A BIM-based fire engineering construction progress and quality management and control method and system

Abstract

The application discloses a kind of based on BIM's fire engineering construction progress and quality control method and system, method includes: the building information model of fire engineering and preset construction plan are loaded, generate time-stamped field state data stream;Field state data stream is registered with point cloud with the building information model and feature mapping, respectively extract progress deviation vector and quality deviation vector, generate deviation fusion matrix;Deviation fusion matrix is input into graph neural network, and output associated risk propagation graph;Based on associated risk propagation graph, generate control instruction set containing progress catch-up strategy and quality problem positioning and issue to construction management terminal, realize the dynamic control of fire engineering construction progress and quality.Utilize the embodiment of the application, can realize the collaborative dynamic control of progress and quality in fire engineering construction, improve deviation timeliness and control decision accuracy.

Description

Technical Field

[0001] This invention belongs to the field of fire protection engineering technology, and in particular to a method and system for controlling the construction progress and quality of fire protection engineering based on BIM. Background Technology

[0002] Fire protection engineering involves cross-disciplinary work involving sprinkler pipes, alarm equipment, and smoke control systems. Construction progress and quality control heavily rely on on-site management and experience-based judgment. Traditional management methods typically employ manual inspections, paper records, and post-construction acceptance, which suffer from problems such as delayed data collection, untimely detection of deviations, and isolated handling of progress and quality issues. In recent years, Building Information Modeling (BIM) technology has been gradually applied to construction process management; however, most methods remain at the level of 3D visualization and static collision detection, failing to dynamically compare on-site measured data with model design values ​​and quantify deviations. More importantly, schedule deviations and quality defects in fire protection engineering often exhibit a coupled propagation effect. For example, delays in pipe installation in one area can cause cascading delays in subsequent equipment commissioning, while installation posture deviations may induce new rework. Existing methods lack the ability to calculate and predict the associated risks between schedule and quality deviations, making coordinated management of both impossible. Summary of the Invention

[0003] The purpose of this invention is to provide a BIM-based method and system for controlling the construction progress and quality of fire protection engineering, in order to overcome the shortcomings of the existing technology, realize the coordinated dynamic control of progress and quality in fire protection engineering construction, and improve the timeliness of deviation detection and the accuracy of control decisions.

[0004] One embodiment of this application provides a BIM-based method for controlling the construction progress and quality of fire protection engineering, the method comprising: Load the building information model and preset construction plan of the fire protection project, and use laser scanners and panoramic cameras to collect component point cloud data and image data at the construction site to generate a time-stamped site status data stream; The on-site status data stream is registered with the building information model using point cloud and feature mapping. By comparing the actual component position and installation posture with the model design value, and combining the process nodes in the preset construction plan, the degree of deviation is determined. The progress deviation vector and quality deviation vector are extracted respectively to generate a deviation fusion matrix. The deviation fusion matrix is ​​input into a graph neural network. A correlation topology graph is constructed with fire protection components as nodes and construction process dependencies as edges. The coupling propagation strength between schedule deviation and quality deviation is calculated through a message passing mechanism, and a correlation risk propagation graph is output. Based on the aforementioned risk propagation graph, a long short-term memory network is used to predict the evolution trend of deviations within the preset construction period. Combined with the resource availability constraints in the preset construction plan, critical path replanning is performed to generate a set of control instructions that includes progress catch-up strategies and quality problem location, which are then sent to the construction management terminal to achieve dynamic control of the construction progress and quality of fire protection engineering.

[0005] Optionally, the loading of the building information model and preset construction plan for the fire protection project, and the collection of component point cloud data and image data at the construction site using a laser scanner and panoramic camera, generate a time-stamped site status data stream, including: Load the building information model file of the fire protection project from the project database, parse the position coordinates and geometric parameters of each component in the model, and read the list of process nodes in the preset construction plan to generate the model-plan basic dataset; Laser scanners are deployed at the construction site to perform multi-station scanning of the constructed area according to the preset scanning stations, and the original point cloud data of each station is obtained. At the same time, panoramic cameras are used to collect 360-degree image data at each scanning station to generate original point cloud and image matching data. The original point cloud and image paired data are subjected to noise reduction filtering and stitching fusion to remove outliers and moving object point clouds, and the multi-site cloud is registered to a unified coordinate system to generate a fused full-site cloud of the construction area. Add a timestamp of the acquisition time to each point in the merged panoramic view cloud of the construction area, and attach the corresponding panoramic image as a texture to it, finally generating a time-stamped on-site status data stream.

[0006] Optionally, the step of performing point cloud registration and feature mapping between the on-site status data stream and the building information model, comparing the actual component positions and installation postures with the model design values, and determining the degree of deviation by combining the process nodes in the preset construction plan, extracting the progress deviation vector and quality deviation vector respectively, and generating a deviation fusion matrix includes: The overall point cloud in the site status data stream is coarsely registered with the design point cloud in the building information model. The initial transformation matrix is ​​estimated using the random sampling consistency algorithm, and then fine registration is performed using the iterative nearest point algorithm to generate the registered site point cloud model. The actual position coordinates and installation attitude angles of the fire protection components are extracted from the registered site point cloud model. The deviation is calculated for each component and compared with the design value of the corresponding component in the building information model to obtain the spatial deviation of each component and generate the component deviation feature vector. Based on the completion time of the process nodes in the preset construction plan, it is determined whether the currently completed components are completed on schedule. The delay time is quantified into a schedule deviation value. At the same time, the over-threshold items in the component deviation feature vector are marked as quality deviation values, and a schedule deviation vector and a quality deviation vector are generated. The schedule deviation vector and the quality deviation vector are aligned by component index and then fused into a unified deviation fusion matrix through principal component analysis. The rows of the matrix represent components and the columns represent deviation types and indicators.

[0007] Optionally, the step of inputting the deviation fusion matrix into a graph neural network, constructing an associated topology graph with fire protection components as nodes and construction process dependencies as edges, calculating the coupling propagation strength between schedule deviations and quality deviations through a message passing mechanism, and outputting an associated risk propagation graph includes: The construction process dependency table of the fire protection project is analyzed. Each fire protection component is treated as a node, and the dependencies between processes are treated as directed edges. The deviation characteristics of each node in the deviation fusion matrix are used as the initial node attributes to generate an associated topology graph structure. The associated topology graph structure is input into the graph neural network. The deviation information of neighboring nodes is aggregated through graph convolutional layers. A multi-head attention mechanism is used to calculate the weight of the impact of progress deviation on quality deviation between different nodes, and a coupling strength matrix between nodes is generated. Based on the inter-node coupling strength matrix, multiple rounds of message passing iterations are performed to propagate the quality deviation of each node to adjacent nodes according to the coupling weight, while the progress deviation is propagated in reverse to generate the node deviation features after propagation. The propagated node deviation features are superimposed with the original edge direction information to generate a risk propagation graph. In the graph, the node color represents the comprehensive risk level, and the edge thickness represents the coupling propagation strength.

[0008] Optionally, based on the associated risk propagation map, a long short-term memory network is used to predict the deviation evolution trend within a preset construction period. Combined with resource availability constraints in the preset construction plan, critical path replanning is performed to generate a control instruction set containing schedule catch-up strategies and quality problem location, which is then distributed to the construction management terminal. This achieves dynamic control of the fire protection engineering construction progress and quality, including: The deviation time series of each node in different time windows is extracted from the associated risk propagation map and input into a pre-trained long short-term memory network model. The model outputs the predicted values ​​of progress deviation and quality deviation at a future preset time and generates a deviation evolution trend prediction curve. Based on the deviation evolution trend prediction curve, identify components that are about to exceed the tolerance threshold and their deviation types, and combine them with the resource availability constraints in the preset construction plan, including the remaining amount of manpower, equipment and materials, to generate a resource conflict list. Based on the resource conflict list, the critical path method is used to replan the construction plan, adjust the execution sequence of the work process and add parallel operations, and generate a schedule catch-up strategy and a list of quality problem locations. The schedule catch-up strategy and the list of quality issues are encapsulated into a set of control instructions, including the component numbers that need to be reworked, the amount of additional resources, and the adjusted schedule milestones. These instructions are then pushed to the on-site engineers through the construction management terminal to achieve dynamic control.

[0009] Another embodiment of this application provides a BIM-based fire protection engineering construction progress and quality control system, the system comprising: The data acquisition module is used to load the building information model and preset construction plan of the fire protection project, and to use laser scanners and panoramic cameras to collect point cloud data and image data of components at the construction site, generating a time-stamped on-site status data stream. The extraction module is used to perform point cloud registration and feature mapping between the on-site status data stream and the building information model. By comparing the actual component position, installation posture and model design value, and combining the process nodes in the preset construction plan, the degree of deviation is determined, and the progress deviation vector and quality deviation vector are extracted respectively to generate a deviation fusion matrix. The construction module is used to input the deviation fusion matrix into the graph neural network, construct an associated topology graph with fire protection components as nodes and construction process dependencies as edges, calculate the coupling propagation strength between schedule deviation and quality deviation through a message passing mechanism, and output an associated risk propagation graph. The generation module is used to predict the deviation evolution trend within a preset construction period based on the associated risk propagation map using a long short-term memory network, and to perform critical path replanning in conjunction with the resource availability constraints in the preset construction plan. It generates a set of control instructions that includes progress catch-up strategies and quality problem location and sends them to the construction management terminal to realize dynamic control of the construction progress and quality of fire protection engineering.

[0010] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0011] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0012] Compared with existing technologies, the present invention provides a BIM-based method for controlling the construction progress and quality of fire protection engineering, which can realize the coordinated and dynamic control of progress and quality in fire protection engineering construction, and improve the timeliness of deviation detection and the accuracy of control decisions. Attached Figure Description

[0013] Figure 1 A hardware structure block diagram of a computer terminal for a BIM-based fire protection engineering construction progress and quality control method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a BIM-based method for controlling the construction progress and quality of fire protection engineering, provided as an embodiment of the present invention. Figure 3 This is a schematic diagram of a BIM-based fire protection engineering construction progress and quality control system provided in an embodiment of the present invention. Detailed Implementation

[0014] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] This invention first provides a BIM-based method for controlling the construction progress and quality of fire protection engineering. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0016] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a BIM-based method for controlling the construction progress and quality of fire protection engineering, provided as an embodiment of the present invention. (See diagram below.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0017] See Figure 2 The present invention provides a BIM-based method for controlling the construction progress and quality of fire protection engineering, which may include the following steps: S201 loads the building information model and preset construction plan of the fire protection project, and uses a laser scanner and panoramic camera to collect component point cloud data and image data at the construction site to generate a time-stamped site status data stream. Specifically, the building information model file of the fire protection project can be loaded from the project database, the position coordinates and geometric parameters of each component in the model can be parsed, and the list of process nodes in the preset construction plan can be read to generate the model-plan basic dataset. The core of this step is to achieve unified retrieval and structured parsing of BIM model data from the design side and planning data from the construction side, stripping redundant rendering data from the BIM model, extracting the core spatial and dimensional parameters required for fire protection construction management, standardizing construction procedure time and logical data, and establishing a benchmark data source for subsequent on-site measurement data comparison. This provides a unified reference standard for subsequent point cloud registration and deviation calculation. The specific implementation method is as follows: The project database is a dedicated storage unit for storing static data throughout the entire lifecycle of a fire protection engineering project. Internally, the database stores data in three main modules: model files, construction plan files, and component ledger files. The data storage format is a standard engineering structured format, and the data read response latency is controlled within 80ms to ensure real-time data retrieval from the front end and prevent data read delays from impacting subsequent overall management processes. The fire protection engineering building information model stored in the database is a detailed construction-specific BIM model, distinct from the rough model in the initial design phase. This model contains comprehensive information on all physical fire protection components on the construction site, covering all categories of construction components such as sprinkler mains, sprinkler branches, fireproof smoke exhaust valves, fire hydrants, fire pumps, and pipe supports, without any missing model components.

[0018] After loading the BIM model file, the system automatically calls the built-in BIM model parsing interface to decompose the model data, extracting the spatial coordinates and geometric parameters of each type of fire protection component. The spatial coordinates are represented using a unified three-dimensional geodetic coordinate system for the construction site, including three basic parameters: x-axis horizontal coordinate, y-axis vertical coordinate, and z-axis vertical elevation coordinate. The coordinate accuracy is fixed at 0.001m. Parameter x_1 represents the horizontal coordinate of the center axis of a single sprinkler pipe, parameter y_1 represents the longitudinal positioning coordinate of the component, and parameter z_1 represents the installation elevation coordinate of the component. These three coordinates together pinpoint the precise installation position of the fire protection component on the construction site. Geometric parameters define the external dimensions and installation specifications of the fire protection component. For pipe components, geometric parameters include the nominal pipe diameter D_1, pipe wall thickness H_1, and standard length of a single pipe section L_1. For valve components, geometric parameters include the valve body's length, width, and height, and the flange mounting hole spacing. For support and hanger components, geometric parameters include the steel section dimensions, hanger length, and horizontal support arm width. All geometric parameters are completely consistent with the fire protection engineering construction design drawings, ensuring the accuracy of the design baseline data.

[0019] After completing the BIM model parameter parsing, the system synchronously reads the process node list corresponding to the preset construction plan in the database. The process node list is a standardized time sequence form for the construction of fire protection engineering sub-items. The core parameters in the list include the planned start time T_s of a single process, the planned completion time T_e of the process, the process's prerequisite dependency number N_1, and the set of construction components corresponding to the process C_1. Among them, parameter T_s represents the time node when a single fire protection construction process officially starts, parameter T_e represents the deadline node when the process must be completed, parameter N_1 is used to indicate that the current process can only start after the completion of the preceding process, reflecting the sequential logic of the construction process on site, and parameter C_1 is used to bind all the fire protection components that need to be installed in the current process, realizing a one-to-one correspondence between construction time and construction physical components. In the example, the planned start time T_s of the fire main pipe installation process is the 12th day after the project starts, and the planned completion time T_e is the 18th day after the project starts. The prerequisite dependency number N_1 is the construction process of the reserved hole in the fire floor slab. The corresponding construction component set C_1 includes 8 sections of DN150 fire sprinkler main pipes and 24 sets of supporting pipe supports.

[0020] The system integrates the parsed component location coordinates, component geometric parameters, process sequence parameters, and process-related component parameters. It aligns the data using the unique component code as an index, removes irrelevant non-fire protection component data such as civil engineering walls, floors, doors, and windows from the model, and removes process data irrelevant to the early civil engineering preparation in the construction plan. Finally, it integrates to form a standardized model-plan basic dataset. This dataset completely retains all design and planning benchmarks for fire protection construction and is the only reference for subsequent comparison of on-site measured data.

[0021] Laser scanners are deployed at the construction site to perform multi-station scanning of the constructed area according to the preset scanning stations, and the original point cloud data of each station is obtained. At the same time, panoramic cameras are used to collect 360-degree image data at each scanning station to generate original point cloud and image matching data. The core of this step is to rely on laser scanning ranging technology and panoramic image acquisition technology to complete the in-situ acquisition of full-area spatial information and visual image information of the completed fire protection components at the construction site. By using multi-site staggered scanning to compensate for the visual blind spots of a single scanning site, and at the same time, to achieve temporal and spatial synchronization of point cloud spatial data and image visual data, a raw data source containing spatial geometric information and real-scene visual information is constructed on site. The specific implementation method is as follows: The laser scanner utilizes industrial-grade outdoor 3D laser scanning equipment, adaptable to complex construction environments such as dust, uneven lighting, and slight obstructions. The equipment's effective scanning range is set from 0.5m to 80m, fully covering typical fire protection engineering construction areas such as fire corridors, equipment rooms, and underground parking garages. It emits 12,000 laser points per second, ensuring ample spatial detail in the original point cloud. The scanning sites at the construction site are planned according to an equilateral triangle layout principle, with a fixed straight-line distance of 6m between adjacent scanning sites. This distance balances the scanning overlap rate and overall scanning efficiency, controlling the overlap area between any two adjacent sites to within 35%. This ensures accurate matching of subsequent point cloud stitching without generating excessive redundant scanning data.

[0022] Multi-station scanning is conducted sequentially according to preset sites, with a fixed scanning time of 90 seconds per site. The scanning field of view is set to 360 degrees horizontally and -45 degrees to +90 degrees vertically, completely covering all fire protection construction components on the ground, walls, and ceiling around the site, without blind spots such as ceiling pipes and bottom supports. After each site laser scan is completed, the device's local storage unit automatically caches the corresponding original point cloud data. The original point cloud of a single site contains two basic types of information: spatial three-dimensional coordinates and laser reflection intensity. The laser reflection intensity parameter I_1 is used to distinguish components of different materials on the construction site. The reflection intensity value range of metal fire protection pipes is 75 to 90, and the reflection intensity value range of concrete walls is 30 to 45. The difference in reflection intensity can initially distinguish fire protection components from civil engineering structures, reducing the difficulty of subsequent component identification.

[0023] At the same moment that each laser scanning station starts scanning, the interconnected panoramic cameras simultaneously begin acquiring 360-degree panoramic images. The system is configured with hardware synchronization trigger signals to control the time deviation between the laser scanning start time and the camera capture time to within 20ms, achieving high-precision time synchronization between spatial point cloud data and real-scene image data. The resolution of a single panoramic camera image is fixed at 6000*3000 pixels, with no image distortion correction error. Each station generates a complete panoramic real-scene image without blind spots, fully recording visual information such as the appearance of fire protection components, construction gaps, and temporary construction debris around the current station.

[0024] After completing the data collection at all preset sites, the system uses the scanning site number as a unique association index to bind the single-site raw point cloud data generated at the same site to the corresponding panoramic image. Each set of bound data corresponds to the same spatial location and the same acquisition time. Finally, multiple sets of raw point cloud and image pairing data are generated in batches. All pairing data are sorted in order according to the site deployment order, preserving the original acquisition time sequence, and providing a one-to-one original data source for subsequent point cloud fusion and image texture attachment.

[0025] The original point cloud and image paired data are subjected to noise reduction filtering and stitching fusion to remove outliers and moving object point clouds, and the multi-site cloud is registered to a unified coordinate system to generate a fused full-site cloud of the construction area. The core of this step is to clean and spatially calibrate the raw, coarse point cloud data collected on-site. This eliminates invalid noise point clouds caused by environmental interference at the construction site, solves the problems of inconsistent coordinate systems and spatial misalignment among point clouds from multiple sites, and stitches together the scattered local point clouds from multiple sites into a complete and continuous point cloud covering the entire construction area, thus restoring the true spatial state of the fire protection components at the construction site. The specific implementation method is as follows: First, a two-stage noise reduction filtering process is performed on each set of original point cloud data. This involves executing two algorithms: statistical filtering and direct-pass filtering. These two algorithms work together to remove noisy point clouds. The first stage is the statistical filtering algorithm, primarily used to identify and remove spatially discrete outliers. Outliers often originate from floating dust in the air at the construction site or invalid laser points reflected from distant irrelevant debris. The algorithm sets the number of neighboring search points K_1 to 30 and the average distance threshold D_1 to 15mm. It calculates the average spatial distance between the current point and its 30 neighboring points. If the average distance of a single point exceeds 15mm, that point is determined to be a spatially discrete outlier and is directly removed from the original point cloud sequence. The second stage is a direct filtering algorithm, which is mainly used to remove point clouds of moving objects caused by personnel walking on the construction site, forklifts transporting materials, and swinging construction cables. The algorithm compares the coordinate offset of the same spatial position in two consecutive frames of point clouds, sets the coordinate offset threshold S_1 to 20mm, and if the offset between frames of point clouds at the same spatial position exceeds 20mm, it is determined to be an invalid point cloud caused by dynamic motion interference, and the batch removal is completed.

[0026] After single-point cloud denoising, multi-site cloud spatial registration and stitching fusion operations are performed. During the acquisition process, each multi-site cloud uses its own scanning site as its local coordinate origin. The coordinate systems of different site point clouds are independent and cannot be directly stitched together. Therefore, a rigid body space transformation algorithm is needed to register and align all local point clouds to a unified global coordinate system. This unified coordinate system adopts the construction site geodetic 3D coordinate system used in the BIM model analysis in step one, ensuring that the site point clouds and the design BIM model coordinate systems are completely homogeneous and eliminating reference coordinate deviations in subsequent registration processes. The algorithm extracts common feature points such as the circular cross-section of fire-fighting pipelines and the outline of rectangular valve bodies within the overlapping areas of adjacent sites, calculates the translation and rotation transformation parameters between the local and global coordinate systems, and uniformly transforms the denoised point clouds of all sites to the global geodetic coordinate system.

[0027] To address the issue of duplicate point cloud data in overlapping areas of multiple site cloud systems, a weighted average fusion algorithm was employed to optimize the overlapping point clouds. The weight coefficient for the center of the overlapping area was set to 0.7, and the weight coefficient for the edge of the overlapping area was set to 0.3. A weighted average was calculated for the coordinates of the duplicate point clouds at the overlapping locations, preserving the complete component details of the overlapping areas while eliminating data redundancy caused by duplicate point clouds. After all site point clouds were registered and fused with overlapping areas, there were no scanning blind spots, no noise interference, and no coordinate misalignment. This resulted in a single, continuous, and complete panoramic point cloud covering the entire construction area. This panoramic point cloud perfectly replicates the actual installation spatial position and outline of all fire protection components on site, with the overall spatial accuracy error controlled within 2mm, meeting the accuracy requirements for subsequent construction deviation detection.

[0028] Add a timestamp of the acquisition time to each point in the merged panoramic view cloud of the construction area, and attach the corresponding panoramic image as a texture to it, finally generating a time-stamped on-site status data stream.

[0029] The core of this step is to supplement pure spatial geometric point clouds without temporal attributes and visual textures with temporal identifiers and real-world visual textures. This upgrades a single spatial point cloud into a multi-dimensional field data stream that includes spatial location, acquisition time, and real-world appearance information. This enables full-domain synchronous recording of the spatial, temporal, and visual states of the construction site, and is suitable for comparing longitudinal progress deviations at different construction periods. The specific implementation method is as follows: First, a high-precision acquisition timestamp is bound to each spatial point within the overall scenic cloud. The timestamp adopts a millisecond-level time sequence encoding format, with the encoding structure being year-month-day-hour-minute-second-millisecond, achieving a timestamp accuracy of 1ms. This allows for precise differentiation of minute changes in the construction status of the construction site within a short time interval. The system retrieves the precise acquisition time corresponding to each scanning station in the second acquisition phase. Combined with the point cloud attribution information during the point cloud stitching and fusion process, the original acquisition station and acquisition time corresponding to each spatial point within the overall scenic cloud are located, and the corresponding millisecond-level timestamp parameter T_ms is written point by point. The timestamp parameter is not uniformly normalized, preserving the actual on-site acquisition time sequence. This accurately reproduces the time differences in the sequential scanning and construction of fire protection components in different areas. In the example, the timestamps of the underground garage sprinkler pipe area are concentrated between 10:20 AM and 10:22 AM on the same day, and the timestamps of the computer room fire pump area are concentrated between 10:25 AM and 10:27 AM on the same day, closely matching the actual operation time sequence of sequential scanning of different areas on site.

[0030] After completing the point-by-point timestamp binding, a panoramic image texture mapping and attachment operation is performed. A 3D point cloud UV texture mapping algorithm is used to achieve a precise one-to-one match between the panoramic image pixel information and the 3D spatial point cloud. The algorithm first establishes a mapping relationship between the 2D pixel coordinate system of the panoramic image and the 3D spatial coordinate system of the panoramic point cloud. Based on the spatial normal vector of the point cloud, it matches the color information of the corresponding pixels in the image. The panoramic image paired in step two is then used as a texture resource and attached to the surface of the panoramic point cloud area generated at the corresponding site. This allows the monochrome point cloud, which originally only possessed spatial coordinates, to have visual texture information such as color, appearance, and surface construction texture completely consistent with the construction site. After texture attachment, the point cloud can intuitively present visible construction appearance quality information such as the integrity of the paint surface of fire-fighting pipelines, the appearance of valve body sealing surfaces, and the welding quality of supports and hangers, compensating for the shortcoming of pure geometric point clouds in being unable to identify appearance quality defects.

[0031] Finally, the system sequentially sorts all 3D point clouds with timestamps and image textures according to a time-series rule from earliest to latest, unifying the data transmission format and data encapsulation rules. This integrates discrete single-point cloud data into a continuous serial data stream structure. Within the data stream, the on-site point cloud states of different construction areas are arranged sequentially according to the acquisition time, without data disorder or timestamp errors. The final result is a standardized on-site status data stream with timestamps. This data stream simultaneously integrates 3D spatial geometric information of fire protection components at the construction site, millisecond-level real-time acquisition information, and real-scene image texture visual information. It can directly perform time-series and visualized data comparison with the BIM design model, supporting subsequent point cloud registration and quantitative calculation of construction deviations.

[0032] S202, the on-site status data stream is registered with the building information model by point cloud and feature mapping. By comparing the actual component position, installation posture and model design value, and combining the process nodes in the preset construction plan, the degree of deviation is determined. The progress deviation vector and quality deviation vector are extracted respectively to generate a deviation fusion matrix. Specifically, the overall point cloud in the field status data stream can be coarsely registered with the design point cloud in the building information model, the initial transformation matrix can be estimated using the random sampling consensus algorithm, and then fine registration can be performed using the iterative nearest point algorithm to generate the registered field point cloud model. The core of this step is to achieve unified alignment of the spatial coordinates between the measured point cloud at the construction site and the BIM design point cloud. First, coarse registration is used to eliminate large positional and orientation offsets between the two sets of point clouds. Then, fine registration is used to eliminate minor spatial errors, solving the problem of inconsistency between the scanning coordinate system at the construction site and the design coordinate system of the BIM model. This provides data support from the same source under the same coordinate system for subsequent calculation of spatial deviations of components. The specific implementation method is as follows: The built-in panoramic point cloud in the site status data stream is a measured point cloud collected from the actual construction site. It contains the three-dimensional spatial discrete point coordinates of all fire protection components, civil engineering structures, and pipelines and cable trays on site. The number of points in a single panoramic point cloud is controlled between 800,000 and 1.2 million, and the spatial coordinate acquisition accuracy of a single point is 0.005m. This point cloud is generated based on the local scanning coordinate system of the construction site, with the origin of the coordinate system being the initial setup position of the laser scanner. The design point cloud in the building information model is a standard design point cloud generated by grid sampling from the BIM fire protection model. Discrete spatial points are sampled evenly according to the contours of the model components. The theoretical accuracy of a single point in the design point cloud can reach 0.001m. It is generated based on the project's globally unified design coordinate system. The origins and three-axis orientations of the two sets of point cloud coordinate systems have natural deviations, making direct spatial comparison impossible. Therefore, coordinate alignment needs to be completed in two stages: coarse registration and fine registration.

[0033] In the coarse registration stage, the random sampling consensus algorithm is used to solve the initial spatial transformation. The core function of this algorithm is to filter out noise outliers and interference from irrelevant civil engineering point clouds in the two sets of point clouds, quickly lock the source matching points of fire protection components, and output the initial transformation matrix to complete the coarse alignment of large-scale coordinates. The random sampling consensus algorithm has three core controllable parameters. Parameter N_1 represents the number of random sampling iterations, set to 1000 in this case. A larger value results in higher stability of the matching results, but also increases computation time. 1000 iterations balance registration accuracy and computational efficiency. Parameter T_1 represents the interior point distance threshold in meters, set to 0.1m in this case. This parameter determines whether the midpoints of two point clouds are from the same source. When the spatial distance between a single point in the measured point cloud and the designed point cloud is less than this threshold, it is considered a valid matching interior point; otherwise, it is considered an interfering exterior point and directly discarded. Parameter S_1 represents the effective interior point percentage threshold, set to 65% in this case. When the proportion of effective interior points obtained in a single iteration exceeds this value, the transformation parameters solved in this iteration are considered valid. During algorithm execution, pairs of points from the same source are randomly extracted from the measured panoramic point cloud and the BIM design point cloud. Rigid body transformation parameters are repeatedly calculated iteratively, and the final output is an initial transformation matrix M_1 containing translation and rotation parameters. This matrix can eliminate more than 90% of the large-scale spatial offset between the two sets of point clouds. After coarse registration, the overall outlines of the two sets of point clouds are basically overlapping, but there are still small residual deviations at the millimeter to centimeter level.

[0034] In the fine registration stage, the iterative nearest-neighbor algorithm is used to refine the coarse registration results, compensating for minor spatial deviations left over from the coarse registration and further improving the point cloud alignment accuracy. The iterative nearest-neighbor algorithm has four core operating parameters: parameter N_2 represents the maximum number of iterations (set to 50 in this case to avoid infinite iterations consuming computational resources); parameter T_2 represents the iteration convergence error threshold (in meters, set to 0.008m in this case; when the overall average spatial deviation of the point cloud calculated in two consecutive iterations is less than this value, registration is directly determined to be converged and iteration stops); parameter K_1 represents the number of nearest neighbor points searched (set to 20 in this case; for each spatial point in the measured point cloud, the 20 closest points in the designed point cloud are searched as candidate matching points to improve matching accuracy); and parameter E_1 represents the maximum displacement correction in a single iteration (in meters, set to 0.05m in this case; this limits the coordinate correction range in a single iteration to avoid excessive offset of the point cloud). The algorithm uses the initial transformation matrix output from the coarse registration as the initial value for iteration. It calculates the spatial distance error between the measured point cloud and the design point cloud point by point, continuously iterating and updating the transformation matrix to reduce the overall matching error until the convergence condition is met. After completing the entire registration process, the original site-wide point cloud is transformed according to the final fusion transformation matrix to obtain a registered site point cloud model with the same coordinate system as the BIM design model. The coordinates of all spatial points in this model are consistent with the BIM design coordinate system, meeting the basic requirements for subsequent comparison of component spatial parameters.

[0035] The actual position coordinates and installation attitude angles of the fire protection components are extracted from the registered site point cloud model. The deviation is calculated for each component and compared with the design value of the corresponding component in the building information model to obtain the spatial deviation of each component and generate the component deviation feature vector. The core of this step is to extract the actual installation parameters of fire protection components from the aligned measured point cloud model, match them one-to-one with the standard design parameters built into the BIM model, quantify the two core spatial installation deviations of fire protection components (position and attitude), transform the concrete spatial installation errors into standardized numerical features, and construct component deviation feature vectors that can be identified by subsequent algorithms. This achieves a digital and quantitative representation of construction and installation quality issues. The specific implementation method is as follows: First, semantic segmentation and feature parameter extraction of fire protection components were performed on the registered on-site point cloud model. Using point cloud semantic segmentation technology, point cloud clusters of different types of fire protection components were distinguished. Independent point cloud regions were created for five common fire protection components: sprinkler pipes, smoke exhaust ducts, fire dampers, fire pump connections, and indoor fire hydrants, to avoid interference between point clouds of different components in the parameter extraction results. For each type of fire protection component, two core measured parameters were extracted: center position coordinates and three-axis installation attitude angles. The component center position coordinates were represented by three-dimensional rectangular coordinates (x, y, z axes), with the unit uniformly in meters, and the coordinate extraction accuracy retained to three decimal places. The installation attitude angles were divided into pitch angle α, roll angle β, and yaw angle γ, with the unit uniformly in degrees, and the angle extraction accuracy retained to one decimal place. The pitch angle represents the vertical rotation angle of the component around the x-axis, the roll angle represents the horizontal tilt angle of the component around the y-axis, and the yaw angle represents the horizontal rotation angle of the component around the z-axis. These three attitude angles completely cover all spatial installation attitude changes of the fire protection components.

[0036] Simultaneously, standard design parameters for fire protection components with the same number and location are retrieved from the Building Information Model (BIM). These design parameters are the theoretically optimal installation parameters preset in the early stages of project modeling, ensuring no construction or installation errors. They correspond to the design center coordinates (X, Y, Z) and design attitude angles (α0, β0, γ0), respectively. After matching the measured parameters with the design parameters one by one, quantitative calculations of spatial deviations for each component are performed. The spatial deviations are divided into two main modules: positional deviation components and attitude deviation components. Positional deviations include x-axis positional deviation Δx, y-axis positional deviation Δy, and z-axis positional deviation Δz, which are the differences between the measured coordinates and the design coordinates, respectively. Attitude deviations include pitch angle deviation Δα, roll angle deviation Δβ, and yaw angle deviation Δγ, which are the differences between the measured attitude angles and the design attitude angles, respectively.

[0037] Taking a commonly used rectangular smoke exhaust duct component in the project as a specific calculation example, the BIM design center coordinates of this smoke exhaust duct are (3.200m, 5.600m, 2.800m), and the design attitude angle is (0.0°, 0.0°, 0.0°). The actual center coordinates of the duct obtained from on-site measurements are (3.221m, 5.587m, 2.812m), and the actual installation attitude angle is (0.3°, -0.2°, 0.4°). After difference calculation, the positional deviations of this smoke exhaust duct are Δx=0.021m, Δy=-0.013m, and Δz=0.012m, respectively, and the attitude deviations are Δα=0.3°, Δβ=-0.2°, and Δγ=0.4°, respectively.

[0038] The six spatial deviation values ​​of a single component are arranged and integrated in a fixed order to form a one-dimensional component deviation feature vector corresponding to that component. The vector dimension is fixed at 6 dimensions, corresponding to the x-axis position deviation, y-axis position deviation, z-axis position deviation, pitch angle deviation, roll angle deviation, and yaw angle deviation, respectively. After completing the deviation calculation and vector construction for all fire protection components in the project according to unified rules, each component generates an independent component deviation feature vector with a uniform format. Each vector uniquely corresponds to a physical fire protection component on the construction site, fully carrying all spatial installation quality deviation information of that component, providing complete original numerical basis for the subsequent classification and judgment of schedule deviations and quality deviations.

[0039] Based on the completion time of the process nodes in the preset construction plan, it is determined whether the currently completed components are completed on schedule. The delay time is quantified into a schedule deviation value. At the same time, the over-threshold items in the component deviation feature vector are marked as quality deviation values, and a schedule deviation vector and a quality deviation vector are generated. The core of this step is to assess the degree of construction progress lag from a time dimension and the degree of construction and installation quality defects from a spatial installation deviation dimension. Based on preset progress tolerance thresholds and quality installation deviation thresholds, it distinguishes between progress deviation data and quality deviation data, separating the two types of deviation information from different dimensions, and constructing independent progress deviation vectors and quality deviation vectors respectively. This achieves the separation and quantification of construction progress issues and construction quality issues. The specific implementation method is as follows: Firstly, construction progress is assessed based on the completion times of the process nodes in the project's pre-set construction plan. Each fire protection component is assigned a unique planned completion time node T_p in the pre-set construction plan; this time node represents the theoretical completion time of the component's installation process. Simultaneously, the system has a built-in global current construction data collection time T_n. The difference between these two time parameters is used as the basis for progress evaluation. The system has a pre-set progress deviation tolerance threshold T_th, in days. In this case, T_th is set to 0.5 days. This parameter represents the reasonable allowable construction period fluctuation for the project. When the difference between the actual completion time and the planned completion time of a component is less than this threshold, the construction progress is considered to be without deviation and within the normal construction fluctuation range; if it exceeds this threshold, a construction progress delay is identified.

[0040] The quantification rule for schedule deviation values ​​is a unified linear quantification calculation. The schedule deviation value ΔT equals the actual completion time of the component minus the planned completion time. A positive result indicates a delay in the process, with a larger value indicating a longer delay; a result of zero indicates on-time completion; and a negative result indicates early completion. All installed fire protection components on the construction site are sequentially traversed, and the schedule deviation value for each component is calculated. All component schedule deviation values ​​are arranged in a sorting order that perfectly matches the component deviation feature vector, ultimately generating a one-dimensional schedule deviation vector. Each element within the vector corresponds to the delay time of a single fire protection component, comprehensively reflecting the construction schedule deviation of all components.

[0041] Secondly, quality deviation judgment is performed on the component deviation feature vector. The system pre-sets a three-dimensional position deviation threshold D_th and a three-axis attitude deviation threshold A_th. The position deviation threshold D_th is set to 0.02m, representing the maximum permissible error of single-axis position installation of the fire protection component; the attitude deviation threshold A_th is set to 0.5°, representing the maximum permissible error of single-axis attitude installation of the fire protection component. The six deviation indicators within the component deviation feature vector are compared with the corresponding thresholds. When the absolute value of the single-axis position deviation is greater than D_th, or the absolute value of the single-axis attitude deviation is greater than A_th, the indicator is judged as an out-of-limit quality deviation item; otherwise, it is judged as qualified with no quality deviation item.

[0042] For each fire protection component, the sum of the absolute values ​​of all out-of-limit indicators within its component deviation feature vector is calculated. This sum is used as the unified quantified quality deviation value for that component; the larger the value, the more severe the installation quality defects. Continuing with the component sorting index that is completely consistent with the schedule deviation vector and component deviation feature vector, the quantified quality deviation values ​​of all components are arranged sequentially to generate a one-dimensional quality deviation vector.

[0043] Using the smoke exhaust duct example from before, the planned completion date for this duct was day 12 of construction, while the actual completion date was day 14, resulting in a schedule deviation of ΔT = 2 days, significantly exceeding the 0.5-day tolerance threshold. Among its six spatial deviations, the x-axis position deviation of 0.021m exceeds the 0.02m position deviation threshold, while the other five indicators are all within the threshold range. Therefore, the component's quality deviation is 0.021m. After quantifying all component deviations in this way, the schedule deviation vector focuses on representing schedule delays, while the quality deviation vector focuses on representing installation spatial defects. Both types of vectors have identical dimensions and perfectly aligned component indices, satisfying the prerequisite matching requirements for subsequent matrix fusion.

[0044] The schedule deviation vector and the quality deviation vector are aligned by component index and then fused into a unified deviation fusion matrix through principal component analysis. The rows of the matrix represent components and the columns represent deviation types and indicators.

[0045] The core of this step is to eliminate data redundancy between the schedule deviation vector and the quality deviation vector, unify the data dimension and data distribution characteristics of the two types of heterogeneous deviation data, and complete the dimensionality reduction and fusion of high-dimensional vectors through principal component analysis algorithm. This integrates the one-dimensional schedule deviation data and one-dimensional quality deviation data into a two-dimensional structured deviation fusion matrix, standardizes the deviation data format, and adapts to the input data requirements of subsequent graph neural networks. The specific implementation method is as follows: First, the component indices of the two sets of vectors are precisely aligned. Both the schedule deviation vector and the quality deviation vector are arranged in the exact same component numbering order, with the i-th element of each vector always corresponding to the fire-fighting component numbered i. During alignment, the order and values ​​of elements within the original vectors are not changed; only the consistency of the lengths of the two sets of vectors is checked to ensure that the dimensions of the schedule deviation vector and the quality deviation vector are completely equal, with no missing component data or misaligned component indices. After alignment, the two sets of one-dimensional vectors are horizontally concatenated to form an initial high-dimensional deviation data matrix. Each row of this initial matrix corresponds to a single fire-fighting component, and each row contains two basic deviation indicators: the component schedule deviation value and the component quality deviation value. The initial matrix retains all original deviation data without any data compression or feature filtering. However, it contains a small amount of data redundancy and noise interference, requiring dimensionality reduction optimization using principal component analysis.

[0046] Principal Component Analysis (PCA) is an unsupervised linear dimensionality reduction algorithm. Its core principle is to transform correlated original deviation indicators into uncorrelated comprehensive principal component indicators through orthogonal transformation, eliminating invalid noise features with low contribution rates, and retaining effective features that reflect the core changing patterns of progress and quality deviations. Simultaneously, it compresses data dimensionality and simplifies the computational burden on subsequent neural networks. During algorithm execution, three core parameters are set: parameter C_1 represents the covariance matrix normalization coefficient, with a fixed value of 1 / (n-1), where n represents the total number of fire-fighting components, used to eliminate the influence of the number of component samples on the covariance calculation results; parameter R_1 represents the cumulative contribution rate threshold of the principal components, set to 95% in this case, meaning that principal components with a cumulative feature value ratio of 95% are retained, ensuring that the dimensionality-reduced data can restore more than 95% of the deviation information of the original data; parameter D_2 represents the target dimensionality reduction dimension. Considering the characteristics of the project's fire-fighting component deviation data, the final fusion matrix column dimension is set to 3 columns to balance data integrity and computational lightweight requirements.

[0047] The algorithm's execution flow consists of five steps: standardization of the original deviation data, covariance matrix calculation, eigenvalue and eigenvector decomposition, principal component selection, and data projection dimensionality reduction. The first step standardizes the initial deviation data after splicing, eliminating computational interference caused by the inconsistency in the dimensions of schedule deviation (time unit) and quality deviation (length unit), mapping all data uniformly to the 0-1 interval. The second step calculates the covariance matrix of the standardized data to uncover the potential linear correlation between schedule deviation and quality deviation. The third step decomposes the covariance matrix to obtain all eigenvalues ​​and corresponding orthogonal eigenvectors; the magnitude of the eigenvalues ​​represents the data information carrying capacity of the corresponding principal components. The fourth step sorts the eigenvalues ​​from largest to smallest and selects principal components sequentially until the cumulative contribution rate reaches 95%. The fifth step projects the original high-dimensional deviation data onto the selected principal component space, completing data dimensionality reduction and fusion.

[0048] After dimensionality reduction, a final deviation fusion matrix is ​​generated. This matrix is ​​a two-dimensional structured matrix, with each row uniquely corresponding to a fire protection component, and the row number perfectly matching the previous component index. Each column corresponds to a type of fused comprehensive deviation index. The three columns respectively correspond to schedule-driven comprehensive deviation indices, quality-driven comprehensive deviation indices, and schedule-quality coupled comprehensive deviation indices, fully covering all deviation characteristics of single schedule deviations, single quality deviations, and cross-coupled deviations. This matrix unifies the format and dimensions of all deviation data, eliminates invalid noise data from the original vectors, retains the core information of actual construction progress and quality deviations at the construction site, and reduces the input data scale of subsequent graph neural networks, ensuring efficient and stable operation of subsequent deviation coupling relationship calculations.

[0049] S203, input the deviation fusion matrix into the graph neural network, construct an associated topology graph with fire protection components as nodes and construction process dependencies as edges, calculate the coupling propagation strength between schedule deviation and quality deviation through message passing mechanism, and output the associated risk propagation graph; Specifically, the construction process dependency table of fire protection engineering can be parsed, each fire protection component can be treated as a node, the dependencies between processes can be treated as directed edges, and the deviation features of each node in the deviation fusion matrix can be used as the initial node attributes to generate an associated topology graph structure. The core of this step is to complete the format conversion of engineering data and graph structure data at the fire protection construction site. This involves transforming the linear, written construction procedure logic and discrete component deviation values ​​into topological graph structured data that can be directly recognized by graph neural networks. This achieves a one-to-one binding of physical fire protection components, construction procedure logic, and schedule and quality deviation data, building a basic graph network carrier adapted for coupled risk analysis in fire protection engineering construction. The specific implementation method is as follows: The fire protection engineering construction procedure dependency table is a dedicated logical ledger of procedures compiled based on the project's construction organization design. Unlike general construction schedule tables, this table focuses on the strong constraints of fire protection construction, comprehensively recording the pre-construction conditions, subsequent linkage requirements, and parallel construction boundaries for all procedures, including pipe pre-embedding, fire damper installation, smoke exhaust fan connection, fire alarm module wiring, and emergency lighting installation. The table's core contains two types of fixed parameters: the pre-related component code S_prev and the subsequent related component code S_next. The parameter S_prev refers to... The parameter S_next refers to the upstream fire protection components that require priority completion and acceptance, while the parameter S_next refers to the downstream fire protection components that must wait for the upstream components to be completed before construction can begin. Both types of parameters use a unified six-digit component code for the project. The first two digits are the fire protection component category code, and the last four digits are the component serial number. In the example, the code for the main fire sprinkler pipe is XF010001, the code for the branch sprinkler pipe is XF010002, and the code for the photoelectric smoke detector is XF020001. The coding rules ensure that each physical fire protection component on site has a unique identifier and there is no problem of duplicate codes.

[0050] The topology graph nodes are divided into units based on the smallest physical fire protection components. No simplification processes such as process merging or component classification are performed to ensure the precision of subsequent risk propagation calculations. In this calculation, 28 core fire protection components that have been installed or are yet to be installed at the construction site are selected as the total nodes of the graph network. The total number of nodes is identified as N, which is fixed at 28. Each independent node corresponds to a unique node number V_n, and the value of n ranges from 1 to 28. This achieves a one-to-one mapping between the virtual nodes in the graph and the real physical components at the construction site, avoiding any discrepancies between the topology structure and the on-site construction conditions.

[0051] The directed edges in the topology graph are automatically generated based on the process dependency table. Unlike undirected connecting edges, this project uses unidirectional directed edges to align with the irreversible sequential logic of construction. Directed edges uniformly point from upstream preceding component nodes to downstream subsequent component nodes, intuitively representing the propagation of construction risks along the process flow. An initial inherent weight w_0 is set for each directed edge, with a value of 0.6. This parameter characterizes the inherent tightness of the process connections, and its value ranges from 0 to 1. A value closer to 1 indicates a higher degree of binding between the two construction processes, meaning deviations in upstream components more easily affect the construction quality and progress of downstream components. A value closer to 0 indicates that the two processes are independent, with no significant propagation effect from deviations. For parallel fire protection construction processes that can be carried out simultaneously on the construction site without sequential constraints, no directed edges are generated between the corresponding component nodes, eliminating invalid risk propagation paths and reflecting real-world construction conditions.

[0052] The initial node attributes directly reuse the deviation fusion matrix data processed by dimensionality reduction mentioned earlier. All features of component progress deviation and quality deviation corresponding to each row of the deviation fusion matrix are directly assigned to the corresponding numbered graph node as the node's initial inherent attributes. In this case, the single node attribute dimension d=6, where d represents the total number of deviation features carried by a single node. The six features are: quantified value of progress lag deviation, quantified value of progress advance deviation, component spatial position installation deviation, component 3D installation posture deviation, component external dimension deviation, and component splicing fit deviation, fully covering the two major control dimensions of construction progress deviation and construction quality deviation. After completing the three operations of creating all nodes, connecting directed edges, and assigning node deviation attributes, a complete initial associated topology graph structure is finally generated. This structure completely replicates the on-site construction process logic and the real-time component deviation status, and can be directly input into the subsequent graph neural network for deviation coupling calculations.

[0053] The associated topology graph structure is input into the graph neural network. The deviation information of neighboring nodes is aggregated through graph convolutional layers. A multi-head attention mechanism is used to calculate the weight of the impact of progress deviation on quality deviation between different nodes, and a coupling strength matrix between nodes is generated. The core of this step is to solve the feature extraction problem of irregular construction topology data by relying on graph convolutional networks, aggregating the global deviation information of adjacent component nodes, and then adaptively quantifying the bidirectional influence of schedule deviation and quality deviation between different components through a multi-head attention mechanism. This transforms the fuzzy coupling relationship between processes into directly computable quantitative weight values, and finally outputs a standardized coupling strength matrix, completing the transformation from qualitative process correlation to quantitative deviation coupling values. The specific implementation method is as follows: This study uses a two-layer graph convolutional neural network (GCN) as the basic computational model. The network consists of three modules: an input layer, a two-layer graph convolutional layer, and an attention fusion layer. The input layer directly receives the associated topology graph generated in the previous step, adapting to the non-Euclidean spatial data characteristics of fire protection construction processes, which have no fixed grid and random node connections. This overcomes the limitation of traditional convolutional neural networks in handling irregular topology data. The two-layer graph convolutional layer is divided into a first-layer graph convolutional layer (GCN_1) and a second-layer graph convolutional layer (GCN_2). The two convolutional layers work together to aggregate deviation information from neighboring nodes at different distances. The first-layer graph convolutional layer (GCN_1) is responsible for the first-order deviation aggregation of direct neighbor nodes, that is, the fusion of deviation data between upstream and downstream component nodes with direct process connections. The convolutional kernel size is set to 3, representing that a single aggregation operation simultaneously includes the deviation data of the current computing node and the two directly related neighboring nodes before and after it. The second-layer graph convolutional layer (GCN_2) is responsible for the second-order deviation aggregation of indirect neighbor nodes, capturing the implicit deviation relationships between long-distance components across a construction process, covering the potential risk impact of long-link construction processes.

[0054] Both graph convolutional layers are configured with a bias normalization coefficient λ=0.01. This parameter is used to constrain the numerical growth rate after the multiple node biases are superimposed, avoiding numerical overflow and bias feature distortion problems after the bias data of multiple neighboring nodes are superimposed, and ensuring that the aggregated node bias features always remain within a stable computational range. After processing by the two graph convolutional layers, each graph node completes the global information fusion of its own original bias, direct neighbor bias, and indirect neighbor bias, obtaining an initial aggregated bias feature containing local and global correlation information.

[0055] The multi-head attention mechanism is the core algorithm for accurately calculating the interaction weights of deviations between nodes. Compared to the single-head attention mechanism, which can only capture single-dimensional correlations, this algorithm can simultaneously mine the bidirectional coupling relationship between schedule deviations and quality deviations from multiple independent feature subspaces. In this study, the number of parallel attention heads is set to h=4, where h represents the number of feature subspaces that are independently processed in parallel. The four attention heads correspond to four coupling scenarios: the negative impact of schedule lag on quality, the positive impact of schedule advancement on quality, the linkage effect of spatial installation deviations, and the linkage effect of posture installation deviations. This comprehensively covers all deviation interaction paths on the construction site. Simultaneously, an attention temperature smoothing coefficient τ=0.2 is configured. This parameter is used to optimize the distribution of attention weights, avoid extremely large weight values ​​for individual component nodes, balance the proportion of mutual influence between nodes across the entire domain, and make the weight calculation results more consistent with the balanced working conditions of multiple components mutually constraining each other on the construction site.

[0056] During the weighting process, the normalized progress deviation features of each node are used as the query vector, and the quality deviation features of the global neighbor nodes are used as the key vector and value vector, respectively. The four attention heads independently complete the interaction score calculation between nodes. Then, the outputs of the four attention heads are concatenated, and the feature dimensions are unified through a fully connected layer. Finally, the bidirectional interaction weights between any two component nodes V_i and V_j are obtained, where W_ij represents the positive influence weight of the progress deviation of node V_i on the quality deviation of node V_j, and W_ji represents the negative influence weight of the progress deviation of node V_j on the quality deviation of node V_i.

[0057] Finally, all bidirectional interaction weights are integrated to generate a node coupling strength matrix. This matrix is ​​a 28-row, 28-column symmetric square matrix, with row and column numbers completely corresponding to the node numbers in the topology graph. The diagonal elements of the matrix are uniformly fixed at 1, representing that the progress deviation of a single component has a 100% inherent impact on its own quality deviation. The off-diagonal elements are dynamic interaction weights output by the multi-head attention mechanism. All matrix element values ​​are uniformly constrained to the range of 0 to 1. The larger the value, the higher the coupling degree between the progress and quality deviations of the corresponding two fire protection components, and the stronger the risk transmission ability. Finally, the normalized node coupling strength matrix is ​​output, providing a precise quantitative calculation basis for subsequent multiple rounds of message passing iterations.

[0058] Based on the inter-node coupling strength matrix, multiple rounds of message passing iterations are performed to propagate the quality deviation of each node to adjacent nodes according to the coupling weight, while the progress deviation is propagated in reverse to generate the node deviation features after propagation. The core of this step is to rely on the quantified coupling strength matrix and simulate the real risk flow pattern on the construction site through the native message passing mechanism of graph neural networks. It forward reconstructs the process of quality defects spreading along the construction process, and backward reconstructs the process of delays forcing rectification of previous processes, leading to secondary quality problems. After multiple rounds of iteration to eliminate local calculation errors in single propagation, an updated version of the node deviation feature that considers the linkage effect of all components in the entire domain is finally obtained. The specific implementation method is as follows: The graph neural network message passing mechanism is the core computational logic for realizing the dynamic flow of risks between nodes in the topology graph. Unlike the simple computation method of directly superimposing fixed values, this mechanism strictly follows the dynamic weights within the coupling strength matrix to complete the weighted transmission of deviation data, perfectly matching the real risk propagation patterns at fire protection construction sites: quality deviations such as pipe misalignment and equipment installation offset in upstream fire protection components will affect the installation accuracy of all downstream connected components along the construction process; schedule deviations such as delays and construction stoppages in downstream components will require rework and re-inspection of upstream components, thereby exacerbating the quality defects of upstream components and forming a two-way closed-loop risk transmission link.

[0059] This iteration process sets two core controllable parameters: the total number of iterations k across the entire domain and the deviation attenuation coefficient γ per iteration. The total number of iterations k=5, where k represents the number of iterations required for complete risk transmission across the entire domain. Considering the industry-wide common characteristic of fire protection engineering with an average of 4 to 6 processes per construction link, 5 iterations can completely cover the risk transmission path of all construction processes in short, medium, and long links. Too few iterations will miss implicit deviation associations in distant components, while too many iterations will cause excessive smoothing of the original local deviation features, losing key risk source information. The deviation attenuation coefficient γ=0.85, which characterizes the natural attenuation ratio of risk propagation between nodes. The farther the component is from the construction site, the weaker the impact of deviation transmission will be. This value represents a 15% reduction in the influence of the deviation after each node propagation, consistent with the general law of natural risk attenuation in construction sites.

[0060] The five rounds of iterative calculations are carried out simultaneously along two parallel paths: forward quality deviation propagation and reverse schedule deviation propagation. Forward propagation follows the original directed edge direction of the topology graph, retrieves the forward coupling weights in the coupling strength matrix, multiplies the current node's quality deviation value with the corresponding weight, and then adds an attenuation coefficient to complete the numerical conversion, thus transmitting the quality deviation risk to the downstream directly connected nodes. Reverse propagation is in the opposite direction to the original directed edge direction, retrieves the reverse coupling weights in the matrix, weights and converts the downstream node's schedule deviation, and then transmits the reverse quality risk caused by the schedule lag to the upstream preceding component nodes.

[0061] After each iteration, the network updates the deviation characteristics of all nodes in real time, adding the deviation increments transmitted from upstream and downstream nodes in the current round, while fully preserving the original basic deviation data of each node to prevent the loss of original deviation information. After all five iterations are completed, the system initiates weak noise filtering, setting the noise filtering threshold to 0.02. Minimal cross-node deviation interference items below this threshold generated during the iteration process are removed, reducing the interference of invalid, small data on the overall deviation characteristics. After complete bidirectional multi-round message passing operations, each graph node integrates four types of data: its own original deviation, the deviation of its direct neighbor nodes, the deviation of distant indirect nodes, and the bidirectional propagation deviation increment. This ultimately generates the node deviation characteristics propagated under the influence of the interconnected effects of all components. This characteristic can fully reflect the interconnected impact of a single component deviation on the entire fire protection construction network.

[0062] The propagated node deviation features are superimposed with the original edge direction information to generate a risk propagation graph. In the graph, the node color represents the comprehensive risk level, and the edge thickness represents the coupling propagation strength.

[0063] The core of this step is to transform network abstract numerical data into engineering visualization graphics. This involves fusing and matching the iteratively updated node deviation values ​​and inter-node coupling propagation strength with the fixed process edge direction information of the original topology diagram to build an intuitive and easy-to-understand visualized risk map. This allows the backend neural network computation data to be directly read and used by on-site construction management personnel, accurately locating risk sources, risk propagation paths, and the intensity of risk impact. The specific implementation method is as follows: First, the two types of core data were fused and overlaid. These two types of basic data were the node deviation features updated after multiple iterations and the original directed edge direction information retained from the initial topology graph. The original edge direction information was retained without modification throughout the process, strictly matching the original sequence of construction procedures to ensure that the risk propagation flow within the visualized graph was completely consistent with the actual construction flow on site, preventing any issues where the risk propagation direction contradicted the construction logic. The data overlay adopted a weighted fusion mode. The final coupling propagation strength between nodes was calculated jointly by the inherent edge weights of the initial procedures and the dynamic coupling weights after iterations. The initial static procedure weights accounted for 30%, while the dynamic deviation coupling weights obtained from network computation accounted for 70%. This approach preserved the inherent correlation attributes of the construction procedures while also reflecting the dynamic risk changes brought about by real-time construction deviations, ensuring that the coupling strength data balanced static process logic and dynamic on-site conditions.

[0064] Subsequently, node visual attribute mapping was carried out, and the comprehensive risk value R of a single component was calculated based on the node deviation characteristics after propagation. The comprehensive risk value adopted a weighted calculation method, with the normalized value of schedule deviation accounting for 40% and the normalized value of quality deviation accounting for 60%, which is in line with the industry standard that quality control takes precedence over schedule control in fire protection engineering construction management. The comprehensive risk value R is divided into five risk levels based on a range of 0 to 1, and each level is assigned a specific node with a corresponding color, enabling intuitive and barrier-free identification of the risk level: Level 1, low risk, R value range 0 to 0.2, node filled with green, indicating that the deviations in component progress and quality are within the compliant tolerance range, requiring no on-site intervention and only routine monitoring; Level 2, relatively low risk, R value range 0.2 to 0.4, node filled with light yellow, indicating that there are slight construction deviations in the components, which will not affect the overall construction acceptance standards and do not require rectification; Level 3, medium risk, R value range 0.4 to 0.6, node filled with dark yellow, indicating that the deviation is approaching the upper limit of the tolerance threshold, requiring an early warning; Level 4, relatively high risk, R value range 0.6 to 0.8, node filled with orange, indicating that the construction deviation has exceeded the standard, requiring localized small-scale rework and rectification; Level 5, high risk, R value range 0.8 to 1.0, node filled with red, indicating that the component deviation is seriously exceeded, belonging to the source of risk across the entire domain, requiring immediate work stoppage and rectification. In the example, the comprehensive risk value R of the main sprinkler pipe node is 0.86, matching the red high-risk node; the comprehensive risk value R of the ordinary emergency lighting component node is 0.18, matching the green low-risk node.

[0065] Finally, a directed edge visual attribute mapping was performed, using the final coupling propagation strength value between nodes as the criterion for determining the thickness of the edge lines. The coupling propagation strength range is also from 0 to 1, and this range is divided into five thickness levels. The larger the coupling propagation strength value, the thicker the corresponding directed edge line, representing higher efficiency in the transmission of deviation risk between the two components and stronger mutual constraints; the smaller the coupling propagation strength value, the thinner the corresponding directed edge line, representing weaker risk transmission between components. At the same time, all directed edges retain their original arrow directions, clearly marking the propagation path of risk from the source component to the upstream and downstream components.

[0066] After completing the visualization mapping and data overlay of node color, edge thickness, and edge direction, a complete association risk propagation map is finally generated. This map fully retains all core information such as the construction process topology, bidirectional risk propagation path, single component risk level, and risk coupling strength between components. It transforms the complex network calculation values ​​in the background into a visual graphic that can be directly used on the engineering site, providing intuitive and accurate visual data support for subsequent construction deviation trend prediction and construction critical path replanning.

[0067] S204. Based on the aforementioned associated risk propagation diagram, a long short-term memory network is used to predict the deviation evolution trend within the future preset construction period. Combined with the resource availability constraints in the preset construction plan, critical path replanning is performed to generate a set of control instructions that includes progress catch-up strategies and quality problem location, and these instructions are sent to the construction management terminal to achieve dynamic control of the construction progress and quality of fire protection engineering.

[0068] Specifically, the deviation time series of each node in different time windows can be extracted from the associated risk propagation map, and input into a pre-trained long short-term memory network model. The model outputs the predicted values ​​of progress deviation and quality deviation at a future preset time, generating a deviation evolution trend prediction curve. The core of this step is to rely on the previously constructed risk propagation map to extract the dynamic deviation data of different fire protection component nodes over time. Then, using a long and short time memory network with time-series data retention capabilities, the step captures the patterns of deviation changes, accurately predicts the trends of schedule and quality deviations in the subsequent construction cycle, and generates a visualized trend curve. This provides time-series data support for subsequent risk identification. The specific implementation method is as follows: Each independent node in the associated risk propagation diagram corresponds to a single fire protection construction component at the construction site, covering all categories of fire protection construction components such as sprinkler pipes, fire dampers, smoke exhaust vents, and fire pumps. All nodes are equipped with progress deviation characteristics and quality deviation characteristics in the continuous time dimension. The time window is the core scale for dividing time-series data. This time, three types of progressive time windows are set: historical backtracking window W_1, transition buffer window W_2, and future prediction window W_3. Among them, W_1 is 14 days, representing the deviation data obtained from continuous monitoring of the component over the past 14 days. W_2 is 2 days, used to smooth the time series gap between historical data and prediction data and eliminate data fluctuation interference caused by temporary construction conditions. W_3 is 9 days, corresponding to the future preset construction period that needs to be predicted this time. Based on three types of time windows, deviation time series are extracted node by node. Each time series contains daily synchronously collected progress deviation and quality deviation values. The progress deviation value is quantified in days, representing the lag time of a single fire protection construction procedure relative to the planned node. The quality deviation value is quantified in millimeters, representing the spatial offset of the actual installation position and posture of the fire protection components relative to the BIM design model. All time series data carry the original collection timestamp to ensure that the time series arrangement is not disordered or misaligned.

[0069] Long Short-Term Memory (LSTM) networks are deep learning time-series prediction models specifically designed to address the problem of long-term dependencies in time-series data. They effectively circumvent the gradient vanishing defect inherent in traditional recurrent neural networks and are well-suited for predicting long-cycle, continuously changing time-series data such as construction deviations. The LSTM network used in this study was a mature model that had undergone offline pre-training. During the pre-training phase, a historical construction deviation time-series dataset of nearly 80 days from similar fire protection engineering projects in the same region was used as training samples. The model's basic hyperparameters were uniformly fixed, with the learning rate (lr_1) set to 0.001. This parameter represents the update step size of the weight parameters in each iteration of the network; a smaller value results in higher model fitting accuracy but slower training speed. 0.001 represents the optimal balance for time-series prediction scenarios. The number of hidden layer neurons (hidden_1) was set to 128 to store historical deviation time-series features, ensuring the network can retain them completely. The model stores subtle deviation changes within a 14-day window; the number of iterative training epochs (epoch_1) is set to 80 to ensure that the model fully fits the inherent variation pattern of fire construction deviations; at the same time, the network has three built-in gate control structures: the forget gate threshold f_1 is set to 0.7, which is responsible for filtering and removing invalid noise deviation data caused by sudden working conditions such as rainy day construction and temporary equipment maintenance in historical time series data; the input gate threshold i_1 is set to 0.6, which is responsible for accurately inputting the latest effective deviation feature data collected each day; and the output gate threshold o_1 is set to 0.8, which is responsible for normalizing the processed time series features and outputting standardized deviation prediction results.

[0070] The extracted time-series deviation sequences of each component node are uniformly normalized to map all schedule and quality deviation values ​​to a range of 0 to 1, eliminating prediction interference caused by the different units of days and millimeters. The normalized time-series data is then batch-input into a pre-trained LSTM model. Using 14 days of historical time-series data as input, the model sequentially outputs predicted schedule deviation values ​​P_pre and predicted quality deviation values ​​Q_pre for three preset construction periods: 3 days, 6 days, and 9 days. Here, P_pre_1 represents the predicted schedule deviation value for the first day of the future, and Q_pre_1 represents the predicted quality deviation value for the first day of the future, and so on, completing the prediction for all time periods. Finally, using construction time as the horizontal axis and deviation quantification value as the vertical axis, we plotted the independent deviation evolution curve of a single component and the comprehensive deviation evolution curve of the entire project. This fully presents three typical evolution trends of deviation: continuous increase, stable fluctuation, and gradual decline. The inflection point of deviation mutation for each curve is clearly marked, which intuitively reflects the coupling change law of schedule deviation and quality deviation as construction progresses. This provides a visualized and quantifiable time-series basis for subsequent identification of over-limit risks.

[0071] Based on the deviation evolution trend prediction curve, identify components that are about to exceed the tolerance threshold and their deviation types, and combine them with the resource availability constraints in the preset construction plan, including the remaining amount of manpower, equipment and materials, to generate a resource conflict list. The core of this step is to accurately screen fire protection components with the risk of exceeding limits based on the predicted deviation evolution curve, and distinguish between two types of deviations: schedule exceedance and quality exceedance. Then, combined with the remaining amount of existing construction resources on the construction site, the resource gap required for rectification of deviations and catching up with the schedule is calculated, the severity of various resource conflicts is quantified, and complete resource conflict data is formed to provide hard resource constraints for subsequent replanning of construction paths. The specific implementation method is as follows: First, a unified construction deviation tolerance threshold is preset for the project, divided into two fixed control indicators: a schedule deviation tolerance threshold T_p and a quality deviation tolerance threshold T_q. The schedule deviation tolerance threshold T_p is set to 2 days, representing the maximum allowable planned delay for a single fire protection construction procedure. A delay exceeding this value is considered an excess of schedule risk. The quality deviation tolerance threshold T_q is set to 3 millimeters, representing the maximum allowable spatial position and orientation deviation of fire protection components during installation. A deviation exceeding this value is considered an excess of quality risk. The slope K of the deviation evolution trend curve is further used to classify the urgency of the risk. The slope K represents the average daily increase in deviation. K greater than 0.5 indicates a rapidly deteriorating deviation, classifying it as a high-risk component about to exceed the limit; K between 0.2 and 0.5 indicates a slow increase in deviation, classifying it as a medium-risk critically exceeding-limit component; and K less than 0.2 indicates a gradual increase in deviation, classifying it as a low-risk, controllable component. At the same time, two types of over-limit warning states are distinguished: when the curve value is in the range of 80% to 100% of the threshold, it is a near over-limit warning; when the curve value exceeds the threshold, it is a complete over-limit warning. Risk screening is carried out on all fire protection components one by one, and the three core information of deviation type, risk level and expected over-limit time corresponding to each risk component are recorded simultaneously.

[0072] The availability constraints of construction site resources are divided into three core categories: manpower, construction equipment, and engineering materials. Each category of resources is equipped with standardized statistical parameters to accurately represent the remaining available resources on site. Human resources are represented by the remaining effective working hours H_res, with the unit being working hours. A single working hour represents one hour of continuous effective work by a professional construction worker. This parameter excludes ineffective working time such as lunch breaks and on-site safety briefings, accurately reflecting the available manpower reserves on site. Construction equipment resources are represented by the remaining effective shifts E_res, with the unit being shifts. A single shift represents eight hours of continuous effective work for fire protection construction scanners and pipe installation tools. The statistical process excludes time occupied by routine equipment maintenance and fault repair. Engineering material resources are represented by the remaining number of standard components M_res, with the unit being pieces, covering standardized consumables for fire protection construction such as fire sprinklers, sealing fittings, and fireproof sealing strips.

[0073] Simultaneously, pre-set standard resource consumption quotas for expedited rectification are established, clearly defining the fixed resource consumption required to resolve deviations: When carrying out expedited progress work, the standard daily manpower consumption quota H_con for a single process is 4 man-hours, the standard equipment consumption quota E_con is 0.5 machine-hours, and there is no additional material consumption; when carrying out rework rectification work for quality deviations, the standard daily manpower consumption quota H_re for a single component is 6 man-hours, the standard equipment consumption quota E_re is 0.8 machine-hours, and the material consumption quota M_re is 1 piece. The total resource requirement for rectifying risky components is subtracted from the existing remaining resources on site to calculate the manpower gap ΔH, equipment resource gap ΔE, and material resource gap ΔM respectively. The conflict level is then classified according to the proportion of the resource gap to the total requirement: a gap of less than 30% is a minor resource conflict, requiring only a small increase in resources to complete the rectification; a gap between 30% and 60% is a moderate resource conflict, requiring adjustments to the work process to supplement resources; a gap greater than 60% is a severe resource conflict, requiring a reconstruction of the construction process path to accommodate existing resources. Finally, all information, including risk component number, deviation type, risk level, three types of resource gap values, and resource conflict level, is integrated to form a complete and coherent resource conflict dataset with no data omissions and no indicator ambiguity processing, supporting the optimization and replanning of subsequent construction critical paths.

[0074] Based on the resource conflict list, the critical path method is used to replan the construction plan, adjust the execution sequence of the work process and add parallel operations, and generate a schedule catch-up strategy and a list of quality problem locations. The core of this step is to utilize the critical path method, a mature approach in engineering, to optimize the existing sequential construction process network based on the current resource conflicts on site. It addresses schedule delays through two methods: adjusting the sequence of processes and adding parallel operations. Simultaneously, it uses the results of previous risk propagation and tracing to accurately pinpoint the root causes of quality problems, outputting feasible schedule recovery plans and layered quality problem tracing results. The specific implementation method is as follows: The Critical Path Method (CPM) algorithm is a core algorithm for project schedule control and process optimization in construction projects. Based on the inherent pre- and post-process dependencies among all construction processes in fire protection engineering, it constructs a complete construction process network topology and automatically calculates four core timing parameters for each process: earliest start time (ES), latest start time (LS), earliest finish time (EF), and latest finish time (LF). These parameters are then used to calculate the total float time (TF) and free float time (FF). The total float time (TF) represents the maximum delay allowed for a single process without affecting the overall project completion date; the free float time (FF) represents the maximum delay allowed for a single process without affecting the normal start of its successor processes. All processes with a total float time of 0 are sequentially connected to form the original critical path of the project. Delays in any process on the critical path directly extend the overall project duration and are the core optimization targets of this process replanning. Non-critical path processes have float time margins and do not directly affect the overall project duration; they are only used as auxiliary optimization targets.

[0075] Based on the resource conflict list obtained in Step 2, priority is given to optimizing and adjusting processes with resource conflicts and excessive deviations on the critical path. Optimization methods are divided into two categories: adjusting the sequence of processes and adding parallel construction operations. For core fire protection processes with strong sequential logical dependencies that cannot be split into parallel operations, such as pipeline pressure testing which must be carried out after pipeline installation, the maximum floating time limit for a single period transfer (TF_max) is set at 1.5 days, relying solely on the floating time of non-critical processes to transfer schedule pressure. This avoids excessive transfer of floating time, which could cause non-critical processes to become new critical processes, leading to secondary schedule risks. For fire protection processes with independent work surfaces and no mandatory sequential dependencies, such as fire box wiring and fireproof sealing of pipeline outer walls, parallel operation modes are directly added. Considering the space constraints of the construction site, the maximum number of parallel processes for the project is set at 3, avoiding on-site construction congestion and new quality hazards caused by overlapping construction due to multiple parallel processes. A single compliant parallel process can reduce the overall construction period (ΔT_c) by 0.8 days per day.

[0076] Based on the results of process optimization, a graded schedule catch-up strategy is generated, and a corresponding catch-up plan is matched according to the risk level of the components: high-risk over-limit components are matched with an expedited catch-up plan that combines emergency resource addition and parallel process operation to quickly curb the continuous deterioration of deviations; medium-risk critical over-limit components are matched with a conventional catch-up plan that combines time sequence fine-tuning and local resource supplementation to smoothly resolve potential schedule risks; low-risk controllable components are matched with a routine inspection and control plan, which does not require adjustment of processes, only an increase in monitoring frequency. Simultaneously, by combining the node coupling propagation strength of the associated risk propagation graph, the precise location of quality problems is achieved in layers. The first layer is the original quality problem, namely the initial installation deviation of the component itself caused by the non-standard operation of the construction personnel or the unqualified acceptance of the component upon arrival, without the influence of external processes. The second layer is the quality problem derived from the coupling of processes, namely the secondary installation offset caused by the deviation generated by the previous construction process propagating to the current component through the topological associated edge. The third layer is the global chain quality risk problem, namely the large-scale construction quality hazard formed by the mutual coupling and diffusion of deviations of multiple nodes. The source starting node, deviation propagation path and rectification difficulty coefficient of each type of quality problem are marked simultaneously. Finally, all schedule optimization plans and quality traceability results are integrated to form a complete schedule catch-up strategy and quality problem location list.

[0077] The schedule catch-up strategy and the list of quality issues are encapsulated into a set of control instructions, including the component numbers that need to be reworked, the amount of additional resources, and the adjusted schedule milestones. These instructions are then pushed to the on-site engineers through the construction management terminal to achieve dynamic control.

[0078] The core of this step is to standardize and structure the previously generated schedule optimization plan, quality rectification plan, and resource replenishment plan into unified control instructions that can be directly executed by on-site construction personnel. This enables real-time push notifications to terminals, constructing a closed-loop control process from data collection, deviation analysis, risk prediction, path optimization to on-site execution. Ultimately, this achieves two-way dynamic control of construction progress and quality. The specific implementation method is as follows: The control instruction set adopts a three-layer modular encapsulation structure. The three modules are interconnected and each performs its own function, fully adapting to the automatic parsing logic of the on-site construction management terminal. These are the component rectification basic information layer, the on-site resource execution layer, and the global schedule control layer. The component rectification basic information layer uniformly enters the numbers of all fire protection components that require rework, rectification, or expedited construction. The project adopts standardized component coding rules, with the coding format being building area code-process code-component serial number. For example, code 2-04-18 represents the 18th smoke exhaust vent component under the fire duct installation process of Building 2. This module fully marks the deviation type, risk level, and standard rectification process requirements corresponding to each rectification component, allowing on-site engineers to intuitively understand the rectification objects and construction standards.

[0079] The on-site resource execution layer accurately calculates the additional manpower, equipment, and material resources required for each rectification task. To avoid additional resource losses caused by unexpected construction conditions on site, a resource redundancy coefficient γ=1.1 is added on top of the theoretical resource gaps ΔH, ΔE, and ΔM. The actual additional manpower hours required are ΔH×γ, the actual additional equipment shifts are ΔE×γ, and the actual additional material quantity is ΔM×γ. At the same time, the precise arrival time of each type of resource, the on-site construction area, and the person responsible for resource handover are clearly defined to eliminate ambiguity in on-site resource allocation and ensure that resources can be accurately matched with the needs of the rectification process.

[0080] The global project schedule control layer outputs all updated schedule nodes after the critical path replanning is completed, including the earliest start time, latest completion time, and interval duration of each process after adjustment. It also outputs the total construction period after the overall project update, calculates the overall schedule reduction ΔT_total by comparing it with the original planned schedule, and simultaneously marks the updated critical path nodes and high-risk component locations that require intensive monitoring. It clarifies the monitoring frequency of laser scanning and panoramic image acquisition at the construction site, increases the monitoring frequency of high-risk component areas to once every 4 hours, and maintains a monitoring frequency of twice a day in regular construction areas, continuously updates the site status data stream, and realizes dynamic iteration of control data.

[0081] After the integration of the three-layer modules, the entire control command set is encapsulated. Data downlink is pushed using a dedicated industrial wireless communication network at the construction site, maintaining a stable transmission bandwidth of 10Mbps. The transmission latency for a single control command is controlled within 200ms, ensuring that control commands are delivered to the on-site engineer's handheld construction management terminal without delay. Upon receiving the command, the terminal automatically splits the data from the three modules, clearly displaying the rework component list, resource addition list, and schedule adjustment list on separate interfaces, eliminating the need for engineers to perform secondary data parsing. The entire control process forms a closed-loop dynamic control link: on-site data collection – deviation fusion analysis – risk coupling assessment – ​​deviation trend prediction – construction path replanning – on-site command execution – new round of on-site data collection. This continuously mitigates the risk of coupled propagation of schedule and quality deviations during construction, dynamically adapting to uncertainties such as changes in on-site resources and operating conditions. Ultimately, the entire process is automated and intelligent, achieving integrated dynamic control of fire protection engineering construction progress and quality.

[0082] Another embodiment of the present invention provides a BIM-based fire protection engineering construction progress and quality control system, see [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to load the building information model and preset construction plan of the fire protection project, and to use a laser scanner and panoramic camera to collect component point cloud data and image data at the construction site, and generate a time-stamped site status data stream. The extraction module 302 is used to perform point cloud registration and feature mapping between the on-site status data stream and the building information model. By comparing the actual component position, installation posture and model design value, and combining the process nodes in the preset construction plan, the degree of deviation is determined, and the progress deviation vector and quality deviation vector are extracted respectively to generate a deviation fusion matrix. The construction module 303 is used to input the deviation fusion matrix into the graph neural network, construct an associated topology graph with fire protection components as nodes and construction process dependencies as edges, calculate the coupling propagation strength between schedule deviation and quality deviation through a message passing mechanism, and output an associated risk propagation graph. The generation module 304 is used to predict the deviation evolution trend within the future preset construction period based on the associated risk propagation map using a long short-term memory network, and to perform critical path replanning in combination with the resource availability constraints in the preset construction plan, generate a set of control instructions including progress catch-up strategies and quality problem location, and send them to the construction management terminal to realize dynamic control of the construction progress and quality of fire protection engineering.

[0083] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0084] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0085] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0086] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A BIM-based method for controlling the construction progress and quality of fire protection engineering, characterized in that, The method includes: Load the building information model and preset construction plan of the fire protection project, and use laser scanners and panoramic cameras to collect component point cloud data and image data at the construction site to generate a time-stamped site status data stream; The on-site status data stream is registered with the building information model using point cloud and feature mapping. By comparing the actual component position and installation posture with the model design value, and combining the process nodes in the preset construction plan, the degree of deviation is determined. The progress deviation vector and quality deviation vector are extracted respectively to generate a deviation fusion matrix. The deviation fusion matrix is ​​input into a graph neural network. A correlation topology graph is constructed with fire protection components as nodes and construction process dependencies as edges. The coupling propagation strength between schedule deviation and quality deviation is calculated through a message passing mechanism, and a correlation risk propagation graph is output. Based on the aforementioned risk propagation graph, a long short-term memory network is used to predict the evolution trend of deviations within the preset construction period. Combined with the resource availability constraints in the preset construction plan, critical path replanning is performed to generate a set of control instructions that includes progress catch-up strategies and quality problem location, which are then sent to the construction management terminal to achieve dynamic control of the construction progress and quality of fire protection engineering.

2. The method according to claim 1, characterized in that, The loading of the building information model and preset construction plan for the fire protection project, and the use of laser scanners and panoramic cameras to collect component point cloud data and image data at the construction site, generate a time-stamped site status data stream, including: Load the building information model file of the fire protection project from the project database, parse the position coordinates and geometric parameters of each component in the model, and read the list of process nodes in the preset construction plan to generate the model-plan basic dataset; Laser scanners are deployed at the construction site to perform multi-station scanning of the constructed area according to the preset scanning stations, and the original point cloud data of each station is obtained. At the same time, panoramic cameras are used to collect 360-degree image data at each scanning station to generate original point cloud and image matching data. The original point cloud and image paired data are subjected to noise reduction filtering and stitching fusion to remove outliers and moving object point clouds, and the multi-site cloud is registered to a unified coordinate system to generate a fused full-site cloud of the construction area. Add a timestamp of the acquisition time to each point in the merged panoramic view cloud of the construction area, and attach the corresponding panoramic image as a texture to it, finally generating a time-stamped on-site status data stream.

3. The method according to claim 2, characterized in that, The process involves point cloud registration and feature mapping between the on-site status data stream and the building information model. By comparing the actual component positions and installation postures with the model design values, and considering the deviation levels in the pre-set construction plan's process nodes, the process extracts progress deviation vectors and quality deviation vectors to generate a deviation fusion matrix, including: The overall point cloud in the site status data stream is coarsely registered with the design point cloud in the building information model. The initial transformation matrix is ​​estimated using the random sampling consistency algorithm, and then fine registration is performed using the iterative nearest point algorithm to generate the registered site point cloud model. The actual position coordinates and installation attitude angles of the fire protection components are extracted from the registered site point cloud model. The deviation is calculated for each component and compared with the design value of the corresponding component in the building information model to obtain the spatial deviation of each component and generate the component deviation feature vector. Based on the completion time of the process nodes in the preset construction plan, it is determined whether the currently completed components are completed on schedule. The delay time is quantified into a schedule deviation value. At the same time, the over-threshold items in the component deviation feature vector are marked as quality deviation values, and a schedule deviation vector and a quality deviation vector are generated. The schedule deviation vector and the quality deviation vector are aligned by component index and then fused into a unified deviation fusion matrix through principal component analysis. The rows of the matrix represent components and the columns represent deviation types and indicators.

4. The method according to claim 3, characterized in that, The deviation fusion matrix is ​​input into a graph neural network to construct an associated topology graph with fire protection components as nodes and construction process dependencies as edges. The coupling propagation strength between schedule deviations and quality deviations is calculated through a message passing mechanism, and an associated risk propagation graph is output, including: The construction process dependency table of the fire protection project is analyzed. Each fire protection component is treated as a node, and the dependencies between processes are treated as directed edges. The deviation characteristics of each node in the deviation fusion matrix are used as the initial node attributes to generate an associated topology graph structure. The associated topology graph structure is input into the graph neural network. The deviation information of neighboring nodes is aggregated through graph convolutional layers. A multi-head attention mechanism is used to calculate the weight of the impact of progress deviation on quality deviation between different nodes, and a coupling strength matrix between nodes is generated. Based on the inter-node coupling strength matrix, multiple rounds of message passing iterations are performed to propagate the quality deviation of each node to adjacent nodes according to the coupling weight, while the progress deviation is propagated in reverse to generate the node deviation features after propagation. The propagated node deviation features are superimposed with the original edge direction information to generate a risk propagation graph. In the graph, the node color represents the comprehensive risk level, and the edge thickness represents the coupling propagation strength.

5. The method according to claim 4, characterized in that, Based on the aforementioned risk propagation map, a long short-term memory network is used to predict the evolution trend of deviations within a preset construction period. Combined with resource availability constraints in the preset construction plan, critical path replanning is performed to generate a set of control instructions including progress catch-up strategies and quality problem localization, which is then distributed to the construction management terminal. This achieves dynamic control of the progress and quality of fire protection engineering construction, including: The deviation time series of each node in different time windows is extracted from the associated risk propagation map and input into a pre-trained long short-term memory network model. The model outputs the predicted values ​​of progress deviation and quality deviation at a future preset time and generates a deviation evolution trend prediction curve. Based on the deviation evolution trend prediction curve, identify components that are about to exceed the tolerance threshold and their deviation types, and combine them with the resource availability constraints in the preset construction plan, including the remaining amount of manpower, equipment and materials, to generate a resource conflict list. Based on the resource conflict list, the critical path method is used to replan the construction plan, adjust the execution sequence of the work process and add parallel operations, and generate a schedule catch-up strategy and a list of quality problem locations. The schedule catch-up strategy and the list of quality issues are encapsulated into a set of control instructions, including the component numbers that need to be reworked, the amount of additional resources, and the adjusted schedule milestones. These instructions are then pushed to the on-site engineers through the construction management terminal to achieve dynamic control.

6. A BIM-based fire protection engineering construction progress and quality control system, characterized in that, The system includes: The data acquisition module is used to load the building information model and preset construction plan of the fire protection project, and to use laser scanners and panoramic cameras to collect point cloud data and image data of components at the construction site, generating a time-stamped on-site status data stream. The extraction module is used to perform point cloud registration and feature mapping between the on-site status data stream and the building information model. By comparing the actual component position, installation posture and model design value, and combining the process nodes in the preset construction plan, the degree of deviation is determined, and the progress deviation vector and quality deviation vector are extracted respectively to generate a deviation fusion matrix. The construction module is used to input the deviation fusion matrix into the graph neural network, construct an associated topology graph with fire protection components as nodes and construction process dependencies as edges, calculate the coupling propagation strength between schedule deviation and quality deviation through a message passing mechanism, and output an associated risk propagation graph. The generation module is used to predict the deviation evolution trend within a preset construction period based on the associated risk propagation map using a long short-term memory network, and to perform critical path replanning in conjunction with the resource availability constraints in the preset construction plan. It generates a set of control instructions that includes progress catch-up strategies and quality problem location and sends them to the construction management terminal to realize dynamic control of the construction progress and quality of fire protection engineering.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Load the building information model file of the fire protection project from the project database, parse the position coordinates and geometric parameters of each component in the model, and read the list of process nodes in the preset construction plan to generate the model-plan basic dataset; Laser scanners are deployed at the construction site to perform multi-station scanning of the constructed area according to the preset scanning stations, and the original point cloud data of each station is obtained. At the same time, panoramic cameras are used to collect 360-degree image data at each scanning station to generate original point cloud and image matching data. The original point cloud and image paired data are subjected to noise reduction filtering and stitching fusion to remove outliers and moving object point clouds, and the multi-site cloud is registered to a unified coordinate system to generate a fused full-site cloud of the construction area. Add a timestamp of the acquisition time to each point in the merged panoramic view cloud of the construction area, and attach the corresponding panoramic image as a texture to it, finally generating a time-stamped on-site status data stream.

8. The system according to claim 7, characterized in that, The extraction module is specifically used for: The overall point cloud in the site status data stream is coarsely registered with the design point cloud in the building information model. The initial transformation matrix is ​​estimated using the random sampling consistency algorithm, and then fine registration is performed using the iterative nearest point algorithm to generate the registered site point cloud model. The actual position coordinates and installation attitude angles of the fire protection components are extracted from the registered site point cloud model. The deviation is calculated for each component and compared with the design value of the corresponding component in the building information model to obtain the spatial deviation of each component and generate the component deviation feature vector. Based on the completion time of the process nodes in the preset construction plan, it is determined whether the currently completed components are completed on schedule. The delay time is quantified into a schedule deviation value. At the same time, the over-threshold items in the component deviation feature vector are marked as quality deviation values, and a schedule deviation vector and a quality deviation vector are generated. The schedule deviation vector and the quality deviation vector are aligned by component index and then fused into a unified deviation fusion matrix through principal component analysis. The rows of the matrix represent components and the columns represent deviation types and indicators.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images