Big data-based prefabricated building engineering project whole cycle management system

Abstract

The application discloses a prefabricated building engineering project whole cycle management system based on big data, and belongs to the technical field of prefabricated building management. The system comprises a big data collection layer, a data fusion layer, a state sensing layer and an analysis and decision layer. The big data collection layer collects multi-source heterogeneous data in the design, production, transportation, construction and operation and maintenance stages and standardizes the data; the data fusion layer is associated and mapped based on a component coding system, generates graph structure data and stores the data into a distributed graph database; the state sensing layer reads the graph structure in real time, detects attribute and relationship changes and dynamically updates a whole cycle state atlas; and the analysis and decision layer predicts component quality risks, progress deviations and cost overruns based on the atlas and outputs early warning information and management suggestions. The system realizes organic fusion and state dynamic tracking of whole cycle data, and improves the timeliness and accuracy of risk prediction.

Description

Technical Field

[0001] This invention relates to the field of prefabricated building management technology, specifically a big data-based full-cycle management system for prefabricated building engineering projects. Background Technology

[0002] In prefabricated building project management, data management in the design, production, transportation, construction, and operation and maintenance phases typically employs independent software systems. Data generated from BIM models, factory manufacturing execution systems, vehicle GPS and RFID, construction sensors, and building automation systems are fragmented in format and semantics, making it impossible to establish a full-cycle correlation at the individual component level. Current technologies generally simply aggregate multi-source data into a centralized data warehouse, failing to establish semantic connections between components regarding assembly relationships, batch relationships, transportation paths, and construction logic, and lacking a unified data model covering the entire process. When schedule delays or quality issues occur, it is difficult to trace the root causes across phases, and risk management is limited to static analysis of isolated phases.

[0003] Existing risk prediction methods rely on periodic reports or single-dimensional feature inputs, failing to utilize the dynamic evolution information of components across multiple stages, resulting in delayed predictions and insufficient accuracy. Prefabricated components are numerous, and the lack of a unified coding mapping mechanism between different systems makes manual alignment prone to data misalignment, hindering automated status monitoring. Furthermore, management tools lack the real-time ability to identify anomalies in component relationships, such as incorrect construction sequences or deviations from transportation routes, and cannot perform comprehensive risk inference based on the interconnected characteristics of all elements.

[0004] To achieve the organic integration of full-cycle data and dynamic perception of status in prefabricated building projects, the data at each stage needs to be woven into a graph structure with components as nodes and semantic relationships as edges according to the component coding system. This structure should support continuous monitoring of node attributes and edge relationships, dynamically update the full-cycle graph reflecting the overall project status, and then accurately predict quality risks, schedule deviations, and cost overruns based on the multi-dimensional features in the graph. Summary of the Invention

[0005] This paper presents a big data-based full-cycle management system for prefabricated building projects. This system integrates multi-source heterogeneous data from the design, production, transportation, construction, and operation and maintenance phases and constructs a graph structure to dynamically perceive changes in project status and predict and warn of component quality risks, schedule deviations, and cost overruns.

[0006] To achieve the above objectives, this invention provides the following technical solution: This invention provides a big data-based full-cycle management system for prefabricated building projects, aiming to solve the problems of data fragmentation and information silos in the design, production, transportation, construction, and operation and maintenance stages of prefabricated buildings, as well as the resulting quality risks, schedule deviations, and cost overruns. This system achieves dynamic monitoring and intelligent early warning of the status at each stage by constructing a data collection, fusion, perception, analysis, and decision-making chain that runs through the entire project lifecycle.

[0007] To achieve the above objectives, the system of this invention mainly includes a big data acquisition layer, a data fusion layer, a state awareness layer, and an analysis and decision-making layer. The big data acquisition layer collects multi-source heterogeneous data in real time from the design, production, transportation, construction, and operation and maintenance phases, and sends the standardized data to the data fusion layer. The data fusion layer receives the standardized data, performs association mapping based on a preset prefabricated building component coding system, generates unified graph structure data, and stores it in a distributed graph database. The state awareness layer reads the graph structure data from the distributed graph database in real time and dynamically updates the full-cycle state map of the prefabricated building project based on changes in the attributes and relationships of component nodes. The analysis and decision-making layer, based on the full-cycle state map, predicts component quality risks, schedule deviations, and cost overruns, and outputs early warning information and management suggestions.

[0008] As a preferred technical solution of this invention, the big data acquisition layer specifically includes data acquisition units for different stages. The design stage data acquisition unit extracts component geometric parameters, material properties, connection node types, and design change records from BIM software to generate a structured data stream for the design stage. The production stage data acquisition unit obtains component production numbers, pouring times, curing temperatures, rebar binding status, and quality inspection results in real time from the factory's MES system to generate time-series data for the production stage. The transportation stage data acquisition unit reads component transportation trajectories, real-time locations, vibration accelerations, and ambient humidity from vehicle-mounted GPS modules and RFID tags to generate spatial trajectory data for the transportation stage. The construction stage data acquisition unit obtains hoisting times, grouting pressures, installation sequences, and acceptance records from on-site tower crane sensors, concrete pump truck flow meters, and construction personnel mobile terminals to generate event stream data for the construction stage. The operation and maintenance stage data acquisition unit periodically collects component deformation displacement, crack width, temperature stress, and energy consumption data from the building automation system to generate periodic sampling data for the operation and maintenance stage. Through these units, the system achieves comprehensive coverage of multi-source data throughout the entire project lifecycle, laying a data foundation for subsequent fusion analysis.

[0009] Furthermore, the big data acquisition layer is equipped with a standardized processing unit. After receiving heterogeneous data from multiple sources, this unit uniformly names fields with the same meaning from different data sources according to a pre-defined data dictionary, converts date and time fields to UTC format, converts numeric fields to floating-point numbers, and performs word segmentation and stop word removal on text fields, ultimately generating standardized JSON structured data. This processing method eliminates format and semantic differences between data sources, ensuring a high degree of consistency in the data entering the data fusion layer.

[0010] In the data fusion layer of this invention, the unified coding mapping unit matches each data field from the design, production, transportation, construction, and operation and maintenance stages with a unique code in the pre-defined prefabricated building component coding system, establishing a field-to-code mapping table. Based on this mapping table, the graph structure generation unit abstracts each component as a graph node, uses assembly relationships, batch associations, transportation path links, and construction logic between components as graph edges, and uses the values ​​of data fields from each stage as node and edge attributes, generating graph structure data based on the Neo4j graph database interface. Thus, previously isolated business data is integrated into a knowledge graph capable of expressing the entire lifecycle of a component. The distributed storage management unit then stores this graph structure data in shards according to the hash values ​​of the component codes across multiple graph database nodes, maintaining a consistent index between nodes to support efficient access to massive amounts of data.

[0011] Preferably, the graph structure generation unit further includes an edge weight calculation subunit, used to calculate a weight value between 0 and 1 for each graph edge based on the physical connection strength level, distance between components, and construction logic tightness. The physical connection strength level is extracted from the BIM model, the distance is calculated from the component installation coordinates, and the construction logic tightness is obtained from the preceding relationship table in the construction organization plan. This weight value provides a quantitative basis for subsequent relationship change detection and critical path identification.

[0012] Furthermore, the production stage data acquisition unit is also equipped with a production cycle synchronization subunit. This subunit subscribes to production event messages for each component published by the factory's MES system via the MQTT protocol, parses the component code, timestamp, and process status contained in the messages, and groups the messages according to the production line number and mold number to generate a time-sorted production cycle sequence. This mechanism can accurately reflect the production rhythm of components and provide data support for refined prediction of schedule deviations.

[0013] The state awareness layer enables real-time insight into the project's state. It comprises a real-time graph structure reading unit, an attribute change detection unit, a relationship change detection unit, and a graph update unit. The real-time graph structure reading unit continuously scans the distributed graph database using a graph traversal engine, acquiring data update events for all component nodes within the most recent preset time window. The attribute change detection unit compares the current attribute value of each component node with its attribute value at the previous timestamp, calculates the attribute difference vector, and marks nodes with difference vectors exceeding a preset threshold as state change nodes. The relationship change detection unit traverses the adjacent edges of all state change nodes, identifying relationship anomalies such as broken assembly relationships, incorrect construction sequences, or deviations from transportation paths by comparing the changes in edge attributes within the time window, and marks the corresponding abnormal relationship edges as state change edges. The graph update unit is responsible for writing the attribute values ​​and timestamps corresponding to the state change nodes and edges into the full-cycle state graph and synchronously updating the version numbers of the corresponding nodes and edges in the graph database. This mechanism ensures that the full-cycle state graph can accurately and in real-time reflect the actual dynamics of the project site.

[0014] Based on the aforementioned full-cycle state map, the analysis and decision-making layer performs three core prediction tasks. The component quality risk prediction unit extracts production parameter vectors, transportation environment vectors, and construction operation vectors for each component from the map. These three vectors are concatenated and input into a pre-trained gradient boosting regression tree model, outputting a quality risk score for the component. This model can capture the combined impact of the nonlinear interaction between the production environment, transportation disturbances, and construction operations on component quality. The schedule deviation prediction unit extracts the completion node time series of each process in the construction phase from the map. It uses the Prophet time series decomposition model to predict the completion time of the remaining processes and compares the predicted completion time with the planned time to obtain the schedule deviation value. The cost overrun prediction unit extracts actual cost data from the phases already occurred from the map. Combining the number of design changes, material price indices, and labor cost fluctuation factors, it uses a multivariate linear regression model to predict the total cost throughout the entire cycle and calculates the overrun ratio compared to the budgeted cost.

[0015] Preferably, the schedule deviation prediction unit also includes a critical path identification subunit. This subunit extracts the dependencies between processes from the completion node time series of each process in the construction phase, constructs a directed acyclic graph, and uses a topological sorting algorithm to mark the processes on the critical path. Then, it performs Prophet model prediction only on the remaining processes on the critical path. This process focuses limited prediction computing power on processes that have a decisive impact on the overall project duration, effectively improving the targeting and efficiency of schedule early warning.

[0016] The early warning output unit is triggered when the quality risk score exceeds the first preset threshold, the schedule deviation exceeds the second preset threshold, or the cost overrun exceeds the third preset threshold. It generates information including the risk type, affected component number, and recommended corrective actions, and pushes this information to the project management terminal via a message queue. This mechanism enables project managers to be aware of potential problems and take intervention measures immediately.

[0017] Furthermore, as a complete technical solution of this invention, the system also includes a visualization interaction layer. This layer receives early warning information and management suggestions from the analysis and decision-making layer via the WebSocket protocol, and renders the full-cycle status map in the browser in the form of a three-dimensional force-directed graph. This three-dimensional force-directed graph allows users to filter and zoom by component code, time range, and status type, and also allows users to intuitively view the complete historical data traceability path of the component from design, production, transportation, construction to operation and maintenance by clicking on nodes in the graph. This interactive method transforms complex multi-dimensional data relationships into intuitive visualization graphics, significantly reducing the threshold for data understanding and decision-making.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows: The data fusion layer uses a unified coding mapping unit to match pre-defined unique component codes to data fields at each stage. The graph structure generation unit abstracts each component into a graph node, extracting assembly relationships, batch associations, transportation path links, and construction logic as graph edges, and using design parameters, production records, transportation environment values, construction operation details, and operation and maintenance monitoring data as node and edge attributes. It then calls the Neo4j graph database interface to generate graph structure data. The distributed storage management unit deploys graph data in shards across multiple nodes based on component code hash values, maintaining consistent indexing. This fusion method weaves fragmented information scattered across various independent systems into a component-centric, full-cycle semantic network. Complete data for any component can be retrieved along graph edges at millisecond levels. When an anomaly occurs in a certain stage, the affected area can be quickly defined along assembly and batch relationship edges, identifying the relevant components to be investigated without repeated cross-system queries. For large projects with tens of thousands of components, the sharded graph storage structure can support low-latency traversal and analysis under high concurrency, avoiding the performance bottleneck of centralized storage.

[0019] The state awareness layer continuously acquires graph structure update events within a preset time window using a graph traversal engine. The attribute change detection unit automatically marks nodes with state changes, such as deviations in curing temperature and excessive vibration acceleration, by calculating the attribute difference vector of each component node and comparing it with a threshold. The relationship change detection unit traverses the adjacent edges of the changed nodes, compares the evolution of edge attributes within the time window to identify relationship anomalies such as broken assembly relationships, incorrect construction sequence, or deviations in transportation paths, and marks the abnormal edges as state change edges. The graph update unit immediately writes the changed attributes and timestamps into the full-cycle state graph and updates the version number. The analysis and decision-making layer extracts component production parameter vectors, transportation environment vectors, and construction operation vectors from the graph, concatenates them, and inputs them into a pre-trained gradient boosting regression tree model to comprehensively capture the joint impact of cross-stage factors on quality and output a quality risk score. The schedule deviation prediction unit uses the Prophet model combined with critical path identification to perform time series prediction only on the remaining processes on the critical path, obtaining the deviation value from the planned time. The cost overrun prediction unit inputs the actual cost of the incurred stage, the number of design changes, material price index, and labor rate fluctuation factor into a multivariate linear regression model to calculate the overrun ratio of the total cost over the entire cycle to the budget. The early warning output unit generates an early warning message containing the risk type, component number, and corrective measures when the risk score, schedule deviation, or overrun ratio exceeds its respective threshold. The dynamic updating of the graph allows instantaneous changes in component status to be displayed in the project-level view and triggers a chain evaluation of related components, compressing the delay in anomaly detection to the minute level. Relationship anomaly detection can capture potential component mismatches caused by transportation path confusion, preventing rework after entering the construction stage. Multi-dimensional feature inputs enable quality risk scoring to reflect the cumulative effect throughout the entire process, from raw material curing to on-site installation, significantly enhancing its ability to identify hidden defects compared to methods relying solely on single-stage quality inspection data. Critical path progress prediction eliminates fluctuations in non-critical processes, making schedule estimates closer to actual progress. A multi-threshold hierarchical early warning mechanism allows managers to prioritize actions and allocate limited resources to the most pressing risk points. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a schematic diagram of the full-cycle management system for prefabricated building engineering projects based on big data. Figure 2 This is a flowchart of multi-source data fusion, graph structure generation, and distributed storage for prefabricated buildings; Figure 3This is a flowchart of the state perception layer dynamically updating the full-cycle state graph; Figure 4 This is a schematic diagram of the visual interaction layer structure. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figure 1 This invention provides a big data-based full-cycle management system for prefabricated building projects. The overall implementation scheme of the system is as follows: The system consists of a big data acquisition layer, a data fusion layer, a status awareness layer, and an analysis and decision-making layer. The big data acquisition layer collects multi-source heterogeneous data in real time from the design, production, transportation, construction, and operation and maintenance stages, standardizes the data, and sends it to the data fusion layer. The data fusion layer receives the standardized data, performs association mapping based on a preset prefabricated building component coding system, generates unified graph structure data, and stores it in a distributed graph database. The status awareness layer reads the graph structure data from the distributed graph database in real time and dynamically updates the full-cycle status map of the prefabricated building project based on changes in the attributes and relationships of component nodes. The analysis and decision-making layer, based on the full-cycle status map, predicts component quality risks, schedule deviations, and cost overruns, and outputs early warning information and management suggestions. Example 1

[0024] In practice, the big data acquisition layer collects multi-source heterogeneous data in real time from the design, production, transportation, construction and operation and maintenance stages, and then sends the multi-source heterogeneous data to the data fusion layer after standardization processing.

[0025] The design phase data acquisition unit interacts with the BIM model through the API interface provided by the BIM software. The unit calls the component geometry parameter extraction function to read the geometric parameters such as length, width, height, volume, and cross-sectional dimensions of each prefabricated building component from the BIM model; it calls the material property extraction function to read the material properties such as concrete strength grade, steel reinforcement type, and elastic modulus of each prefabricated building component; it calls the connection node type extraction function to read the connection node types between each prefabricated building component, including sleeve grouting connection, grout-anchored lap connection, and bolted connection; and it calls the design change record extraction function to read the design change records, including design change number, change time, change content, and change reason. The design phase data acquisition unit organizes the extracted component geometry parameters, material properties, connection node types, and design change records according to timestamps to generate a structured data stream for the design phase. Each record in the structured data stream includes a component code, data field name, data field value, and acquisition timestamp.

[0026] The production phase data acquisition unit collects data through the OPC UA server interface provided by the factory MES system. The production phase data acquisition unit sends subscription requests to the OPC UA server, subscribing to the component production number node, pouring time node, curing temperature node, rebar binding status node, and quality inspection result node. When a node value changes, the OPC UA server pushes a data packet containing the component code, timestamp, and node value to the production phase data acquisition unit. The production phase data acquisition unit groups the received data packets according to the component code, and arranges the pouring time, curing temperature, rebar binding status, and quality inspection result corresponding to each component code in chronological order, generating production phase time-series data. The rebar binding status node has a binary value range, where 0 indicates binding is not complete and 1 indicates binding is complete; the quality inspection result node includes three states: qualified, unqualified, and pending re-inspection. The production cycle synchronization subunit in the production phase data acquisition unit subscribes to production event messages for each component published by the factory MES system via the MQTT protocol. The production cycle synchronization subunit establishes a connection with the MQTT Broker of the factory's MES system and subscribes to messages on the topic " / production / events / #". Upon receiving a production event message, the subunit parses the component code field, timestamp field, and process status field in the message body. The process status field's values ​​include mold preparation, rebar placement, concrete pouring, vibration, curing, demolding, and finished product inspection. The subunit groups the production event messages based on the production line number and mold number fields, arranging messages under the same production line number and mold number in ascending order of timestamp, generating a time-sorted production cycle sequence. Each record in the production cycle sequence contains a component code, process status, start timestamp, end timestamp, and production line number.

[0027] The data acquisition unit during the transportation phase communicates with the vehicle-mounted GPS module and RFID tags. The unit acquires the vehicle's latitude and longitude coordinates and speed once per minute via the GPS module installed on the transport vehicle, and generates the component transportation trajectory using a map matching algorithm. The unit also reads RFID tags attached to prefabricated building components using an RFID reader installed on the transport vehicle, obtaining the component code and read timestamp. Each time an RFID tag is read, the unit records its real-time location, determined by the latitude and longitude coordinates returned by the GPS module. Furthermore, the unit collects vibration acceleration data using a triaxial vibration sensor installed in the transport vehicle's cargo compartment at a sampling frequency of 100Hz, calculating the root mean square value of vibration acceleration every 5 minutes; and collects ambient humidity data using a humidity sensor at a sampling frequency of once per minute. The unit aligns the component transportation trajectory, real-time location, root mean square value of vibration acceleration, and ambient humidity data according to time to generate spatial trajectory data for the transportation phase. Each record in this spatial trajectory data includes the component code, timestamp, longitude, latitude, speed, root mean square value of vibration acceleration, and ambient humidity.

[0028] The construction phase data acquisition unit connects to on-site tower crane sensors, concrete pump truck flow meters, and construction personnel's mobile terminals via a wireless network. The unit acquires the lifting start time, lifting end time, and RFID identification information of the lifted components from the tower crane sensors, merging the start and end times into a single lifting time record and associating it with the component code. It also acquires the grouting start time, grouting end time, and grouting flow rate from the concrete pump truck flow meter, calculating the total grouting pressure and average grouting pressure to generate grouting pressure data. Furthermore, the unit obtains installation sequence records and acceptance records from the construction personnel's mobile terminals. The installation sequence record is automatically generated and sorted by the timestamps generated after the personnel scan the component codes using their mobile terminals. The acceptance record includes acceptance items, acceptance results, and acceptance personnel information. Finally, the unit organizes the lifting time, grouting pressure, installation sequence, and acceptance records according to the chronological order of events, generating construction phase event stream data. Each event in the construction phase event stream data includes the event type, component code, event timestamp, specific event attributes, and the event source device number.

[0029] The operation and maintenance phase data acquisition unit performs regular data collection through the BACnet / IP interface provided by the building automation system. At the start of each sampling period, the operation and maintenance phase data acquisition unit sends a read request to read the data from displacement sensors, crack meters, temperature sensors, and energy meters installed on prefabricated building components. The displacement sensor returns the component deformation displacement value, the crack meter returns the crack width value, the temperature sensor returns the temperature stress value, and the energy meter returns the cumulative energy consumption value. The sampling period is set to once every 24 hours. The operation and maintenance phase data acquisition unit marks the collected component deformation displacement, crack width, temperature stress, and energy consumption data with the sampling timestamp, to generate operation and maintenance phase periodic sampling data. Each record in the operation and maintenance phase periodic sampling data includes the component code, sampling timestamp, component deformation displacement, crack width, temperature stress, and energy consumption.

[0030] The standardization processing unit receives the above structured data stream from the design phase, time series data from the production phase, spatial trajectory data from the transportation phase, event stream data from the construction phase, and periodic sampling data from the operation and maintenance phase. The standardization processing unit reads a preset data dictionary, which stores the mapping relationship between field names extracted from different data sources and standard field names, for example, mapping both "length" and "component length" to "component length", and mapping both "pouring temperature" and "curing temperature" to "curing temperature". The standardization processing unit traverses each input data record, and replaces the data field name with the standard field name according to the data dictionary. After that, the standardization processing unit identifies the date and time field in the data record, extracts the original string of the date and time field, parses it according to the format of the original string, and uniformly converts the date and time field into UTC format, with the converted format being "YYYY-MM-DDTHH:MM:SS.sssZ". The standardization processing unit identifies the numeric fields in the data record, checks the data type of the numeric fields, and if they are of integer or string type, forcibly converts them to double-precision floating-point type, retaining six decimal places. The standardization processing unit identifies the text fields in the data record, and performs word segmentation operation and stop word removal operation on the text fields; the word segmentation operation adopts the forward maximum matching algorithm to split the text string into a word sequence; the stop word removal operation deletes stop words appearing in the word sequence according to a pre-constructed stop word list, which contains words with no actual engineering meaning such as "de", "shi", "zai". The standardization processing unit assembles the data records after unified field names, unified date and time formats, unified numeric types and text preprocessing into JSON structure data. The top layer of the JSON structure data includes a component code key, a data source key, an acquisition timestamp key and an attribute key. The value corresponding to the attribute key is a nested JSON object, and each key-value pair in the nested JSON object is a standard field name and the corresponding processed data value. The standardization processing unit sends the generated standardized JSON structure data to the data fusion layer. Example 2

[0031] In specific implementation, please refer to Figure 2 The data fusion layer receives standardized JSON structured data, performs association mapping based on a preset prefabricated building component coding system, generates unified graph structure data, and stores it in a distributed graph database.

[0032] The unified coding mapping unit preloads a preset prefabricated building component coding system. This system employs a hierarchical coding structure, comprising project, building, floor, component type, and serial number segments. The project segment is 4 characters, the building segment is 3 characters, the floor segment is 3 characters, the component type segment is 4 characters, and the serial number segment is 5 characters, separated by underscores. The unified coding mapping unit parses the standardized JSON structure data received from the big data acquisition layer, extracting the name of each data field contained in the attribute keys. It then iterates through each data field name, searching for a unique code corresponding to that field name in the prefabricated prefabricated building component coding system's field mapping configuration table. This table predefines the mapping relationships between design-stage, production-stage, transportation-stage, construction-stage, and operation-maintenance-stage data field names and their corresponding standard field identifiers in the component coding system. The unified coding mapping unit stores the correspondence between data field names and unique codes in a field-to-code mapping table. The field-to-code mapping table is stored in key-value pairs, with the key being "component code_data field name" and the value being the standard field identifier corresponding to the unique code.

[0033] The graph structure generation unit generates graph structure data based on the field-to-code mapping table. It iterates through each standardized JSON structure data entry, extracts the component code value corresponding to the component code key, uses the component code value as the unique identifier of the graph node, and calls the node creation interface provided by the Neo4j graph database to create a graph node labeled "Component". The unit extracts all standard field names and corresponding data values ​​from the JSON structure data attribute keys, uses the standard field names as node attribute names, and the data values ​​as node attribute values, and calls the Neo4j graph database interface to write the node attributes into the graph node. The unit creates graph edges based on the assembly relationships between components. It obtains assembly relationship data between components from the BIM model, including the primary component code, secondary component code, and connection type. The unit calls the Neo4j graph database relationship creation interface to create a directed edge between the graph node corresponding to the primary component code and the graph node corresponding to the secondary component code, labeled "ASSEMBLY", with the edge attribute including the connection type. The graph structure generation unit obtains batch association data from the production stage data. This batch association data includes a list of all component codes within the same production batch. The unit creates an undirected edge between any two component codes corresponding to graph nodes within the same production batch, labeled "BATCH". The unit also obtains transportation path links from the transportation stage spatial trajectory data. These links consist of the sequence of component codes loaded by the same transport vehicle in a single transport task. The unit creates a directed edge between graph nodes corresponding to two component codes in adjacent loading sequences within the transportation path link, labeled "TRANSPORT". The edge attributes include the transportation trajectory number and transportation sequence number. Finally, the unit obtains construction logic data from the construction organization plan. This construction logic data includes preceding and subsequent component codes. The unit creates a directed edge between the graph node corresponding to the preceding component code and the graph node corresponding to the subsequent component code, labeled "LOGIC". The edge attributes include the construction process name and the planned interval in days. The graph structure generation unit writes the data field values ​​of each stage corresponding to the above assembly relationship edge, batch association edge, transportation path link edge, and construction logic edge as edge attributes into each graph edge.

[0034] The edge weight calculation sub-unit in the graph structure generation unit calculates a weight value between 0 and 1 for each graph edge. This sub-unit extracts the physical connection strength level between components from the BIM model. The physical connection strength level is divided into five levels according to the importance of the connection node in the structural load-bearing system, represented by 1, 2, 3, 4, and 5, with higher values ​​indicating higher connection strength. The edge weight calculation sub-unit normalizes the physical connection strength level by dividing the value by 5 to obtain the normalized physical connection strength level. , The value range is [0.2, 1.0]. The edge weight calculation sub-unit extracts the installation coordinates of each component from the BIM model. The installation coordinates are point coordinates in three-dimensional space. Calculate the Euclidean distance between the installation coordinates of the two components. Euclidean distance The calculation formula is ,in , , For the installation coordinate components of the first component, , , The installation coordinate components for the second component are used. The edge weight calculation sub-element finds the maximum value among all component pairs' distances as... The edge weight calculation subunit extracts the construction logic tightness from the prerequisite relationship table of the construction organization plan. The construction logic tightness is divided into three levels according to the degree to which two components must be constructed consecutively in the construction process: tight, moderate, and loose, represented by values ​​2, 1, and 0 respectively. The edge weight calculation subunit normalizes the construction logic tightness by dividing the value by 2, obtaining the normalized construction logic tightness. , The value range is {0, 0.5, 1.0}. The edge weight calculation sub-unit calculates the edge weight according to the following formula. :

[0035] in, This is the weighting coefficient for the physical connection strength level, with a value of 0.4; This is a weighting coefficient based on distance, with a value of 0.3. This is the weighting coefficient for the tightness of the construction logic, with a value of 0.3. The value of 0.4 is based on the fact that the physical connection strength level directly reflects the reliability of the structural force transmission path between components, and has the most significant impact on the overall safety of prefabricated buildings, so it should account for the largest proportion in the edge weight. The value of 0.3 is based on the fact that the distance affects the order of component hoisting and on-site logistics efficiency, and has a certain impact on the construction progress, but its importance is secondary. The value of 0.3 is based on the fact that the tightness of the construction logic determines the mandatory requirements for overlapping work processes, which also has a substantial impact on schedule control, consistent with the weighting of distance. The edge weight calculation sub-unit will calculate the edge weights. Write it as an edge attribute to the corresponding graph edge, with the edge attribute name being "weight".

[0036] The distributed storage management unit stores the graph structure data generated by the graph structure generation unit into the distributed graph database cluster. The distributed storage management unit extracts the component code value of each graph node, calculates the MD5 hash value of the component code value, converts the first 8 hexadecimal characters of the MD5 hash value into a 32-bit integer, and takes the modulo of this 32-bit integer. The modulo is the number of graph database nodes in the distributed graph database cluster, resulting in a sharding index. The distributed storage management unit stores the graph node and all graph edges connected to it as a single storage unit in the graph database node corresponding to the sharding index. During storage, the distributed storage management unit maintains a consistency index between nodes. The consistency index is a lookup table of all component codes and sharding indexes, and it is replicated to every graph database node in the distributed graph database cluster. When the graph structure data is updated, the distributed storage management unit first updates the sharding graph database nodes containing the graph nodes and edges, and then synchronizes the updated consistency index information to all graph database nodes through a distributed consistency protocol. This ensures that when any graph database node receives a cross-sharding query request, it can locate the shard containing the target graph node through the consistency index, completing the cross-sharding graph traversal operation. Example 3

[0037] In specific implementation, please refer to Figure 3 The state awareness layer reads graph structure data from the distributed graph database in real time and dynamically updates the full-cycle state map of the prefabricated building project based on changes in the attributes and relationships of component nodes.

[0038] The real-time graph structure reading unit continuously scans the graph structure data in the distributed graph database using a graph traversal engine, obtaining data update events for all component nodes within the most recent preset time window. The real-time graph structure reading unit is configured with a preset time window, the length of which is set to 300 seconds. The graph traversal engine establishes a long connection with the distributed graph database cluster based on the Bolt protocol of the Neo4j graph database. The graph traversal engine sends Cypher query statements to the distributed graph database at a fixed interval of 30 seconds. The query statement is "MATCH(n:Component)WHERE n.updateTimestamp> `lastScanTime` represents the timestamp of the last scan. The graph traversal engine receives data update events from the distributed graph database for all component nodes that meet the conditions. Each data update event includes a node unique identifier field, a node attribute key-value pair set field, and a node update timestamp field. The real-time graph structure reading unit assembles all the acquired data update events into a data update event list and passes the data update event list to the attribute change detection unit.

[0039] The attribute change detection unit internally maintains a node attribute history snapshot table. This table is stored in memory using a hash table structure. The key of the hash table is the node's unique identifier, and the values ​​are the node attribute vector and its corresponding timestamp. After receiving a list of data update events, the attribute change detection unit iterates through each event. It extracts the node's unique identifier from each event, sorts the node attribute key-value pairs lexicographically by attribute name, and sequentially extracts all attribute values ​​to form the current attribute vector. The unit then searches the node attribute history snapshot table using the node's unique identifier as the key. If a corresponding historical record is found, the attribute vector for the previous timestamp is extracted from the record. If no record is found, the current attribute vector and timestamp are directly stored in the node attribute history snapshot table, ending the processing of that data update event. After obtaining the current attribute vector and the attribute vector for the previous timestamp, the attribute change detection unit calculates the attribute difference vector. The attribute difference vector contains the first attribute vector and the second attribute vector. The calculation method for the nth component is as follows: The nth component of the current attribute vector is... The value of the component minus the value of the previous timestamp attribute vector. The absolute value of the difference between the nth component value is used as the nth component value of the attribute difference vector. There are 1 component, of which This represents the index of the component in the attribute vector. The increments from 1 to the total number of dimensions of the attribute vector. The attribute change detection unit calculates the magnitude of the attribute difference vector, which is represented as... The calculation formula is:

[0040] in, This represents the total number of dimensions of the attribute difference vector, which is the total number of attributes of the component node. The first attribute difference vector represents the first attribute difference vector. Each component value; This represents the magnitude of the attribute difference vector calculated from the current attribute vector and the attribute vector at the previous timestamp.

[0041] The attribute change detection unit measures the magnitude of the attribute difference vector. Compared with a preset threshold, which is set to 0.15, the threshold is compared. The basis for setting the preset threshold to 0.15 is that all component node attribute values ​​have been normalized to the [0,1] range by the standardization processing unit of the big data acquisition layer before entering the attribute change detection unit. Statistical analysis of historical component state change records shows that over 97% of changes in modulus values ​​below 0.15 originate from sensor measurement noise or minor fluctuations in environmental parameters, and do not reflect substantial changes in the physical or engineering state of the component. When the modulus value is greater than or equal to 0.15, at least one dimension of the attribute value has shifted by more than 15%, corresponding to changes in the component's geometric dimensions, strength indicators, or construction parameters that require attention. When the value exceeds a preset threshold of 0.15, the attribute change detection unit marks the component node corresponding to the node's unique identifier as a state change node, and updates the current attribute vector and current timestamp of the state change node to the storage location corresponding to the node's unique identifier in the node attribute history snapshot table. If the threshold of 0.15 is not exceeded, the attribute change detection unit only updates the timestamp corresponding to the unique identifier of the node in the node attribute history snapshot table to the current timestamp, and does not mark the component node as a state change node. The attribute change detection unit summarizes the unique identifiers of all marked state change nodes into a state change node list and sends the state change node list to the relationship change detection unit.

[0042] The relation change detection unit receives a list of nodes with state changes and performs an adjacency edge traversal operation on the unique identifier of each node in the list. The relation change detection unit executes a Cypher query through the Bolt interface of the Neo4j graph database, with the query template being "MATCH(n:Component{id:' The `nodeId` is replaced with the actual value of the unique identifier of the state-changed node. The distributed graph database returns edge data for all adjacent edges connected to the state-changed node. Each adjacent edge's edge data includes an edge label field, an edge attribute key-value pair set field, and an edge update timestamp field. The relationship change detection unit identifies relationship anomalies for each returned adjacent edge by comparing the changes in edge attributes within a preset time window. The relationship change detection unit reads the edge label of the adjacent edge. When the edge label is "ASSEMBLY", the relationship change detection unit extracts the connection type attribute value from the edge attributes and compares the connection type attribute value within the current preset time window with the connection type attribute value within the previous preset time window. If the connection type attribute value changes from a specific connection type string representing a valid connection to an empty string or the "DISCONNECTED" marker, the relationship change detection unit determines that the adjacent edge has experienced an assembly relationship breakage anomaly and marks the adjacent edge as a state-changed edge. When the edge label is "LOGIC", the relationship change detection unit extracts the construction process name attribute value from the edge attributes and compares it with the process name field specified in the construction organization plan. Simultaneously, it extracts the actual completion timestamp from the edge attributes and calculates the difference between the actual time interval and the planned interval in days. If the construction process name attribute value is inconsistent with the process name specified in the construction organization plan, or if the difference in actual time interval exceeds the preset allowable deviation range, the relationship change detection unit determines that the adjacent edge has an abnormal construction sequence and marks it as a state change edge. The preset allowable deviation range is set to ±20% of the planned interval in days based on the floating time difference in the construction organization plan. When the edge label is "TRANSPORT", the relationship change detection unit extracts the transportation sequence number attribute value from the edge attributes and compares it with the pre-planned transportation path sequence table in the transportation stage spatial trajectory data. If the absolute value of the difference between the transportation sequence number attribute value and the corresponding position number in the preset path sequence table is greater than 2, the relationship change detection unit determines that the adjacent edge has an abnormal transportation path deviation and marks it as a state change edge. The relation change detection unit summarizes all marked state change edges into a state change edge list, and then transmits the state change node list and the state change edge list together to the graph update unit.

[0043] The graph update unit internally maintains a full-cycle state graph, stored in an independent graph database namespace named "LifecycleGraph". After receiving the list of state change nodes and the list of state change edges, the graph update unit processes each state change node in the list. The unit queries the full-cycle state graph for the graph node corresponding to its unique identifier. If a matching node is found, it writes the attribute name and value from the current attribute vector of the state change node into the node's attributes, writes the end timestamp of the current preset time window into the state change time attribute, and increments the version number attribute by 1. If no matching node is found, the unit creates a new graph node in "LifecycleGraph", using the node's unique identifier as its unique identifier. It writes the current attribute vector of the state change node into the new node's attributes, writes the end timestamp of the current preset time window into the new node's state change time attribute, and sets the version number attribute to 1. The graph update unit processes each state change edge in the state change edge list. It searches the full-cycle state graph for edges whose start and end points are unique identifiers of the two endpoints of the state change edge, and the edge labels are the same as the edge labels of the state change edge. If a corresponding edge is found, the edge attribute set of that edge is overwritten and updated with the edge attribute set of the state change edge, the state change time attribute is updated to the end timestamp of the current preset time window, and the edge version number attribute value is incremented by 1. If no corresponding edge is found, the graph update unit creates a new edge between the start and end nodes in the "LifecycleGraph", sets the edge label to the edge label of the state change edge, writes it into the edge attribute set of the state change edge, sets the state change time attribute to the end timestamp of the current preset time window, and sets the edge version number attribute to 1. After completing the write operation of the full-cycle state graph, the graph update unit synchronously writes back the version number attribute values ​​of all updated graph nodes and edges in the full-cycle state graph to the version number attribute fields of the corresponding component nodes and adjacent edges in the distributed graph database. The synchronous write-back operation is completed through the transaction batch update mechanism of Neo4j graph database, with a batch size of 100 records. Example 4

[0044] In practice, the analysis and decision-making layer uses the full-cycle state map generated by the state perception layer to predict component quality risks, schedule deviations and cost overruns, and outputs early warning information and management suggestions.

[0045] The component quality risk prediction unit extracts the production parameter vector, transportation environment vector, and construction operation vector for each component from the full-cycle state map. Using the component code as the query condition, the unit locates the corresponding graph node in the full-cycle state map and reads the attribute values ​​related to the production stage from the node's attributes. The production parameter vector includes five dimensions: pouring time interval, mean curing temperature, standard deviation of curing temperature, rebar binding status identifier, and quality inspection result identifier. The pouring time interval dimension is obtained by extracting the difference between the start timestamp of the mold preparation process and the end timestamp of the finished product inspection process from the time series data of the production stage. The mean curing temperature dimension extracts the arithmetic mean of all curing temperature samples during the curing period. The standard deviation of curing temperature dimension extracts the standard deviation of all curing temperature samples during the curing period. The rebar binding status identifier dimension takes a value of 0 or 1. The quality inspection result identifier dimension takes a value of 0, 1, or 2, corresponding to unqualified, qualified, and pending re-inspection, respectively. The transportation environment vector includes three dimensions: mean root mean of vibration acceleration, maximum root mean of vibration acceleration, and mean ambient humidity. The construction operation vector includes four dimensions: hoisting time deviation, mean grouting pressure, standard deviation of grouting pressure, and installation sequence deviation. The hoisting time deviation dimension is obtained by subtracting the planned hoisting time from the actual hoisting time of the component. The mean grouting pressure dimension is extracted as the arithmetic mean of all sampled grouting pressure values ​​during the grouting process. The standard deviation of grouting pressure dimension is extracted as the standard deviation of all sampled grouting pressure values ​​during the grouting process. The installation sequence deviation dimension is obtained by subtracting the absolute value of the installation sequence number specified in the construction organization plan from the actual installation sequence number of the component.

[0046] The component quality risk prediction unit concatenates the production parameter vector, transportation environment vector, and construction operation vector in the order of production parameter vector first, transportation environment vector in the middle, and construction operation vector last, forming a joint feature vector containing twelve dimensions. The component quality risk prediction unit loads a pre-trained gradient boosting regression tree model, implemented using the XGBoost framework. The model hyperparameters are set as follows: maximum tree depth of 6, learning rate of 0.05, number of weak learners of 200, subsampling ratio of 0.8, column sampling ratio of 0.8, regularization parameter lambda of 0.1, and regularization parameter alpha of 0.05. The training process of the pre-trained gradient boosting regression tree model is as follows: collect full-cycle data of components from historical prefabricated building projects, extract the production parameter vector, transportation environment vector, and construction operation vector for each historical component from the historical full-cycle data, and record the quality defect level detected for the first time during the operation and maintenance phase of each historical component as a training label. The quality defect level is divided into a continuous score from 0 to 100 according to the severity of the defect, where 0 represents no defect and 100 represents a severe defect. The 12-dimensional joint feature vector of historical components is used as input features, and the quality defect level is used as the target value. The dataset is divided into training and validation sets in an 8:2 ratio. The mean squared error loss function is minimized on the training set, and an early stopping strategy is employed: training terminates when the mean squared error on the validation set no longer decreases after 20 consecutive iterations. The component quality risk prediction unit inputs the 12-dimensional joint feature vector of the current component into a pre-trained gradient boosting regression tree model. The pre-trained gradient boosting regression tree model outputs a quality risk score for the component, which is a continuous value between 0 and 100.

[0047] The schedule deviation prediction unit extracts the completion node time series of each process in the construction phase from the full-cycle state map. Using the construction process name as the node query condition, the unit searches the full-cycle state map for nodes tagged "construction process," and reads the planned start time, planned end time, actual start time, and actual end time of the process from the node's attributes. The unit calculates the difference between the actual end time and the planned end time of the process, marking processes with an empty difference or actual end times later than the current system time as incomplete processes, and marking processes with actual end times as completed processes. The actual end times of completed processes are then arranged in ascending order to form the completion node time series.

[0048] The critical path identification subunit in the schedule deviation prediction unit extracts the dependencies between processes from the completion node time series of each process in the construction phase and constructs a directed acyclic graph (DAG). The critical path identification subunit reads the process dependency table from the construction organization plan. Each record in the process dependency table contains a predecessor process name field and a successor process name field. The critical path identification subunit uses each process name as a node in the DAG, and the directed edges from predecessor process names to successor process names are the directed edges, with the weight of the edge being the planned duration days of the predecessor process. The critical path identification subunit uses a topological sorting algorithm to perform topological sorting on the directed acyclic graph constructed from process dependencies. The steps of the topological sorting algorithm are as follows: Calculate the in-degree value of each node in the directed acyclic graph, and add all nodes with an in-degree value of zero to the queue; Take a node from the head of the queue and add it to the topological sequence; Traverse all successor nodes of the node, decrement the in-degree value of each successor node by 1; If the in-degree value of a successor node is reduced to 0, add the successor node to the tail of the queue; Repeat the above steps until the queue is empty. If the length of the topological sequence is equal to the total number of nodes in the directed acyclic graph, then the topological sorting is complete. After completing the topology sorting, the critical path identification subunit calculates the earliest start time and latest start time for each node according to the topology sequence. The earliest start time is calculated recursively from the first node, where the earliest start time of the current node equals the maximum of the earliest finish times of all predecessor nodes, and the earliest finish time equals the earliest start time plus the planned duration in days. The latest start time is calculated recursively from the last node, where the latest finish time of the current node equals the minimum of the latest start times of all successor nodes, and the latest start time equals the latest finish time minus the planned duration in days. The critical path identification subunit compares the earliest start time and latest start time of each process node and marks the process nodes with the same earliest start time and latest start time as processes on the critical path.

[0049] The schedule deviation prediction unit only uses the Prophet time series decomposition model to predict the remaining activities on the critical path. The unit collects the actual completion timestamps of all completed activities on the critical path, converts these timestamps into days from the project start date, and assigns one time series record to each completed activity. The feature column of the time series record is an integer number representing the completed activity name, and the target value is the actual number of days completed. The unit initializes the Prophet time series decomposition model, setting its parameters as follows: growth is set to linear, seasonality_mode is set to additive, changepoint_prior_scale is set to 0.05, and seasonality_prior_scale is set to 10.0. The unit inputs the time series records of completed activities on the critical path into the Prophet time series decomposition model. The model decomposes the time series into three parts: a trend component, a periodic component, and a holiday component. The trend component is a piecewise linear function, the periodic component has a weekly cycle, and the holiday component is empty. The schedule deviation prediction unit calls the `predict` method of the Prophet time series decomposition model, setting future time points as the planned duration days of the remaining activities on the critical path. The `predict` method outputs the predicted completion days by adding the predicted trend and periodicity values ​​for the future time points. The schedule deviation prediction unit compares the predicted completion days with the planned duration days of the remaining activities to calculate the schedule deviation value. The schedule deviation value is the difference between the predicted completion days and the planned duration days, expressed in days.

[0050] The cost overrun prediction unit extracts actual cost data for the current stage from the full-cycle state map. It retrieves nodes labeled "cost record" from the full-cycle state map and reads the values ​​for three dimensions: actual cost in the design stage, actual cost in the production stage, and actual cost in the transportation stage from the node's attributes. The unit also extracts the number of design changes from the full-cycle state map, obtained by counting the total number of event nodes labeled "design change" in the full-cycle state map. Furthermore, the unit obtains the material price index and labor cost fluctuation factor for the current month through an external data interface. The material price index is based on the building materials price index published by the National Bureau of Statistics, and the labor cost fluctuation factor is calculated as the ratio of the average daily wage of construction workers in the project location to the average daily wage in the base period. Finally, the unit constructs the input feature vector for a multivariate linear regression model, which includes six feature dimensions: actual cost in the design stage, actual cost in the production stage, actual cost in the transportation stage, number of design changes, material price index, and labor cost fluctuation factor. The expression for the multivariate linear regression model is as follows:

[0051] in, This indicates the predicted total cost over the entire lifecycle; This indicates the actual cost during the design phase; This represents the actual cost during the production stage; This indicates the actual cost during the transportation phase; Indicates the number of design changes; This represents the material price index; This represents the labor cost rate fluctuation factor; The intercept term is the mean of the regression residuals of the total cost of 30 similar prefabricated building projects in history relative to the cost of each stage, which is set to 253.7 after statistical calculation. The regression coefficient for the actual cost during the design phase is 1.42. The regression coefficient for the actual cost during the production stage is 1.18. The regression coefficient for the actual cost during the transportation phase is 1.05. The regression coefficient for the number of design changes is 8.73. The regression coefficient for the material price index is 0.67. The regression coefficient for the labor cost rate fluctuation factor is 0.91. This regression coefficient is obtained using the least squares method, based on historical data of actual costs at each stage and total final settlement costs from 30 similar prefabricated building projects. The cost overrun prediction unit substitutes the values ​​of the six feature dimensions into the multivariate linear regression model expression to calculate the predicted total cost for the entire project lifecycle. The budget cost is read from the project management database, and the overspending ratio is calculated. The overspending ratio is the percentage obtained by subtracting the budget cost from the predicted total cost for the entire cycle, dividing by the budget cost, and then multiplying by 100%.

[0052] The early warning output unit internally maintains a first preset threshold, a second preset threshold, and a third preset threshold. The first preset threshold is set for the quality risk score, with a value of 70. This value is based on the fact that a quality risk score between 0 and 100 corresponds to a critical threshold requiring special rectification of quality defects. The second preset threshold is set for the schedule deviation value, with a value of 14 days. This value is based on the fact that construction contracts for prefabricated building projects typically stipulate that a deviation from a milestone date exceeding two weeks constitutes a delay. The third preset threshold is set for the cost overrun percentage, with a value of 5%. This value is based on the fact that project cost management regulations stipulate that a budget deviation exceeding 5% must be reported to the project manager for record-keeping and the cost control process must be initiated. The early warning output unit receives the quality risk score output by the component quality risk prediction unit. When the quality risk score exceeds the first preset threshold of 70, a quality risk early warning message is generated. The early warning output unit also receives the schedule deviation value output by the schedule deviation prediction unit. When the schedule deviation value exceeds the second preset threshold of 14 days, a schedule deviation early warning message is generated. The early warning output unit receives the overspending ratio output by the cost overspending prediction unit. When the overspending ratio exceeds the third preset threshold of 5%, a cost overspending early warning message is generated. Each early warning message includes a risk type field, an affected component number field, and a recommended corrective action field. The recommended corrective action field queries the action knowledge base according to the risk type: the recommended corrective action for quality risk is "conduct on-site non-destructive testing of the component and initiate the component supplier quality backtracking process"; the recommended corrective action for schedule deviation is "add construction teams or adjust construction techniques to compress single-process operation time in subsequent processes of the critical path"; the recommended corrective action for cost overspending is "review the procurement plan for the non-critical stage and initiate cost optimization schemes, prioritizing the adjustment of resource input for non-critical paths". The early warning output unit serializes the generated early warning message into JSON format and pushes it to the project management terminal through a message queue using Apache Kafka, with the topic name "project-alert-topic". Example 5

[0053] In specific implementation, please refer to Figure 4 The visualization and interaction layer receives early warning information and management suggestions from the analysis and decision-making layer via WebSocket, and renders the full-cycle status map in the browser in the form of a three-dimensional force-guided graph. It supports filtering and zooming by component code, time range and status type, and allows users to view the complete historical data traceability path of the component by clicking on the node.

[0054] The visualization interaction layer includes a WebSocket client module, which is automatically initialized when the browser loads the visualization page. The WebSocket client module reads the pre-configured analysis and decision-making layer message queue subscription address, which is formatted as "ws: / / <analysis and decision-making layer server IP address>:<port number> / alert". The WebSocket client module initiates a WebSocket handshake request to the analysis and decision-making layer server, with the request header containing an "Upgrade" field with the value "websocket" and a "Connection" field with the value "Upgrade". After establishing a long-lived WebSocket connection with the analysis and decision-making layer server, the WebSocket client module subscribes to the warning information and management recommendations pushed by the warning output units in the analysis and decision-making layer. Simultaneously, the warning output units in the analysis and decision-making layer push the same warning information and management recommendations to the visualization interaction layer via the long-lived WebSocket connection. The pushed data is in JSON string format, containing fields for risk type, affected component number, recommended corrective action, and push timestamp. Upon receiving the JSON string, the WebSocket client module calls the JavaScript engine's `JSON.parse` method to parse the JSON string into a warning information object, which contains attributes for risk type, affected component number, recommended corrective action, and push timestamp. The WebSocket client module stores the warning information object into the warning information queue in the browser's memory. The warning information queue adopts a first-in-first-out array structure with a capacity limit of 100 items. When the capacity limit is exceeded, the first element of the queue is automatically removed.

[0055] The visualization interaction layer includes a 3D force-directed graph rendering module, which is implemented based on the Three.js graphics library. The 3D force-directed graph rendering module retrieves graph data from the "LifecycleGraph" namespace, the graph database containing the full-cycle state graph, via an HTTP GET request. The request parameters include the graph identifier and the return format, which is specified as JSON. Upon receiving the graph data, the 3D force-directed graph rendering module parses each graph node into a sphere object in 3D space. The center coordinates of the sphere object are initialized to randomly distributed coordinate values ​​in 3D space. The radius of the sphere object is scaled according to the degree centrality of the graph node. The degree centrality is the number of edges connected to the graph node in the full-cycle state graph. The radius of the sphere object is proportional to the degree centrality value, with a base radius of 5 Three.js scene units. For every increase of 1 in degree centrality, the radius increases by 0.5 Three.js scene units. The 3D force-directed graph rendering module creates a material object for each graph node. The material object uses the MeshPhongMaterial material, and the color is set according to the engineering stage to which the graph node belongs: light blue for design stage nodes, light green for production stage nodes, orange for transportation stage nodes, yellow for construction stage nodes, and light purple for maintenance stage nodes. The module parses each graph edge in the graph data into a cylindrical line in 3D space. The two ends of the cylindrical line connect to the center coordinates of the source graph node and the center coordinates of the target graph node, respectively. The thickness of the cylindrical line is scaled according to the edge weight attribute value, which is read from the edge attributes of the full-cycle state graph. The thickness of the cylindrical line is proportional to the edge weight attribute value, with a base thickness of 0.5 Three.js scene units. For every 0.1 increase in the edge weight attribute value, the thickness increases by 0.2 Three.js scene units.

[0056] The 3D force-directed graph rendering module simulates force-directed layout in a Three.js scene. The force model of force-directed layout includes three types of forces: Coulomb repulsion between nodes, Hooke's attraction caused by edges, and central attraction. The Coulomb repulsion is calculated as follows: for any two graph nodes... and Graph nodes Received from graph nodes Coulomb repulsion vector The direction is from the graph node The center coordinates point to the graph node The center coordinates, The formula for calculating the size is:

[0057] in, The value is the Coulomb force constant, set to 500. The reason for setting the value to 500 is that the scale range of the three-dimensional spatial coordinate system is approximately 1000×1000×1000 Three.js scene units. A Coulomb force constant of 500 can generate a repulsive force of about 0.05 units between two nodes with a spacing of 100 scene units, which is sufficient to offset the excessive overlap of the initial random distribution without causing the nodes to scatter. and Graph nodes Graph Nodes The charge value is set to the degree centrality value of the graph node plus 1. For graph nodes Center coordinates and graph nodes The Euclidean distance between the center coordinates, The calculation formula is for graph nodes. The center coordinates of each component are reduced by the map nodes. The square root of the sum of the squares of the components of the central coordinate.

[0058] The calculation method for Hooke's gravity is as follows: for each source graph node connected by an edge... and target graph nodes Source graph nodes The Hooker gravitational vector from the edge The direction is from the source graph node The center coordinates point to the target graph node The center coordinates, The formula for calculating the size is: ,in The elasticity coefficient is set to 0.001. Source graph node With target graph nodes The Euclidean distance between them Let be the natural length of the edge. The calculation is dynamically calculated based on the edge weight attribute value, and the formula is as follows: , This represents the edge weight attribute value.

[0059] The central gravity is calculated as follows: for each graph node in the scene, the central gravity vector is... The direction is from the center coordinates of the graph node to the origin of the three-dimensional coordinate system. The formula for calculating the size is: , where is the central gravitational coefficient, with a value set to 0.0001. is the Euclidean distance from the center coordinates of the graph node to the origin of the 3D spatial coordinate system. In each iteration of the force-directed layout, the 3D force-directed graph rendering module accumulates the Coulomb repulsive force vector, Hooke's gravitational force vector, and central gravitational force vector acting on each graph node to obtain the resultant force vector. This resultant force vector is multiplied by a time step of 0.01 and then superimposed onto the graph node's velocity vector. The velocity vector is then multiplied by a damping coefficient of 0.95 for attenuation. Finally, the velocity vector is superimposed onto the center coordinates of the graph node to complete one position update. The 3D force-directed graph rendering module continues iterating until the average magnitude of the velocity vectors of all graph nodes is less than 0.001 Three.js scene units, at which point the force-directed layout calculation is complete.

[0060] The visual interaction layer includes a filtering and zooming control module, which presents three filtering controls on the browser page: a component code filtering input box, a time range filtering slider, and a status type filtering dropdown list. The component code filtering input box receives the component code string entered by the user. The filtering and zooming control module listens for the `oninput` event of the component code filtering input box in real time. When the event is triggered, the module iterates through all graph nodes in the 3D force-guided graph rendering module, performing exact and prefix matching between the node's unique identifier and the string in the component code filtering input box. Nodes that successfully match are set to `visible`, while nodes that fail to match are set to `hidden`. Simultaneously, the visibility of graph edges connected to nodes that fail to match is also set to `hidden`. The time range filtering slider allows the user to drag and set the start and end timestamps. The minimum value of the time range filtering slider is the project start timetamp, and the maximum value is the current project timestamp. The filtering and scaling control module monitors the `onchange` event of the time range filtering slider in real time, iterates through all graph nodes in the 3D force-guided graph rendering module, and reads the state change time attribute of each graph node. Graph nodes with state change time attributes between the start and end timestamps have their visibility attribute set to `visible`, while graph nodes outside the time range have their visibility attribute set to `hidden`. The visibility attribute of graph edges connected to graph nodes outside the time range is also synchronously set to `hidden`. The state type filtering drop-down list provides four options: "All," "Quality Risk," "Schedule Deviation," and "Cost Overrun." The filtering and scaling control module monitors the `onchange` event of the status type filter dropdown list in real time. When "All" is selected, the visibility attribute of all graph nodes is set to `visible`. When "Quality Risk" is selected, the filtering and scaling control module retrieves all warning information objects with the risk type attribute "Quality Risk" from the front-end warning information queue, extracts the affected component number attribute from the warning information object, sets the node color of all graph nodes corresponding to the affected component number attribute to red highlight, and keeps the color of the remaining graph nodes to their original stage color while reducing the transparency to 0.2. When "Schedule Deviation" is selected, the filtering and scaling control module retrieves all warning information objects with the risk type attribute "Schedule Deviation" from the front-end warning information queue, extracts the graph nodes corresponding to the affected component number attribute to red highlight, and reduces the transparency of the remaining graph nodes to 0.2. When "Cost Overrun" is selected, the same filtering and highlighting logic is executed.The filtering and zoom control module also binds to the mouse wheel event to realize the zoom control of the 3D scene. For each tick of the mouse wheel, the camera object of the Three.js scene moves forward or backward by 50 Three.js scene units along the line of sight. The zoom range of the camera object is limited to between 200 and 5000 Three.js scene units away from the center of the scene.

[0061] The visualization interaction layer includes a node click tracing module, which listens for user mouse click events on graph nodes in the 3D force-guided graph. The module uses Three.js's Raycaster ray detector to detect the intersection between the mouse click location and the graph node. When a mouse click event is detected hitting a graph node, the module retrieves the unique identifier of the clicked node. Using this unique identifier as a query parameter, the module initiates a Cypher query to the distributed graph database. The Cypher query statement is "MATCH path=(n:Component {id:'$nodeId'})-[*0..]-()RETURN path", which returns the subgraph data of all depth paths of that graph node in the graph structure data. The module then displays the returned subgraph data hierarchically according to path depth, rendering the tracing path information in a tree structure in the browser's sidebar. The root node of the tree structure is the clicked graph node. Below the root node, the first level of child nodes consists of the graph nodes directly adjacent to the clicked node and their corresponding edge labels. The second level of child nodes consists of the graph nodes directly adjacent to the first level of child nodes and their corresponding edge labels, and so on, until all subgraph data has been traversed. Each node in the tree structure displays its component code attribute value, node label type, and state change time attribute value. Each edge displays its edge label and edge weight attribute value. The node click tracing module allows users to click any node in the tree structure, setting that node as the new center node, initiating a new Cypher query to the distributed graph database, and refreshing the tracing path tree display.

[0062] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A big data-based full-cycle management system for prefabricated building projects, characterized in that: include: The big data acquisition layer is used to collect multi-source heterogeneous data in real time from the design, production, transportation, construction and operation and maintenance stages, and then send the data to the data fusion layer after standardization processing. The data fusion layer is used to receive the standardized data, perform association mapping based on the preset prefabricated building component coding system, generate unified graph structure data, and store it in a distributed graph database. The state awareness layer is used to read graph structure data from the distributed graph database in real time and dynamically update the full-cycle state map of the prefabricated building project based on the changes in the attributes and relationships of the component nodes. The analysis and decision-making layer is used to predict component quality risks, schedule deviations, and cost overruns based on the full-cycle status map, and output early warning information and management suggestions. The state-aware layer includes: The real-time graph structure reading unit is used to continuously scan the graph structure data in the distributed graph database through the graph traversal engine to obtain the data update events of all component nodes within the most recent preset time window; The attribute change detection unit is used to compare the current attribute value of each component node with the attribute value of the previous timestamp, calculate the attribute difference vector, and mark the node whose difference vector exceeds the preset threshold as a state change node. The relationship change detection unit is used to traverse all adjacent edges of the state change node, identify relationship anomalies such as assembly relationship breakage, construction sequence error or transportation path deviation by comparing the change of edge attributes within the time window, and mark the relationship anomaly edge as a state change edge. The graph update unit is used to write the attribute values ​​and timestamps corresponding to the state change nodes and state change edges into the full-cycle state graph, and synchronously update the version number of the node and edge in the graph database.

2. The big data-based full-cycle management system for prefabricated building projects according to claim 1, characterized in that, The big data acquisition layer includes: The design phase data acquisition unit is used to extract component geometric parameters, material properties, connection node types and design change records from BIM software to generate a structured data flow for the design phase. The production stage data acquisition unit is used to obtain component production number, pouring time, curing temperature, rebar binding status and quality inspection results from the factory MES system in real time, and generate production stage time series data; The transportation phase data acquisition unit is used to read the component transportation trajectory, real-time location, vibration acceleration and ambient humidity from the vehicle-mounted GPS module and RFID tag, and generate spatial trajectory data during the transportation phase. The construction phase data acquisition unit is used to acquire hoisting time, grouting pressure, installation sequence and acceptance records from on-site tower crane sensors, concrete pump truck flow meters and construction personnel mobile terminals, and generate construction phase event flow data; The operation and maintenance phase data acquisition unit is used to periodically collect component deformation displacement, crack width, temperature stress and energy consumption data from the building automation system to generate periodic sampling data for the operation and maintenance phase.

3. The big data-based full-cycle management system for prefabricated building projects according to claim 1, characterized in that, The data fusion layer includes: The unified coding mapping unit is used to match each data field in the design, production, transportation, construction and operation and maintenance stages with the unique code in the preset prefabricated building component coding system to establish a mapping table from field to code. The graph structure generation unit is used to take each component as a graph node according to the mapping table, take the assembly relationship, batch association, transportation path link and construction logic between components as graph edges, and take the values ​​of data fields of each stage as node attributes and edge attributes, and generate graph structure data based on the Neo4j graph database interface. The distributed storage management unit is used to divide the graph structure data into shards according to the hash values ​​of the component codes and store them to multiple graph database nodes, and maintain a consistent index between the nodes.

4. The big data-based full-cycle management system for prefabricated building projects according to claim 1, characterized in that, The analytical decision-making layer includes: The component quality risk prediction unit is used to extract the production parameter vector, transportation environment vector and construction operation vector of each component from the full-cycle state map, and input the three sets of vectors into the pre-trained gradient boosting regression tree model to output the quality risk score of the component. The schedule deviation prediction unit is used to extract the completion node time series of each process in the construction stage from the full-cycle state map, use the Prophet time series decomposition model to predict the completion time of the remaining processes, and compare the predicted completion time with the planned time to obtain the schedule deviation value. The cost overrun prediction unit is used to extract actual cost data of the stages that have occurred from the full-cycle status map, combine the number of design changes, material price index and labor cost fluctuation factor, predict the total cost of the whole cycle through a multivariate linear regression model, and calculate the overrun ratio with the budget cost. The early warning output unit is used to generate information including risk type, affected component number and recommended corrective measures when the quality risk score exceeds the first preset threshold, the schedule deviation value exceeds the second preset threshold or the overspending ratio exceeds the third preset threshold, and push it to the project management terminal through a message queue.

5. The big data-based full-cycle management system for prefabricated building projects according to claim 1, characterized in that, The big data acquisition layer also includes: The standardization processing unit is used to receive the multi-source heterogeneous data, and then, according to a preset data dictionary, uniformly name fields with the same meaning in different data sources, unify date and time fields to UTC format, unify numeric fields to floating-point type, perform word segmentation and stop word removal operations on text fields, and generate standardized JSON structure data.

6. The big data-based full-cycle management system for prefabricated building projects according to claim 2, characterized in that, The production stage data acquisition unit also includes: The production cycle synchronization subunit is used to subscribe to production event messages for each component published by the factory MES system via the MQTT protocol, parse the component code, timestamp, and process status contained in the message, group the message according to the production line number and mold number, and generate a production cycle sequence sorted by time.

7. The big data-based full-cycle management system for prefabricated building projects according to claim 3, characterized in that, The graph structure generation unit further includes: The edge weight calculation sub-unit is used to calculate a weight value between 0 and 1 for each edge of the drawing based on the physical connection strength level, distance between components, and construction logic tightness. The physical connection strength level is extracted from the BIM model, the distance is calculated from the component installation coordinates, and the construction logic tightness is obtained from the prerequisite relationship table in the construction organization plan.

8. The big data-based full-cycle management system for prefabricated building projects according to claim 4, characterized in that, The schedule deviation prediction unit further includes: The critical path identification subunit is used to extract the dependencies between processes from the time series of completion nodes of each process in the construction phase, construct a directed acyclic graph, use a topological sorting algorithm to mark the processes on the critical path, and only perform Prophet model prediction on the remaining processes on the critical path.

9. The big data-based full-cycle management system for prefabricated building projects according to claim 1, characterized in that, Also includes: The visualization and interaction layer is used to receive the early warning information and management suggestions from the analysis and decision-making layer via WebSocket, and render the full-cycle state map in the form of a three-dimensional force-guided graph in the browser. It supports filtering and zooming by component code, time range and state type, and allows users to view the complete historical data traceability path of the component by clicking on the node.

Images