures that the collected data directly reflects the core operation and connection status of each interface, avoiding invalid data collection, reducing computational consumption in subsequent data processing, and improving the efficiency and accuracy of interface parameter adjustment decisions.

[0022] Preferably, in order to overcome the limitations of existing interface adaptive adjustment which only adjusts fixed tolerances, and considering the different application scenarios of modular housing, such as the interface requirements of conventional buildings and high-rise buildings, the interface parameters are autonomously optimized to solve the problems of fixed adaptation, poor environmental adaptability, and lack of feedback adjustment in traditional interfaces. The adaptive adjustment of interface parameters specifically includes: Real-time status data of the construction site is collected through the multi-dimensional micro sensors; The real-time status data is transmitted to a dynamic interface generator and transformed into node feature vectors and edge feature vectors of the dynamic graph neural network. The node feature vectors and edge feature vectors are then fused using a temporal attention mechanism to obtain the fused feature vectors. The node features focus on the building module itself, including module type, geometric dimensions, material properties, and real-time status data. The edge features focus on the interface connections between modules, including connection type (mechanical / electrical / spatial), interface parameters such as the initial value of the locking torque of the mechanical interface, the communication protocol type of the electrical interface, the initial parameters of the boundary of the spatial interface, and data reflecting the real-time stress state in the real-time status data (such as interface locking stress and pressure distribution). The fused feature vector is dynamically modeled, and parameterized adjustment suggestions are generated for issues such as excessive real-time stress at the interface, tolerance adaptation deviation, and changes in environmental parameters. These parameterized adjustment suggestions include: suggestions for adjusting locking torque and angle deflection, suggestions for switching electrical communication protocols, and suggestions for adjusting morphological parameters. The parameterized adjustment suggestions are verified to ensure their assemblability and manufacturability, and to avoid conflicts between the adjustment suggestions and the design scheme and module parameters. The verified parameterized adjustment suggestions are then sent to various interfaces for adaptive adjustment of interface parameters. The real-time status data collected after adjustment is fed back to the dynamic interface generator through a preset data interaction protocol for parameter optimization of the interface adaptation algorithm.

[0023] In one specific embodiment, corresponding intelligent actuators are configured for the three types of interfaces to achieve adaptive adjustment of interface parameters and ensure matching accuracy. The mechanical interface is equipped with a micro-electric adjustment mechanism that automatically fine-tunes the position and force of the intelligent locking device based on received locking torque and angle deflection adjustment suggestions, achieving precise adjustments at the ±0.1mm level to ensure the stability and security of the interface connection. For the electrical interface, a retractable contact probe is integrated, which automatically adjusts the probe's extension length based on received electrical communication protocol switching suggestions to ensure good contact and optimize the position of the wireless charging coil to improve charging efficiency. For the spatial interface, a deformable seal made of shape memory alloy is used, which automatically deforms the gap and shape of the corresponding module junction based on received morphological parameter adjustment suggestions, balancing sealing and flexibility.

[0024] In this embodiment, the input dimensions of node and edge features in the dynamic graph neural network structure are first optimized, addressing the shortcomings of traditional methods that only focus on the module's own parameters and ignore the influence of multiple factors, thus improving the algorithm's modeling accuracy of the interface's dynamic state. Then, through a temporal attention mechanism, the weights of core features directly related to interface adaptation, such as interface stress, displacement deviation, and ambient temperature, are significantly enhanced, improving the targeting and effectiveness of feature fusion. Furthermore, based on the specific adaptation requirements of the reconfigurable interface, targeted parameterized adjustment suggestions are generated and implemented only after verification, solving the problems of vague and unimplementable traditional adjustment suggestions.

[0025] In a specific embodiment, in addition to the three types of adjustments mentioned above, the parameter adjustment of the interface material is also included. That is, based on the ambient temperature collected by the installed temperature sensor, the thermal expansion and contraction compensation parameters of the interface material are automatically adjusted for high temperatures such as exceeding 50°C and low temperatures such as below -20°C, so as to avoid loosening or damage to the interface due to changes in ambient temperature and to adapt to the assembly needs of conventional and extreme building scenarios.

[0026] Preferably, the parameter optimization of the interface adaptation algorithm specifically includes: The received real-time status data is compared and analyzed to generate feedback evaluation results, which are used to determine whether the adaptive adjustment of interface parameters meets the standards. Based on a preset reinforcement learning reward mechanism, the interface installation success rate and the later stability of module connection are used as reward evaluation indicators. The reward value is obtained by combining the feedback evaluation results. The reward value includes positive reward value and negative reward value. The interface success rate is used to measure the accuracy of interface adaptation, and the later stability of module connection is used to measure the long-term reliability of interface adaptation. Specifically, the interface success rate is used to measure the convenience and accuracy of interface adaptation, and is generally required to be no less than 99%. It is defined as the ratio of the number of interfaces that are successfully installed in one session to the total number of interfaces in each round of interface adaptation within a certain period. The module connection stability in the later stage is defined as the duration for which parameters such as stress and displacement monitored by multi-dimensional micro sensors remain within the standard evaluation threshold range after interface adjustment. It is used to measure the long-term reliability of interface adaptation. The obtained reward value is input into the dynamic graph neural network interface adaptation algorithm integrated in the dynamic interface generator, triggering parameter iteration of the dynamic graph neural network interface adaptation algorithm.

[0027] In this embodiment, feedback evaluation results are generated by accurately analyzing real-time status data, and reward values ​​are generated by combining reinforcement learning reward mechanism to achieve algorithm iteration and detection. This effectively solves the problems of blind iteration, difficulty in quantifying adaptation effect, and easy assembly and manufacturing conflicts after adaptation in traditional interface adaptation algorithms, and significantly improves the accuracy, stability and feasibility of interface adaptation.

[0028] In a specific embodiment, after the above parameters are iterated, in addition to verifying assemblability and manufacturability, the optimized algorithm model also needs to be applied to the next round of interface adaptation. This process forms a continuous closed loop until the interface adaptation accuracy and fault tolerance reach the preset targets, achieving dynamic iteration and continuous optimization of the algorithm model.

[0029] It should also be added that the process of obtaining feedback and evaluation results includes: Real-time status data is preprocessed to ensure data availability, including but not limited to data cleaning, standardization, and outlier removal, where outliers refer to invalid data caused by sensor malfunctions. Extract core feature parameters, including node feature parameters and edge feature core parameters. Node feature core parameters focus on the building module itself, including module type, geometric dimensions, material properties (elastic modulus, tensile strength, heat resistance parameters), and interface strain values, displacement deviation values, and ambient temperature in real-time status data. Edge feature core parameters focus on the interface connection between modules, including connection type, interface basic parameters, and data reflecting the real-time stress state in real-time status data (such as interface locking stress and pressure distribution). Screening and reinforcement learning reward calculation, interface adaptation algorithm parameter iteration, key evaluation data directly related to the selection and reinforcement learning reward calculation, interface adaptation algorithm parameter iteration, including but not limited to interface actual status data, installation success rate, and module connection later stability; The pre-stored standard evaluation thresholds in the dynamic interface generator are compared with the acquired key evaluation data, and the deviation between the two is calculated. If the deviations of all key evaluation data meet the compliance conditions, the feedback evaluation result is recorded as a qualified evaluation result; otherwise, it is recorded as a unqualified evaluation result.

[0030] The specific process for obtaining reward values ​​is as follows: The specific values ​​of the reward evaluation indicators are normalized to a quantitative value between 0 and 1. Then, the normalized results are multiplied by their corresponding weights (generally set to 0.5) and summed. If the result is a qualified evaluation, the sum is the positive reward value. If the result is an unqualified evaluation, the sum minus the evaluation penalty coefficient is the negative reward value. The evaluation penalty coefficient ranges from (0.1 to 0.5) and can be obtained from a mapping table based on the reward evaluation indicators. The mapping table is established based on historical reward evaluation indicators and the proportion of stable duration. The proportion of stable duration is the ratio of the module connection's later-stage stability (stability duration) to the preset duration. The positive reward value is the reward for a qualified evaluation result. The closer the value is to 1, the better the adaptation effect. When the interface installation success rate is 100% and the module connection's later-stage stability far exceeds the preset requirements, the positive reward value is capped at 1. The negative reward value is the reward value for a feedback evaluation result that is unqualified. The closer the value is to -1, the worse the adaptation effect. When serious problems such as installation failure or excessive stress occur, the negative reward value is taken as the lower limit of -1.

[0031] Preferably, the lightweight modeling process includes: Using the standard point cloud data exported from the BIM model as a reference, the construction site point cloud data is registered using a lightweight adaptive point cloud registration algorithm to obtain the registered construction site point cloud data and the registration transformation matrix. Using the BIM model as the modeling framework, and integrating the interface adaptation parameters, the LOD hierarchical technology is used to perform layered lightweight modeling to obtain the initial LOD layered model. Specifically, the initial LOD layered model is obtained, including: dividing the model into multiple detailed levels according to the needs of collaborative management of building construction, such as the overall site topography outline layer, the overall building structure frame layer, the building module and interface fine structure layer, and the full-dimensional high-fidelity detail layer, and setting the display rules for each level. Using the BIM model as a framework, the interface adaptation parameters are accurately mapped to the interface structure of each level. At the same time, site environmental information data is integrated, and each level is modeled in turn and non-core redundant data is removed to obtain the initial LOD layered model. Based on the registration transformation matrix, the registered construction site point cloud data and the initial LOD layered model are spatially fused and parameter calibrated to obtain the LOD layered model; By connecting the real-time status data to the LOD hierarchical model through the digital mainline, a real-time data synchronization link between the physical entity and the model is established, thus completing the construction of a digital twin model for collaborative housing construction.

[0032] In a specific embodiment, spatial fusion and parameter calibration specifically include: The initial LOD layered model is transformed in its overall spatial coordinates based on the registration transformation matrix, so that the model's three-dimensional spatial position and orientation are consistent with the physical entity at the construction site. The geometric features of the registered construction site point cloud data are compared with the geometric features of each level of the initial LOD layered model. For the deviation areas, local geometric corrections are made in combination with interface adaptation parameters and real-time status data. The site environment information data and interface adaptation parameters are added to the corresponding levels and structures of the initial LOD layered model to achieve dual fidelity of model geometry and attribute parameters, thus obtaining the LOD layered model.

[0033] In this embodiment, a lightweight modeling technique based on Level of Detail (LOD) is employed. Different detail levels are defined according to the actual needs of collaborative building construction management. This effectively reduces model data volume and rendering computation while ensuring the fidelity of core structures and key parameters. Simultaneously, a lightweight adaptive point cloud registration algorithm is used, extracting features and optimizing the registration process through a lightweight convolutional neural network, significantly reducing the computational cost of point cloud registration. Compared to traditional full-dimensional high-detail modeling and conventional point cloud registration algorithms, this significantly shortens model construction and point cloud processing time, improves overall modeling efficiency, and enables rapid construction and smooth interaction of the digital twin model.

[0034] It is important to understand that the core objective of point cloud registration is to accurately match the point cloud data at the construction site with the standard point cloud data in terms of spatial location and geometric shape. The interface areas (mechanical / electrical / spatial) of the modular house are the key feature areas for point cloud registration. The interface adaptation parameters provide an accurate feature quantification benchmark for the point cloud registration of this area. If there is only the point cloud registration result, the model can only restore the overall spatial shape of the modular house, and the interface area will be blurry. If there is only the interface adaptation parameters, the parameters do not have a corresponding spatial carrier and cannot be integrated into the 3D model. Only by integrating the two can the LOD-level model be lightweight and high-fidelity.

[0035] Preferably, the step of performing point cloud registration on the construction site point cloud data using a structurally lightweight adaptive point cloud registration algorithm to obtain a registration transformation matrix and registered construction site point cloud data includes: Step 1: Perform point cloud preprocessing on the collected construction site point cloud data (obtained through drone aerial photography and laser scanning equipment) and the exported standard point cloud data to obtain construction site point cloud preprocessed data and standard point cloud preprocessed data. The point cloud preprocessing includes point cloud filtering, downsampling and coordinate normalization. Point cloud filtering involves using statistical filtering algorithms to remove noise points, isolated points, and invalid data from the point cloud, preventing outliers from affecting subsequent registration accuracy. Downsampling employs a voxel downsampling method, setting a reasonable grid size such as 5mm×5mm×5mm to reduce the amount of point cloud data and algorithm computation while preserving the core geometric features of the point cloud, thus adapting to the computing power limitations of mobile devices or edge nodes. Coordinate normalization involves performing coordinate normalization on the two types of downsampled point cloud data, using the geodetic coordinate system commonly used in building construction as a unified coordinate system to eliminate coordinate deviations. Step 2: Input the preprocessed point cloud data of the construction site and the preprocessed point cloud data of the standard point cloud into a lightweight convolutional neural network (such as a MobileNetV3 variant) to extract local features and obtain local feature vectors of the construction site and standard local feature vectors. Then, through the introduced attention mechanism, feature enhancement of key building structures (such as beam-column joints and interface areas) is performed, and the enhanced feature vectors of the site point cloud and standard point cloud are output. Specifically, during the local feature extraction process, the focus is on capturing the geometric shape, spatial location, and texture features of the point cloud. The introduced attention mechanism, based on the architectural structural characteristics of modular houses, automatically assigns feature weights, focusing on enhancing the feature weights of point clouds corresponding to key structures and weakening the influence of point cloud features in non-critical areas (such as temporary protective areas). Step 3: Perform coarse registration on the enhanced feature vector of the on-site point cloud and the enhanced feature vector of the standard point cloud. For example, use the nearest neighbor feature matching algorithm to match the two types of feature vectors, construct feature matching pairs, and obtain the initial registration transformation matrix and the initial feature matching pairs. The core purpose of coarse registration is to reduce the spatial position deviation between the two types of point cloud data within a preset reasonable range, such as not exceeding 5cm. Step 4: Filter the initial feature matching pairs, such as by using the RANSCAN algorithm (RANSAC with Consensus and Neighborhood) to obtain valid feature matching pairs. During the filtering process, based on the geometric characteristics of the modular house building structure, preset matching distance thresholds and angle thresholds are used. Through iterative sampling consistency checks, valid feature matching pairs that conform to the building structure logic (such as vertical beams and columns, parallel interfaces, and other geometric relationships) are filtered out. Step 5: Perform fine registration on the effective feature matching, such as using the Iterative Closest Point (ICP) algorithm to output a registration transformation matrix. The fine registration includes iteratively optimizing the initial registration transformation matrix based on Euclidean distance (with the goal of minimizing the Euclidean distance error between corresponding points of the two types of point cloud data), dynamically adjusting the iteration step size by combining an adaptive compensation optimization mechanism, and introducing local geometric constraints to constrain and correct the initial registration transformation matrix to ensure that the registration result conforms to the building structure logic. This achieves dynamic adaptive adjustment of the registration parameters, transforming point cloud registration from "mathematical coordinate alignment" to "structural logic alignment," ensuring that the registration result fits the actual construction and assembly requirements of modular houses, and improving the registration accuracy. The iteration step size is adjusted according to the trend of Euclidean distance error in each iteration. When the error is large, such as exceeding the preset upper limit threshold of distance error, the iteration step size is increased to speed up the convergence. When the error is less than the preset lower limit threshold of distance error, the iteration step size is decreased to improve the registration accuracy. The adjustment amount is based on the deviation between the Euclidean distance error and the corresponding threshold, which is obtained from the established table. The table stores the relationship between the deviation and the adjustment amount. Step Six: Verify the accuracy of the registration results, such as the root mean square error (RMSE). Determine if the obtained registration error meets the threshold condition. If it does, output the registration transformation matrix. If the registration error is less than or equal to ±2cm, it meets the accuracy requirements for modular house point cloud registration. Synchronize the registration results to the digital twin modeling unit for calibration and updating of the digital twin model. Based on the registration transformation matrix, perform a three-dimensional spatial coordinate transformation on the preprocessed construction site point cloud data to generate registered construction site point cloud data. Otherwise, return to Step Five and re-perform precise registration until the registration error meets the condition. The registration result is the spatial matching result between the construction site point cloud data after coordinate transformation based on the registration transformation matrix output in Step Five and the standard point cloud data. The constraint correction process includes: The local geometric constraints include orthogonal constraints, parallel constraints, dimensional constraints, and topological connection constraints; orthogonal constraints: determine whether the mating surfaces of beams, columns, wall panels, and mechanical interfaces are perpendicular to the reference coordinate system (e.g., 90°); parallel constraints: determine whether the splicing seams of the same module and the routing of electrical lines remain parallel; dimensional constraints: determine whether the gaps of mechanical interfaces, the cross-sectional dimensions of beams and columns, and the contour dimensions of spatial interfaces are within the preset tolerance range; topological connection constraints: determine whether the relative positional relationship between the module interface and beams, columns, and walls conforms to the design topological logic; The actual geometric features (such as actual angle values, actual spacing values, and actual positional relationships) of key building structures are extracted from the finely registered point cloud data of the construction site, and the deviations are compared with the local geometric constraints. The obtained geometric constraint deviations are weighted to obtain a comprehensive geometric constraint deviation value. For example, if the interface spacing deviation exceeds 5mm, a high weight such as 0.4 is given; if the deviation is only 2mm, a low weight such as 0.1 is given. With the goal of minimizing the comprehensive value of the geometric constraint deviation, a constraint objective function is constructed. The initial registration transformation matrix is ​​then iteratively optimized using the gradient descent method until the comprehensive value of the geometric constraint deviation meets the preset geometric constraint deviation threshold, and the registration transformation matrix is ​​output.

[0036] In this embodiment, a lightweight convolutional neural network simplifies the network structure and reduces the number of convolutional layers and parameter scale, thereby reducing computational consumption while ensuring feature extraction accuracy. This adapts to lightweight deployment requirements and enables real-time registration on mobile devices or edge nodes. Initial registration provides a stable initial pose for subsequent fine registration operations, avoiding slow convergence and getting stuck in local optima. Feature matching pair screening focuses on eliminating abnormal matching pairs caused by point cloud noise, site occlusion, and feature similarity, reducing the interference of abnormal matching pairs on registration accuracy. Constraint corrections are used to ensure that the registration results conform to the building structure logic. The method of this application achieves adaptive, high-precision registration between construction site point cloud data and standard point cloud data. The generated registration transformation matrix includes the actual spatial coordinates and three-dimensional shape of each interface area on the construction site, allowing the interface adaptation parameters to be accurately mapped to the corresponding physical interface positions on the construction site, better reflecting the actual conditions of the construction site.

[0037] It is also necessary to add that, based on the rotation parameters of the registration transformation matrix, the spatial angle of the original coordinates of each point in the preprocessed construction site point cloud data is adjusted; according to the scaling parameters of the registration transformation matrix, the distance between the point and the origin of the coordinate system is adjusted to adapt to the scale of the standard point cloud data; according to the translation parameters of the registration transformation matrix, the spatial position of the point is adjusted so that the construction site point cloud data as a whole is aligned with the coordinate system of the standard point cloud data; the coordinate integration of the transformed points is the registered construction site point cloud data.

[0038] Furthermore, local geometric constraints are introduced to correct the initial registration transformation matrix. The specific process is as follows: First, geometric constraint rules for the core structure of the building are extracted from standard point cloud data to form a constraint threshold library. Then, the initial registration transformation matrix is ​​applied to the preprocessed construction site point cloud data, and the deviation between the geometric features (angles, spacing, dimensions) of the core structure and the constraint threshold library is calculated. Based on the deviation, the rotation, translation, and scaling parameters of the initial registration transformation matrix are adjusted in reverse. For example, if the mechanical interface docking angle deviation is 3°, the rotation parameters of the matrix are adjusted to return the angle to 90°. After correction, the geometric feature deviation is recalculated until the constraints are met. Without local geometric constraints, the initial registration transformation matrix may only achieve overall mathematical coordinate alignment of the point cloud, but the geometric relationship of the core structure of the building deviates from the actual design. Introducing geometric constraints can limit the registration deviation to within the allowable tolerance range for construction, providing accurate data for subsequent lightweight modeling and actual assembly.

[0039] Preferably, to adapt to the specific characteristics of modular housing construction scenarios and solve the registration deviation problem caused by point cloud jitter and occlusion during construction, the lightweight adaptive point cloud registration algorithm is improved for specific scenarios. The step of performing point cloud registration on the construction site point cloud data using the structural lightweight adaptive point cloud registration algorithm to obtain a registration transformation matrix and registered construction site point cloud data includes: Based on the original step one, jitter compensation is added to the point cloud preprocessing; the jitter compensation means that a jitter offset model is established by collecting equipment vibration data (vibration frequency, amplitude, vibration direction) from vibration sensors on the hoisting equipment, and the spatial offset of the point cloud data at the construction site caused by equipment jitter is calculated to correct the offset of the jittered point cloud data at the construction site. The establishment of the jitter offset model includes: constructing a time-series correlation-based jitter offset model based on collected equipment vibration data and construction site point cloud data. The model takes equipment vibration data as input and point cloud spatial offset as output. It is trained by using historical equipment vibration data and corresponding point cloud spatial offsets to fit the mapping relationship between point cloud spatial offset and point cloud offset; spatial offset calculation, that is, inputting real-time collected equipment vibration data into the jitter offset model, and calculating the spatial offset (including offset values ​​in the X, Y, and Z coordinate axes) of the construction site point cloud data of the corresponding frame due to equipment jitter by combining the collection timestamp of the corresponding frame point cloud; offset correction, that is, for each point in each frame point cloud, performing reverse compensation calculation according to the offset value of the corresponding coordinate axis to restore the true spatial position of the point cloud; after the correction is completed, the corrected point cloud is initially verified, and point clouds that still have abnormalities after correction are removed. In step two, the temporal feature capture capability of the lightweight convolutional neural network is optimized by adding a temporal convolutional layer between the local feature extraction layer and the feature fusion layer of the lightweight convolutional neural network. That is, the lightweight convolutional neural network also includes a temporal convolutional layer, and the processing of the temporal convolutional layer includes: Extract the static local spatial geometric feature vector of a single frame from the local feature vector of the construction site, input it into the temporal convolutional layer and use the convolution kernel (the convolution kernel size is set to 3 and the stride is set to 1) to perform temporal correlation operation, capture the feature change trend of adjacent frames, filter jitter noise and compensate for inter-frame offset, and output the temporal correlation features of the site. Among them, the extracted vectors are the core feature subsets that characterize the spatial geometric properties of a single frame point cloud, such as the outline size of the interface, the angle / length of the beams and columns, and the position offset of the point cloud in the interface area. The on-site temporal correlation features are fused with the single-frame static local spatial geometric feature vector to generate an on-site dynamic feature vector that represents the dynamic state of point clouds under jitter. The fused features, which include the on-site dynamic feature vector and the single-frame static local spatial geometric feature vector, are used to enhance the key structures of the modular house through an introduced attention mechanism to obtain the on-site point cloud enhanced feature vector. During feature enhancement, the corresponding feature dimensions are given higher weights (e.g., the interface feature weight is increased to 0.8, and the ordinary wall feature weight is 0.2), while the feature weights of noise / occlusion areas are weakened to obtain the corresponding point cloud enhanced feature vector. Based on the original step five, the adaptive step size optimization strategy is adjusted, that is, the iteration step size and the number of iterations are dynamically adjusted by combining the adaptive compensation optimization mechanism. The specific process is as follows: Based on the equipment vibration data and the preprocessed point cloud data of the construction site, shake-sensitive areas with large vibration amplitude and easy point cloud shift are identified, and non-shake-sensitive areas with small vibration amplitude and easy point cloud shift are identified. Shake-sensitive areas include modular house interface areas, beam-column node areas, etc., and the corresponding point cloud coordinate range is marked. Input the amplitude of the jitter-sensitive area into the preset mapping table and query the jitter weight coefficient: the larger the jitter amplitude of the area, the larger the jitter weight coefficient is set to strengthen the registration accuracy constraint of the key area; the jitter weight coefficient of the non-jitter-sensitive area is set to 1.0; the mapping table is established based on the vibration amplitude and the proportion of the point cloud data coordinate range in historical time periods. In each iteration, the local registration error of the jitter-sensitive region and the local registration error of the non-jitter-sensitive region are multiplied by the corresponding jitter weight coefficient and then summed to obtain the registration error. Based on the registration error and its rate of change, the adjustment amounts for the iteration step size and the number of iterations are determined and adjusted until the iteration termination condition is met, i.e., the number of iterations reaches a preset number (e.g., 80 times), or the weighted correction registration error is no greater than ±1.5cm. If either condition is met, the iteration terminates. If the registration error still does not meet the standard after reaching the maximum number of iterations, the iteration step size is readjusted based on the current error state, and fine registration is performed again. If the difference in registration error between two consecutive iterations is less than a preset threshold (e.g., 0.1mm), it is determined that the error has stabilized and is no longer decreasing significantly, and the iteration can be terminated early to avoid unnecessary iterations consuming computational power.

[0040] Specifically, if the rate of change of the registration error in this iteration is large, such as exceeding 5%, maintain the current iteration step size and appropriately reduce the number of iterations to accelerate the convergence rate; if the rate of change of the registration error is small, such as not exceeding 1%, or if the registration error is large, dynamically increase the iteration step size and appropriately increase the number of iterations to ensure that the registration accuracy gradually converges.

[0041] The remaining steps are consistent with the original point cloud registration process. Through the above improvements, stable registration under hoisting vibration scenarios can be achieved, adapting to the real-time registration requirements of the hoisting process.

[0042] In this embodiment, in jitter scenarios such as hoisting, the point cloud data at the construction site is constantly changing due to equipment jitter. Jitter compensation helps restore the true spatial position of the point cloud and reduces the impact of jitter on the point cloud quality. The newly added temporal convolutional layer does not affect the core feature extraction function of the lightweight convolutional neural network, and can efficiently perform temporal correlation processing on local features to effectively solve the point cloud offset and blurring problems caused by jitter scenarios. In addition, even after jitter compensation, there are still small dynamic offsets, which cause the Euclidean distance error of each ICP iteration to fluctuate greatly, and the convergence speed is slower than in static scenarios. If the basic number of iterations is maintained, it is easy to have insufficient iterations and the registration error cannot reach the expected result. Increasing the number of iterations can improve the probability of the algorithm escaping local optima and meet the real-time registration and construction accuracy requirements of jitter scenarios such as hoisting. Furthermore, after introducing the jitter weight coefficient, the registration error convergence requires more iteration steps to gradually approach the optimal solution, ensuring the stability and accuracy of the registration transformation matrix. Overall, the point cloud registration algorithm maintains its lightweight characteristics, is compatible with mobile devices / edge node deployment, does not require additional computing power costs, and balances real-time performance and accuracy requirements.

[0043] Preferably, to address the issue of missing point cloud data and incomplete features in the occluded area due to obstructions on the module surface during construction (i.e., scaffolding, construction equipment, or personnel), which affects registration accuracy, the lightweight adaptive point cloud registration algorithm is improved for specific scenarios. The step of performing point cloud registration on the construction site point cloud data using the structural lightweight adaptive point cloud registration algorithm to obtain a registration transformation matrix and registered construction site point cloud data includes: Based on the original step one, occlusion recognition and completion are added to the point cloud preprocessing. The execution process is as follows: An occlusion recognition reference dataset is constructed based on point cloud data from the construction site and standard point cloud data. The dataset includes information that is prone to occlusion, such as the location of scaffolding, the coordinates of construction equipment, and the work area of ​​construction workers. The coordinate-normalized point cloud data from the construction site is input into a lightweight semantic segmentation algorithm (such as SqueezeSegV3) that incorporates occlusion feature parameters for semantic segmentation. The initial occluded point cloud region is identified and marked. The initial occluded point cloud region is then calibrated using the occlusion recognition parameter dataset to obtain the occluded point cloud region. Specifically, an occlusion feature enhancement module is added to the lightweight semantic segmentation network, incorporating exclusive morphological feature parameters of scaffolding (pole-like and grid-like features), construction equipment (irregular block-like and rigid features), and construction workers (flexible and dynamic point-like features) to construct a feature template library for three types of occlusion. This improved algorithm accurately identifies occluded point cloud regions and marks their coordinate range, occlusion type, and occlusion area through feature template matching and local feature extraction. By comparing the coordinate range of key building structures in standard point cloud data, misidentified key building structure regions are eliminated, ensuring the accuracy of occlusion identification and avoiding misjudging core structures such as interfaces and beams as occluded areas. Differential completion operations are performed based on the occlusion type of the occlusion point cloud region to generate preprocessed point cloud data of the construction site. The differential completion operations include: If the occlusion is caused by a person and the occlusion duration is less than the preset occlusion duration (e.g., less than 30 seconds), then the occlusion is completed using temporal interpolation based on the unoccluded point cloud data of the construction site in the adjacent frame; if the occlusion is caused by a person and the occlusion duration is not less than the preset occlusion duration, i.e., the dwell time is relatively long, then the occlusion is completed based on the standard point cloud data corresponding to that area. If the obstruction is caused by mobile construction equipment (temporary or rigid obstruction), the obstruction is completed based on the point cloud data of the construction site and the standard point cloud data. That is, the size, shape and placement angle of the equipment are obtained from the point cloud data of the construction site, the background point cloud features of the equipment in the unobstructed state are extracted from the continuous frame point cloud data, and the geometric features of the corresponding area in the standard point cloud data are combined with the rigid feature removal and background point cloud reconstruction to complete the point cloud of the obstructed area, ensuring that the completion result matches the working conditions around the equipment. If the obstruction is caused by fixed construction equipment (fixed or grid-like obstruction), the data is supplemented based on standard point cloud data. For example, if the obstruction is caused by scaffolding, the key structural features of the building in the corresponding area are extracted from the standard point cloud data based on the scaffolding's erection parameters (pole diameter, spacing, erection height). The method of mesh feature stripping + key structure point cloud reconstruction is adopted. First, the grid-like point cloud features of the scaffolding are stripped. Then, based on the geometric constraints of the standard point cloud, the key structural point cloud of the building obstructed by the scaffolding is supplemented to ensure that the accuracy of the supplemented key structures (interfaces, beams and columns) meets the construction requirements. Based on the original step two, the weight allocation of the attention mechanism is adjusted, focusing on enhancing the feature weights of occluded point cloud regions and key structures. Specifically, a feature fusion layer is added after the local feature output layer of the original lightweight convolutional neural network and before the feature enhancement layer of the attention mechanism. The processing includes: Min-Max normalization is used to standardize both the local feature vectors of the construction site and the standard local feature vectors, unifying their numerical ranges and avoiding fusion bias caused by different numerical magnitudes. This results in standardized local feature vectors for the construction site and standard local feature vectors. Adaptive weighting coefficients are introduced to weight and fuse the two types of vectors. For example, the feature variance of the standardized local feature vectors (reflecting the stability of local features in the construction site point cloud; smaller variance indicates more stable features and higher weights) and the structural confidence of the standardized standard local feature vectors are calculated. The ratio of the feature variance to the sum of the two is used as the weighting coefficient for the former, and the ratio of the structural confidence to the sum of the two is used as the weighting coefficient for the latter. This yields a fused feature vector that retains the real-time features of the construction site point cloud while strengthening the standard structural features, effectively compensating for feature loss in occluded areas and improving the completeness and reliability of the features.

[0044] In step three, coarse registration is performed based on the fused feature vector and the standardized local feature vector.

[0045] The remaining steps are consistent with the original point cloud registration process. Through the above improvements, the problem of missing point clouds caused by occlusion is solved, ensuring the accuracy and completeness of the registration results in occluded scenarios.

[0046] In this embodiment, occlusion recognition and completion are added to ensure the integrity of the point cloud data and avoid registration failures caused by missing point clouds. The newly added feature fusion layer does not affect the core local feature extraction function of the lightweight convolutional neural network on the original point cloud (including the completed point cloud), and can efficiently fuse standard point cloud features and completed point cloud features, making up for the feature defects in occluded point cloud regions and improving the accuracy of feature matching. The method of this application enables high-precision registration of point clouds without removing occluders, improving construction efficiency.

[0047] Preferably, to address the impact of terrain undulations on point cloud registration in complex terrain scenarios, the lightweight adaptive point cloud registration algorithm is improved for specific scenarios. The step of performing point cloud registration on the construction site point cloud data using the structural lightweight adaptive point cloud registration algorithm to obtain a registration transformation matrix and the registered construction site point cloud data includes: Based on the original step one, terrain feature extraction is added to the point cloud preprocessing. The terrain feature extraction includes: The region growing algorithm is used to segment the filtered point cloud data of the construction site and the standard point cloud data, distinguish the terrain point cloud (site ground point cloud) from the building component point cloud (modular house related point cloud), mark the coordinate range of the terrain point cloud, and extract the site terrain feature data and standard terrain feature data, including slope, undulation and elevation. The offset between the on-site terrain feature data and the standard terrain feature data is calculated. If the offset does not meet the preset offset qualification conditions, the coordinates of the filtered construction site point cloud data are compensated in reverse based on the offset. The point cloud data of complex terrain is corrected to a coordinate system parallel to the BIM model terrain reference plane. Otherwise, no additional processing is performed. The preset offset qualification conditions are usually defined as slope offset not exceeding 3°, undulation offset not exceeding 5cm, and elevation offset not exceeding 3cm. These conditions can be adjusted according to the actual working conditions of special terrain to ensure that the compensated terrain features are consistent with the standards.

[0048] Based on the original step five, the local geometric constraints are optimized. That is, local geometric constraints with superimposed terrain constraints are introduced to constrain and correct the initial registration transformation matrix, ensuring that the registration result conforms to the actual terrain and the logic of the building structure. At the same time, the iteration step size change rate is dynamically adjusted in combination with an adaptive compensation optimization mechanism. That is, it is dynamically adjusted according to the degree of terrain undulation. In areas with large terrain undulation, i.e., "steep slopes", the step size change rate is increased to accelerate the convergence speed, while in areas with gentle terrain, the step size change rate is decreased to improve the registration accuracy. The specific process is as follows: obtain the deviation between the extracted undulation degree and the standard undulation degree, look up the iteration step size change rate in the established change rate mapping table based on the deviation, and adjust the iteration step size based on the iteration step size change rate. The change rate mapping table is established in advance by a set of personnel, and the mapping table establishes the relationship between the relevant deviations and the iteration step size change rate. The processing of terrain constraints includes: recording the registration transformation matrix after local geometric constraints as the intermediate registration transformation matrix, calculating the terrain feature data of the point cloud after registration based on the matrix, comparing the calculated terrain feature data with the preset constraint threshold, and if both meet the constraint requirements, then the intermediate registration transformation matrix is ​​used as the output result of step five; otherwise, differential constraint correction is performed. The differentiated constraint correction includes: if the slope constraint threshold is exceeded, adjusting the rotation parameters of the intermediate registration transformation matrix to gradually correct the slope of the building component point cloud until the slope meets the preset constraint threshold; if the elevation constraint threshold is exceeded, adjusting the overall translation parameters of the intermediate registration transformation matrix to correct the elevation of the building component point cloud until the elevation meets the preset constraint threshold; if the undulation constraint is exceeded, adjusting the local translation parameters of the intermediate registration transformation matrix to fine-tune the local elevation of the building component point cloud. The above constraint thresholds are set by preset personnel, such as a module interface slope not exceeding 2°, beam / column verticality not exceeding 1°, and elevation deviation not exceeding ±3cm. The overall translation parameters apply to the entire construction site point cloud (all building component point clouds), achieving overall translation adjustment of the point cloud in the elevation direction. The local translation parameters are translation components set for local areas of the point cloud (only local terrain areas exceeding the undulation constraint threshold), acting only on the building component point cloud in specific local areas (such as building component areas corresponding to local terrain protrusions or depressions), without affecting other areas of the point cloud.

[0049] If two parameters exceed the threshold, determine the adjustment amount of each parameter in the order of adjustment. After adjusting the previous parameter, the deviation of the next parameter needs to be recalculated and adjusted. The adjustment order is slope, elevation, and undulation. After the overall slope and elevation are corrected, perform fine-tuning of the Z-axis translation in the local area to ensure that the building components conform to the terrain as a whole and can adapt to local undulations, avoiding the impact of local adjustments on the overall posture.

[0050] The remaining steps are the same as the original point cloud registration. Through the above improvements, the registration requirements of complex terrain scenes are adapted to ensure that the registration results fit the site conditions of complex terrain.

[0051] In a specific embodiment, before the original step three, terrain feature matching is added, that is, based on the terrain feature parameters, the point cloud data of the construction site and the standard point cloud data are matched until the corresponding initial attitude deviation meets the attitude deviation condition. Specifically, a feature point matching algorithm is used to select key feature points, such as slope abrupt change points and elevation extreme points, from the two types of terrain point cloud data after downsampling to construct terrain feature matching pairs. Valid matching pairs are selected and abnormal matching pairs are eliminated by calculating the Euclidean distance and the angle between the normal vectors of the key feature points. Based on the valid matching pairs, the initial attitude adjustment matrix of the two types of point cloud data is calculated using a random sampling consensus algorithm. The initial attitude deviation of the two types of point clouds is calculated to determine whether the attitude deviation conditions are met, such as translation deviation not exceeding 8cm and rotation deviation not exceeding ≤3°.

[0052] In this embodiment, by adding terrain feature extraction and matching, the problems of uneven point cloud distribution, difficult initial registration, and slow convergence of fine registration caused by large slopes and significant topographic undulations in complex terrains (mountains and hills) are solved, significantly improving the efficiency and accuracy of initial registration. The newly added slope and elevation constraints, combined with adaptive step size optimization, ensure that the registration results conform to both the building structure logic and the actual conditions of complex terrain. The entire point cloud registration process can achieve high-precision registration without large-scale leveling of complex terrain, supporting the precise construction of modular houses and the calibration of digital twin models in complex terrain.

[0053] In addition, to address the issues of limited computing power and small storage capacity of mobile devices (phones, AR glasses), model quantization and pruning techniques are used to perform lightweight compression on the lightweight point cloud registration algorithm of the structure, reducing the amount of computation and storage. The collected point cloud data from the construction site is processed in segments, processing only the data of the current area to reduce the amount of data processing. At the same time, historical data is cached in the cloud, and mobile devices only retain the point cloud data required at present, saving storage space.

[0054] This application also provides a digital twin model construction system for collaborative management of building construction, and the method for constructing a digital twin model for collaborative management of building construction includes: The design data acquisition module is used to acquire standard design data for building construction. The standard design data includes standardized component library data and external input requirement data. The standardized component library data consists of predefined building modules and their parameterized attribute data. The external input requirement data reflects the construction requirements of the building construction participants. The design scheme and interface standard formulation module is used to generate a parametric design scheme based on the standard design data through a parametric modeling algorithm, and at the same time formulate a dynamic reconfigurable interface standard for the preset interface supported by the dynamic interface generator. The interface parameter adaptive adjustment module is used to adaptively adjust the interface parameters based on the parameterized design scheme and the dynamic reconfigurable interface standard by using a dynamic interface generator that integrates a dynamic graph neural network interface adaptation algorithm to obtain interface adaptation parameters. The digital twin model construction module is used to integrate standard point cloud data extracted from the BIM model with the collected point cloud data from the construction site, and to perform lightweight modeling processing by combining the interface adaptation parameters with the LOD hierarchical method and the lightweight adaptive point cloud registration algorithm to obtain a digital twin model that is synchronized with the physical entity of the building construction.

[0055] In this embodiment, the matching accuracy between the digital twin model and the physical construction entity is effectively improved. Lightweight modeling and efficient point cloud registration significantly improve modeling efficiency and reduce computing power consumption. Based on this model, industry chain participants can achieve collaborative planning, dynamic monitoring, and precise guidance for the entire process, including construction design schemes, component production, and on-site assembly. This enhances the collaborative efficiency and control capabilities of the entire construction process and facilitates intelligent and refined management of the construction process.

[0056] In one specific embodiment, see Figure 3 By applying the digital twin model construction method for collaborative management of housing construction, a modular housing industry chain collaborative system that is perceptive, learnable, collaborative, and optimizable is constructed, providing a complete system-level solution for the development of modular buildings towards personalization, efficiency, and safety. Specifically, the modular housing industry chain collaborative system includes a digital twin modeling unit, an industry chain collaborative platform, and an intelligent decision-making unit.

[0057] The digital twin modeling unit is used to build and update digital twin models.

[0058] Specifically, the supply chain collaboration platform employs a blockchain-based distributed and microservice architecture to execute data processing flows that enable secure cross-enterprise data sharing, end-to-end data interaction, and business module collaboration. This platform includes a design collaboration module, a production scheduling module, a logistics tracking module, an assembly guidance module, and a privacy protection module. The design collaboration module supports online collaborative editing and version control among supply chain participants, enabling collaborative optimization and efficient iteration of parametric design schemes. These participants include design institutes, component factories, logistics companies, and construction units. The production scheduling module dynamically optimizes factory production plans based on real-time data from digital twin feedback, improving production efficiency and capacity utilization. The logistics tracking module combines RFID and blockchain technology to achieve full lifecycle traceability of components from production and transportation to assembly, ensuring traceable component quality and controllable processes. The assembly guidance module generates AR-based construction guidance instructions to assist on-site construction personnel in completing module assembly, improving assembly accuracy and efficiency.

[0059] Specifically, the intelligent decision-making unit includes a speed prediction module, a resource optimization module, an anomaly detection module, and an energy consumption optimization module. The schedule prediction module predicts the critical path of the project based on an LSTM neural network. Input data includes collected time-series data on project task duration, dependencies, and resource usage. After feature encoding and model inference, it outputs a prediction of the total project duration and critical nodes, providing a basis for schedule adjustments. The resource optimization module uses a multi-objective genetic algorithm to solve for the optimal configuration of equipment, manpower, and materials. It uses equipment utilization, labor costs, and material losses as optimization objectives, encoding equipment, manpower, and material configurations as chromosomes to generate an initial population. After iterative optimization, it outputs the optimal resource configuration scheme. The anomaly detection module identifies construction risks through deviation analysis between digital twins and actual data. It compares collected real-time status data with standard evaluation thresholds to determine the existence of construction risks. The energy consumption optimization module dynamically adjusts the building operation strategy based on reinforcement learning. It constructs a state vector based on equipment status, personnel configuration, material inventory, and environmental parameters. The operation strategy is output through a reinforcement learning agent and continuously optimized using the lowest energy consumption as the reward function. The operation strategy is iteratively updated based on real-time status data.

[0060] In one specific embodiment, the digital twin model construction module further includes: a multiphysics modeling and simulation module for performing coupled analysis of structural mechanics, thermal performance, and acoustic environment, and using a lightweight adaptive point cloud registration algorithm to fuse and generate high-precision point cloud data; and a scheme pre-drilling module for visually displaying the construction process and operation and maintenance effects of different parametric design schemes based on VR / AR technology.

[0061] It should be added that the intelligent decision-making unit constructs a federated reinforcement learning topology architecture and configures a local model training module, a global model aggregation module, and a simulation verification module. The federated reinforcement learning topology uses a digital twin modeling unit as the core aggregation node, industry chain participants as hierarchical collaborative nodes, and a multiphysics simulation module as the simulation verification node, each responsible for parameter transfer and aggregation, local model training, and policy verification, respectively; the specific process is as follows: The local model training module is used to issue training instructions to each hierarchical collaborative node, control each hierarchical collaborative node to train and evaluate the model on local private data (synchronously provided by the industry chain collaborative platform), generate local model parameters containing policy gradients and upload them to the core aggregation node. The global model aggregation module is used to receive local model parameters of each hierarchical collaborative node relayed by the core aggregation node using a secure multi-party computation and secure aggregation protocol. After aggregating the node weights determined by the adaptive weight adjustment mechanism, it generates a global optimization strategy model and sends it to the simulation verification module. The simulation verification module is used to perform full-process simulation verification on the received global optimization strategy model, determine whether to trigger the retraining of each hierarchical collaborative node, and send the global optimization strategy model to the modular design unit, digital twin modeling unit and industry chain collaborative platform after the global optimization strategy model converges. The node weights are obtained through a method that includes: acquiring node evaluation indicators for each hierarchical collaborative node; and then, through a weighted fusion (i.e., multiplying and summing) of these indicators by the core aggregation node to obtain the node weights for each hierarchical collaborative node. The node evaluation indicators include data quality indicators and role importance indicators. The data quality indicators include data integrity (e.g., module pass rate data for production nodes, scheme compliance data for design nodes), data real-time performance (e.g., transportation location data for logistics nodes, construction progress data for assembly nodes), and data accuracy (e.g., the accuracy of real-time status data collected by sensors), with a corresponding node evaluation weight typically of 0.6. Role importance indicators are set by pre-defined personnel based on the core importance of the participants in the industry chain. Generally, design institutes and construction units are set to 0.3, and component factories and logistics companies are set to 0.2, with a corresponding node evaluation weight of 0.4. All of the above data are set by pre-defined personnel and are adjustable. If the data quality indicators of a hierarchical collaborative node improve, such as an increase in the module pass rate or improved data integrity of a production node, the node weight is automatically increased, giving its model parameters a higher influence in the global aggregation; conversely, it is decreased.

[0062] In a specific embodiment, based on the above-mentioned federated reinforcement learning topology, a dual privacy protection mechanism is added. That is, before each hierarchical collaborative node uploads its local model parameters, it performs noise-adding and desensitization processing on the local model parameters to avoid inferring the original data from the local model parameters. The core aggregation node performs hierarchical control over the access permissions of each hierarchical collaborative node, and each hierarchical collaborative node can only obtain global policy parameters related to its own business.

[0063] It is important to understand that the aforementioned federated reinforcement learning topology and the microservice architecture, blockchain distributed architecture, and collaborative work of various business modules of the industry chain collaboration platform not only retain the core functions of data security sharing and cross-enterprise collaboration of the original platform, but also improve the accuracy and efficiency of collaborative optimization, achieving a dual improvement in privacy protection and global collaborative optimization, and adapting to the full-process collaborative needs of the modular housing industry chain.

[0064] In summary, this application, in the collaborative management of this industry chain, constructs a federated reinforcement learning topology, combines blockchain technology with a differential privacy protection mechanism, and enables industry chain participants to conduct full-process collaborative optimization without disclosing original business data. The adaptive adjustment mechanism of node weights ensures the accuracy of global collaborative optimization, and the closed-loop simulation verification mechanism prevents optimization strategies from deviating from reality. This breaks down information silos in the industry chain, improves collaborative efficiency, reduces the risk of data leakage, and ensures the security and reliability of data sharing among multiple parties in the industry chain.

[0065] Example 1: In the event of a public health emergency (such as an outbreak or public health emergency response), the rapid deployment, commissioning, and use of modular medical units must be completed within 72 hours. The system and method provided in this application are used, with the following specific configuration: Module configuration: It adopts standardized medical modules with a 95% prefabrication rate, including three functional units: treatment area, isolation area, and equipment area. Each module is equipped with a dynamic reconfigurable interface to adapt to rapid assembly requirements. Computing power and sensing configuration: Deploy 5G edge computing nodes to achieve real-time processing of on-site point cloud data and real-time window data, with a processing latency of <50ms, and support the real-time operation of lightweight point cloud registration algorithms; Equip a drone swarm (3 drones per square kilometer) for site terrain scanning and modeling, and collect high-precision on-site point cloud data; Collaboration and Simulation Configuration: Build an industry chain collaboration platform and use a federated reinforcement learning topology to achieve distributed collaborative optimization; enable the multiphysics simulation module of the digital twin modeling unit to carry out structural stability simulation analysis.

[0066] The specific implementation process is as follows: 1) Dynamic Design and Interface Adaptation Phase (0-6 hours): Through the design collaboration module of the industry chain collaboration platform, input site GIS data and emergency medical function requirements, and automatically generate an elevated foundation scheme that conforms to the terrain slope based on a standardized component library; through the scheme pre-drilling module, simulate the medical process, optimize the module layout, and reduce the crossover rate of medical staff and patients to below 2%; combined with the dynamic interface generator, generate a module connection scheme with seismic nodes such as seismic resistance level 8, and preset interface locking torque and angle adjustment parameters to ensure subsequent assembly compatibility.

[0067] 2) Digital Twin Modeling and Real-time Synchronization Stage (6-12 hours): Launch a drone swarm to perform a comprehensive scan of the deployment site and collect high-precision point cloud data; use the optimized lightweight adaptive point cloud registration algorithm of this application to preprocess the collected point cloud data (remove jittery points and fill in occluded areas), perform high-precision registration between the BIM model and the actual point cloud, automatically adjust the coordinates of the module positioning points, and construct a high-precision digital twin model of the site; through the multiphysics simulation module, simulate the stability of the cabin structure under different extreme weather conditions (wind speed 15m / s, rainfall 50mm / h) and predict potential construction risks.

[0068] 3) Federated Collaborative Production and Logistics Phase (12-48 hours): Based on a federated reinforcement learning topology, each hierarchical collaborative node trains a production scheduling and logistics optimization model on local data. Through a secure aggregation protocol, parameter sharing and global model aggregation are achieved to generate the optimal production and logistics plan. The distributed factory simultaneously starts the production of standardized medical modules. Through blockchain technology combined with RFID tags, the module quality is traced throughout the entire process. Intelligent transport vehicles deliver according to the optimal route planned by the digital twin model. After route optimization, transportation time is saved by 18%, ensuring that the modules arrive at the deployment site on time.

[0069] 4) AR-assisted assembly and closed-loop optimization phase (48-72 hours): Construction personnel wear AR glasses and use the assembly guidance module of the industry chain collaboration platform to visualize the module hoisting path and interface docking benchmarks. During the assembly process, the dynamic interface generator collects real-time status data and automatically adjusts interface parameters through the interface adaptation algorithm. It verifies the electrical interface connection status, such as the voltage detection response time of 0.2s, to ensure that the interface is successfully adapted on the first try. The construction progress and assembly quality data are updated to the command center's large screen in real time and simultaneously fed back to the digital twin model to achieve real-time monitoring and closed-loop optimization of the assembly process, ensuring that all deployments are completed within 72 hours.

[0070] The technological benefits include: First, significantly improved deployment efficiency: Deployment time is reduced by 67% compared to traditional emergency building deployments, successfully achieving rapid deployment, commissioning, and use of the modular unit within 72 hours, meeting emergency response requirements. Second, improved assembly precision and reliability: Relying on interface adaptation algorithms, the first-time installation qualification rate of module interfaces reaches 99.8%, significantly reducing rework rates and ensuring rapid deployment of the modular unit. Third, optimized operation and maintenance energy consumption: Real-time monitoring of the air conditioning system's operating status through a digital twin model optimizes operating parameters, resulting in a 22% reduction in operation and maintenance energy consumption after completion, achieving energy conservation and consumption reduction. Fourth, improved collaborative efficiency: Through federated reinforcement learning topology and blockchain technology, collaboration across the entire process of design, production, logistics, and assembly is achieved, with smooth connections between each link, eliminating information silos, and improving overall collaborative efficiency by over 40%.

[0071] Example 2: With the acceleration of urbanization, the demand for high-rise residential construction is increasing. Traditional high-rise prefabricated residential buildings suffer from problems such as low module hoisting efficiency, poor load adaptability, high risk of multi-disciplinary collisions, and difficulty in collaborative construction. To address these pain points, this application aims to achieve efficient construction of a 32-story prefabricated residential building with 28 different module types and a maximum single module weight of 12 tons, balancing structural safety and construction efficiency. The system and configuration provided in this application are as follows: Sensing and Adaptive Configuration: Pressure sensors are embedded at all module connections to collect interface stress and load data in real time; a dynamic interface generator is configured to enable real-time adaptive adjustment of module interfaces with a response time of <5s; Lifting and Modeling Configuration: Deploy a tower crane (lifting capacity 16t, working radius 45m) and equip it with a drone swarm for on-site point cloud acquisition; adopt a lightweight adaptive point cloud registration algorithm to achieve real-time registration between the BIM model and the on-site point cloud, supporting lifting path optimization and collision detection; Collaboration and Simulation Configuration: Build a cross-enterprise collaborative platform to connect design institutes, module manufacturers, logistics companies, and construction units, and use a federated reinforcement learning topology to achieve collaborative optimization; enable the multiphysics simulation module to carry out simulations of load distribution, structural stability, and multi-disciplinary collision detection.

[0072] The specific implementation method is as follows: 1) Pressure sensors are implanted at the connection points of each module to collect load data in real time during the construction phase and later use (sampling frequency 10Hz) and transmit it synchronously to the digital twin model; the digital twin modeling unit calculates the overall load distribution of the building in real time, and automatically adjusts the module support points and interface locking force in combination with the interface adaptation algorithm.

[0073] 2) Point cloud data of the construction site is collected by a drone swarm. A lightweight adaptive point cloud registration algorithm is used to correct point cloud jitter deviations during hoisting, accurately acquiring data on tower crane position, module position, and on-site obstacle distribution. Tower crane performance parameters (lifting capacity 16t, working radius 45m), module parameters (weight, dimensions), and on-site environmental parameters (wind speed, terrain) are input into the hoisting optimization algorithm. Combined with digital twin simulation, considering wind speed limitations (operation is suspended when >10m / s) and module positioning accuracy requirements (±5mm), the optimal hoisting sequence and hoisting path are generated. During construction, the hoisting path is visualized through AR glasses, and the module positioning deviation is compared in real time to guide hoisting operations, reducing the number of tower crane relocations by up to 40%.

[0074] 3) Based on the digital twin modeling unit, construct a digital twin model that includes multiple disciplines such as architecture, structure, and electromechanical systems; simulate the installation process of electromechanical pipelines, structural components, and module interfaces; and discover various conflict points in advance and automatically generate avoidance solutions through the coupled analysis of multiphysics simulation modules.

[0075] 4) Real-time collaboration between design, production, and construction is achieved based on a federated reinforcement learning topology.

[0076] The technical benefits include four aspects: First, it enables real-time monitoring and dynamic adjustment of the load, ensuring no safety accidents occur during the construction phase; second, it reduces the number of tower crane relocations by 40%, shortens the single module hoisting time by 25%, and improves the overall hoisting efficiency by 30%, ensuring that the construction cycle is controlled within 18 months; third, it avoids 23 conflict points in advance, reduces the rework rate by 90%, achieves module positioning accuracy of ±5mm, and achieves a 99.5% first-time assembly qualification rate for interfaces.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for constructing a digital twin model for multi-party collaboration in housing construction, characterized in that, Includes the following steps: Obtain standard design data for housing construction. The standard design data includes standardized component library data and external input requirement data. The standardized component library data consists of predefined building modules and their parameterized attribute data. The external input requirement data reflects the needs of the housing construction participants. Based on the standard design data, a parametric design scheme is generated through a parametric modeling algorithm. At the same time, a dynamic reconfigurable interface standard for preset interfaces supported by a dynamic interface generator is established. The dynamic interface generator is used to adjust the connection method of the building module according to the external input requirement data. Based on the parametric design scheme and the dynamic reconfigurable interface standard, the interface parameters are adaptively adjusted by a dynamic interface generator that integrates a dynamic graph neural network interface adaptation algorithm to obtain the interface adaptation parameters. Based on the standard point cloud data extracted from the BIM model and the collected point cloud data from the construction site, the interface adaptation parameters are integrated, and a lightweight modeling process is performed using the LOD hierarchical method combined with a lightweight adaptive point cloud registration algorithm to obtain a digital twin model synchronized with the physical entity of the building construction.

2. The method for constructing a digital twin model for collaborative management of building construction according to claim 1, characterized in that, The preset interfaces supported by the dynamic interface generator include mechanical interfaces, electrical interfaces, and spatial interfaces. The dynamic interface generator supports preset interfaces that integrate multi-dimensional micro sensors. The multi-dimensional micro-sensors include a pressure sensor, a displacement sensor, an angle sensor, and a temperature sensor mounted on the mechanical interface; a signal strength sensor and a voltage and current sensor mounted on the electrical interface; and a laser rangefinder and a contour sensor mounted on the spatial interface.

3. The method for constructing a digital twin model for collaborative management of building construction according to claim 2, characterized in that, The adaptive adjustment of the interface parameters specifically includes: Real-time status data of the construction site is collected through the multi-dimensional micro sensors; The real-time state data is transmitted to the dynamic interface generator and transformed into node feature vectors and edge feature vectors of the dynamic graph neural network. The node feature vectors and edge feature vectors are then fused using a temporal attention mechanism to obtain the fused feature vectors. The fused feature vector is dynamically modeled to generate parameterized adjustment suggestions, which include: suggestions for adjusting locking torque and angle deflection, suggestions for switching electrical communication protocols, and suggestions for adjusting morphological parameters. The parameterized adjustment suggestions are verified, and the verified parameterized adjustment suggestions are sent to various interfaces for adaptive adjustment of interface parameters. The real-time status data collected after adjustment is fed back to the dynamic interface generator for parameter optimization of the interface adaptation algorithm.

4. The method for constructing a digital twin model for collaborative management of building construction according to claim 3, characterized in that, The parameter optimization of the interface adaptation algorithm specifically includes: The received real-time status data is compared and analyzed to generate feedback evaluation results, which are used to determine whether the adaptive adjustment of interface parameters meets the standards. Based on a preset reinforcement learning reward mechanism, the interface installation success rate and the later stability of module connection are used as reward evaluation indicators. The reward value is obtained by combining the feedback evaluation results. The reward value includes positive reward value and negative reward value. The interface success rate is used to measure the accuracy of interface adaptation, and the later stability of module connection is used to measure the long-term reliability of interface adaptation. The obtained reward value is input into the dynamic graph neural network interface adaptation algorithm integrated in the dynamic interface generator, triggering parameter iteration of the dynamic graph neural network interface adaptation algorithm.

5. The method for constructing a digital twin model for collaborative management of building construction according to claim 1, characterized in that, The lightweight modeling process includes: Using the standard point cloud data exported from the BIM model as a reference, the construction site point cloud data is registered using a lightweight adaptive point cloud registration algorithm to obtain the registration transformation matrix and the registered construction site point cloud data. Using the BIM model as the modeling framework, and integrating the interface adaptation parameters, layered lightweight modeling is performed to obtain the initial LOD layered model; Based on the registration transformation matrix, the registered construction site point cloud data and the initial LOD layered model are spatially fused and parameter calibrated to obtain the LOD layered model; By connecting the real-time status data to the LOD hierarchical model through the digital mainline, a real-time data synchronization link between the physical entity and the model is established, thus completing the construction of the digital twin model.

6. The method for constructing a digital twin model for collaborative management of building construction according to claim 5, characterized in that, The point cloud registration of the construction site point cloud data is performed using a lightweight adaptive point cloud registration algorithm to obtain a registration transformation matrix and the registered construction site point cloud data, including: Step 1: Perform point cloud preprocessing on the collected construction site point cloud data and the exported standard point cloud data to obtain construction site point cloud preprocessed data and standard point cloud preprocessed data. The point cloud preprocessing includes point cloud filtering, downsampling and coordinate normalization. Step 2: Input the preprocessed point cloud data of the construction site and the preprocessed point cloud data of the standard point cloud into a lightweight convolutional neural network to extract local features and obtain local feature vectors of the construction site and standard local feature vectors. Then, through the introduced attention mechanism, feature enhancement of the key structure of the building is performed, and the enhanced feature vectors of the site point cloud and the standard point cloud are output. Step 3: Perform coarse registration between the enhanced feature vector of the on-site point cloud and the enhanced feature vector of the standard point cloud to obtain the initial registration transformation matrix and the initial feature matching pair; Step 4: Filter the initial feature matching pairs to obtain valid feature matching pairs; Step 5: Perform fine registration on the effective feature matching and output the registration transformation matrix. The fine registration includes iterative optimization of the initial registration transformation matrix based on Euclidean distance, dynamic adjustment of the iteration step size combined with an adaptive compensation optimization mechanism, and constraint correction of the initial registration transformation matrix by introducing local geometric constraints. Step 6: Perform accuracy verification on the registration result and determine whether the obtained registration error meets the threshold condition. If it does, output the registration transformation matrix and perform three-dimensional spatial coordinate transformation on the preprocessed construction site point cloud data based on the registration transformation matrix to generate the registered construction site point cloud data. Otherwise, re-perform accurate registration until the registration error meets the condition. The local geometric constraints include orthogonal constraints, parallel constraints, dimensional constraints, and topological connectivity constraints; The constraint correction process includes: extracting the actual geometric features of the key building structures from the finely registered point cloud data of the construction site, comparing the deviation with the local geometric constraints, and then assigning weights to obtain a comprehensive value of geometric constraint deviation. A constraint objective function is constructed with the goal of minimizing the comprehensive value of geometric constraint deviation. The initial registration transformation matrix is ​​iteratively optimized based on the gradient descent method until the comprehensive value of geometric constraint deviation meets the preset geometric constraint deviation threshold, and the registration transformation matrix is ​​output.

7. The method for constructing a digital twin model for collaborative management of building construction according to claim 6, characterized in that, In step one, the point cloud preprocessing also includes jitter compensation, and the jitter compensation process includes: Vibration data of the equipment collected by vibration sensors on the hoisting equipment at the construction site is input into a trained jitter offset model, which outputs the spatial offset of the construction site point cloud data. Based on the spatial offset of the construction site point cloud data, the coordinate-normalized construction site point cloud data is compensated to obtain preprocessed construction site point cloud data. The equipment vibration data includes vibration frequency, amplitude, and vibration direction. The jitter offset model is used to fit the mapping relationship between the spatial offset of the point cloud data and the point cloud offset. In step two, the lightweight convolutional neural network further includes a temporal convolutional layer, and the processing procedure of the temporal convolutional layer includes: Extract a single-frame static local spatial geometric feature vector from the local feature vector of the construction site, input it into a temporal convolutional layer to perform temporal correlation operation, output the temporal correlation feature of the site, fuse the temporal correlation feature of the site with the single-frame static local spatial geometric feature vector to generate the dynamic feature vector of the site, and use the fused feature containing the dynamic feature vector of the site and the single-frame static local spatial geometric feature vector to enhance the key structure of the modular house through the introduced attention mechanism to obtain the enhanced feature vector of the site point cloud; In step five, the fine registration includes iteratively optimizing the initial registration transformation matrix based on Euclidean distance, dynamically adjusting the iteration step size and number of iterations using an adaptive compensation optimization mechanism, and introducing local geometric constraints to constrain and correct the initial registration transformation matrix; the adjustment of the iteration step size and number of iterations includes: Based on the equipment vibration data and the preprocessed point cloud data of the construction site, vibration-sensitive areas and non-vibration-sensitive areas are identified. The amplitude of the vibration-sensitive areas is input into a preset mapping table, and the vibration weight coefficient is queried. In each iteration, the local registration error of the vibration-sensitive area and the local registration error of the non-vibration-sensitive area are multiplied by the corresponding vibration weight coefficient and then superimposed to obtain the registration error. Based on the registration error and its rate of change, the iteration step size and the number of iterations are dynamically adjusted until the iteration termination condition is reached.

8. The method for constructing a digital twin model for collaborative management of building construction according to claim 6, characterized in that, In step one, the point cloud preprocessing further includes occlusion recognition and completion, which includes: An occlusion recognition reference dataset is constructed based on construction site point cloud data and standard point cloud data. The coordinate-normalized construction site point cloud data is input into a lightweight semantic segmentation algorithm for semantic segmentation, identifying and labeling initial occluded point cloud regions. These initial occluded point cloud regions are then calibrated using the occlusion recognition parameter dataset. Based on the occlusion type of the obtained occlusion point cloud regions, a differential completion operation is performed to generate preprocessed construction site point cloud data. The differential completion operation includes: If the obstruction is caused by a person and the obstruction duration is less than the preset obstruction duration, the data is completed based on the unobstructed point cloud data of the construction site in the adjacent frame. If the obstruction is caused by a person and the obstruction duration is not less than the preset obstruction duration, the data is completed based on the standard point cloud data corresponding to the obstructed point cloud area. If the obstruction is caused by mobile construction equipment, the data will be supplemented based on the point cloud data of the construction site and the standard point cloud data corresponding to the obstructed point cloud area. If the obstruction is caused by fixed construction equipment, the obstruction is completed based on the standard point cloud data corresponding to the obstructed point cloud area. In step two, the lightweight convolutional neural network further includes a feature fusion layer, the processing of which includes: Min-Max normalization is used to standardize the local feature vectors of the construction site and the standard local feature vectors respectively. An adaptive weight coefficient is introduced to weight and fuse the two types of vectors after standardization to obtain a fused feature vector. In step three, coarse registration is performed based on the fused feature vector and the standardized local feature vector.

9. The method for constructing a digital twin model for collaborative management of building construction according to claim 6, characterized in that, In step one, the point cloud preprocessing includes point cloud filtering, terrain feature extraction, downsampling, and coordinate normalization. The terrain feature extraction includes: A region growing algorithm is used to segment the filtered construction site point cloud data and standard point cloud data, extract the site terrain feature data and standard terrain feature data and calculate the offset between the two. If the offset does not meet the preset offset qualification conditions, the coordinates of the filtered construction site point cloud data are subjected to coordinate inverse compensation calculation based on the offset. Otherwise, no additional processing is performed. The site terrain feature data includes slope, undulation and elevation. In step five, the fine registration includes iteratively optimizing the initial registration transformation matrix based on Euclidean distance, dynamically adjusting the iteration step size change rate by combining an adaptive compensation optimization mechanism, and constraining and correcting the initial registration transformation matrix by introducing local geometric constraints with superimposed terrain constraints. The adjustment process for the iteration step size change rate includes: Obtain the deviation between the extracted fluctuation and the standard fluctuation, query the iteration step size change rate in the established change rate mapping table based on the deviation, and adjust the iteration step size based on the iteration step size change rate. The processing of terrain constraints includes: The registration transformation matrix after local geometric constraints is denoted as the intermediate registration transformation matrix. The terrain feature data of the point cloud after registration based on this matrix is ​​calculated. The calculated terrain feature data is compared with the preset constraint threshold. If both meet the constraint requirements, the intermediate registration transformation matrix is ​​used as the output result of step five. Otherwise, differential constraint correction is performed to obtain the output result of step five. The differentiated constraint correction includes: if the slope constraint threshold is exceeded, adjusting the rotation parameters of the intermediate registration transformation matrix; if the elevation constraint threshold is exceeded, adjusting the overall translation parameters of the intermediate registration transformation matrix; and if the undulation constraint is exceeded, adjusting the local translation parameters of the intermediate registration transformation matrix.

10. A digital twin model construction system for collaborative management of building construction, employing the digital twin model construction method for collaborative management of building construction as described in any one of claims 1-9, characterized in that, include: The design data acquisition module is used to acquire standard design data for building construction. The standard design data includes standardized component library data and external input requirement data. The standardized component library data consists of predefined building modules and their parameterized attribute data. The external input requirement data reflects the construction requirements of the building construction participants. The design scheme and interface standard formulation module is used to generate a parametric design scheme based on the standard design data through a parametric modeling algorithm, and at the same time formulate a dynamic reconfigurable interface standard for the preset interface supported by the dynamic interface generator. The interface parameter adaptive adjustment module is used to adaptively adjust the interface parameters based on the parameterized design scheme and the dynamic reconfigurable interface standard by using a dynamic interface generator that integrates a dynamic graph neural network interface adaptation algorithm to obtain interface adaptation parameters. The digital twin model construction module is used to integrate standard point cloud data extracted from the BIM model with the collected point cloud data from the construction site, and to perform lightweight modeling processing by combining the interface adaptation parameters with the LOD hierarchical method and the lightweight adaptive point cloud registration algorithm to obtain a digital twin model that is synchronized with the physical entity of the building construction.

Images