A processing method and system based on engineering project data

Abstract

The application discloses a kind of based on engineering project data processing method and system, it is related to data processing technical field.The steps executed in sequence include the following: step one, based on uniform engineering coding construction engineering knowledge graph;Step two, construct graph signal enhanced engineering knowledge graph in engineering knowledge graph: collect the multi-source data in the implementation process of engineering project, obtain graph signal enhanced engineering knowledge graph;Step three, identify the structured abnormal area in graph;Step four, based on the structured abnormal area and graph neural network model identified, and according to the joint risk prediction result, corresponding node is divided into risk grade and output.The present application can quickly find potential problems in complex engineering system, improve risk identification accuracy and response efficiency, significantly enhance the intelligent level and decision scientificity of engineering project management.

Description

A method and system for processing engineering project data Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for processing engineering project data. Background Technology

[0002] With the widespread application of information and intelligent technologies in the engineering construction field, engineering project management is gradually shifting from traditional manual statistics and experience-based judgment to data-driven intelligent decision-making. Existing engineering project management systems generally possess functions such as schedule planning, cost control, and quality and safety information recording, enabling data sharing and basic information integration among multiple departments. However, engineering project data is characterized by multi-source heterogeneity, wide spatiotemporal distribution, inconsistent sampling periods, and significant differences in data accuracy, leading to considerable difficulties in data fusion, status identification, and risk prediction. Currently, the industry mainly conducts risk analysis through database association, key indicator analysis, or single machine learning models. While these methods can improve the automation level of engineering management to some extent, they still have several technical shortcomings.

[0003] Most existing schedule management and risk prediction technologies rely on single time series analysis methods or regression prediction models. For example, a common practice is to build a time series model based on project schedule data and use historical deviation data to predict future schedule risks. This type of method works well when there are few input features and the data structure is relatively simple, but it cannot fully consider the logical dependencies between work processes, resource allocation relationships, and spatial distribution characteristics in a project. Once the project is large or complex, schedule deviations between different work processes may affect each other through multiple dependency chains. A single time series model cannot identify these implicit topological constraints, resulting in significant prediction errors. On the other hand, methods for analyzing project monitoring and quality risks typically employ anomaly detection mechanisms based on statistical thresholds. For example, by setting upper or lower thresholds for monitoring indicators, alarms are triggered for data points exceeding the thresholds. This method is simple to implement but is easily affected by noisy data and local environmental changes, and lacks the ability to model the spatial correlation between monitoring points and the structural dependencies between components. Because different components, processes and monitoring points are topologically coupled, a single point exceeding the limit does not necessarily represent the occurrence of a risk event. However, multiple joint anomalies are often the manifestation of systemic problems in engineering. Existing threshold detection mechanisms cannot identify such structured anomalies. Summary of the Invention

[0004] In view of this, the present invention provides a method and system for processing engineering project data. By constructing an engineering knowledge graph under a unified engineering coding system and combining graph signal processing and graph neural network models on the graph structure, it achieves integrated processing of multi-source engineering data fusion, anomaly identification, and risk prediction. The method first organizes components, processes, contract terms, monitoring points, equipment, personnel, and events into a traceable topological structure using unified engineering coding. Then, dynamic information such as schedule deviations, monitoring values, and cost consumption are mapped into node signals and edge signals, forming a computable graph signal-enhanced engineering knowledge graph. By performing graph signal smoothing and high-pass filtering operations, local continuous abnormal regions can be identified. Graph neural networks are then used to jointly predict delay risks, cost overrun risks, and quality risks in these structured abnormal regions, thereby generating risk level results at the component and process levels.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for processing engineering project data includes the following steps executed in sequence:

[0007] Step 1: Construct an engineering knowledge graph based on unified engineering coding;

[0008] Step 2: Construct a graph signal-enhanced engineering knowledge graph in the engineering knowledge graph: Collect multi-source data during the implementation of engineering projects, normalize the collected multi-source data, and map the normalized data into graph signals attached to nodes and associated edges in the engineering knowledge graph to obtain the graph signal-enhanced engineering knowledge graph.

[0009] Step 3: Perform graph signal smoothing and graph signal high-pass filtering on the graph signal enhancement engineering knowledge graph to identify structured anomalous regions in the graph;

[0010] Step four: Based on the identified structured anomaly regions and graph neural network models, generate joint risk prediction results for nodes within the structured anomaly regions, and classify and output the risk levels of the corresponding nodes according to the joint risk prediction results.

[0011] Furthermore, Step One specifically includes: establishing a unique and unified engineering code for each component, process, contract clause, monitoring point, equipment, personnel, and event involved in the project. This unified engineering code is obtained by concatenating the project number, professional code, component or process category code, serial number field, and version field in a preset order. The field lengths are set to 8 digits for the project number, 2 digits for the professional code, 4 digits for the component or process category code, 4 digits for the serial number field, and 2 digits for the version field. Based on the established unified engineering codes, a set of nodes in the project knowledge graph is generated. This set of nodes includes at least component nodes, process nodes, contract clause nodes, monitoring point nodes, equipment nodes, and personnel nodes. The system consists of nodes and event nodes, where each node stores a corresponding unified project code and attribute fields related to the project status. Based on the actual relationships within the project, a set of associated edges is established between nodes. This set of associated edges includes at least sequence edges representing process sequence constraints, subordinate edges representing the subordination relationship between components and processes, configuration edges representing the relationship between processes and equipment / personnel configuration, monitoring edges representing the correspondence between processes and monitoring points, constraint edges representing the correspondence between processes and contract terms, and event edges representing the scope of event impact. This enables the project knowledge graph to form a connected topology, where any component node can be traced back to at least one contract term node and at least one monitoring point node through at most five associated edges.

[0012] Furthermore, step two specifically includes: during the implementation of the project, with a fixed sampling period of 10 minutes, synchronously collecting schedule deviation data, monitoring data, and cost consumption data within the current period from the schedule management system, monitoring system, and cost management system; and performing unit unification and numerical normalization processing on the collected schedule deviation data, monitoring data, and cost consumption data so that all data are converted to a closed interval range of 0 to 1.

[0013] Furthermore, step two specifically includes: for each sampling period, mapping the difference between the actual completion ratio and the planned completion ratio corresponding to the component node and the process node to a schedule deviation signal; mapping the monitoring measurement results corresponding to the monitoring point node to a monitoring signal; and mapping the comparison results of the current cost expenditure and the target cost corresponding to the process node and the component node to a cost consumption signal; forming a signal vector of length 12 by the schedule deviation signal, the monitoring signal, and the cost consumption signal in a fixed order and attaching it to the corresponding component node and process node; generating a signal vector of length 6 by the data directly related to the engineering relationship between nodes according to the same normalization rule and attaching it to the corresponding associated edge to obtain a graph signal enhancement engineering knowledge graph; and generating a complete snapshot of the graph signal enhancement engineering knowledge graph for each sampling period and updating the current version in real time.

[0014] Furthermore, step three specifically includes: For the current version of the graph signal enhancement engineering knowledge graph, for each component node and process node, perform the following signal smoothing operation as shown in the figure: Select adjacent nodes connected by sequential edges, subordinate edges, and monitoring edges from the set of direct adjacent nodes of the target node; if the number of selected adjacent nodes is not less than 2, then sum the signal vector of the target node with the signal vectors of all selected adjacent nodes element by element, divide by the number of selected adjacent nodes plus 1, and obtain the smoothed signal vector; if the number of selected adjacent nodes is 1, then average the signal vectors of the target node with the signal vectors of the selected adjacent nodes element by element, and obtain the smoothed signal vector; if there are no selected adjacent nodes, then keep the original signal vector as the smoothed signal vector; subsequently, for each component node and process node, subtract the original signal vector from the corresponding smoothed signal vector element by element and take the absolute value to obtain the high-pass differential signal vector. When any element in the high-pass differential signal vector is greater than 0.20, this node is marked as an abnormal node.

[0015] Furthermore, step three specifically includes: after completing the anomaly marking of all component nodes and process nodes, a continuous set of nodes in the graph that are connected by sequential edges and subordinate edges and are all marked as anomaly nodes will be identified as a candidate anomaly chain; when the number of anomaly nodes in the identified candidate anomaly chain is greater than or equal to 3 and the total planned time of the process corresponding to the span of the identified candidate anomaly chain is greater than or equal to 24 hours, the identified candidate anomaly chain and the monitoring point nodes directly connected to any node in the identified candidate anomaly chain through monitoring edges will be defined together as a structured anomaly region.

[0016] Furthermore, step four specifically includes: establishing a graph neural network model, which comprises a first graph convolutional layer, a second graph convolutional layer, and an output layer connected sequentially. The first graph convolutional layer expands the input node signal vector of length 12 into a node feature vector of length 16. The second graph convolutional layer compresses the node feature vector of length 16 into a node feature vector of length 8. The output layer maps the node feature vector of length 8 into three output values: delay risk score, cost overrun risk score, and quality risk score. In the first graph convolutional layer, for structured anomalous regions and... For adjacent component nodes and process nodes, the input signal vector of each node is averaged element-wise with the input signal vectors of its adjacent nodes connected by sequential edges and subordinate edges to obtain an intermediate feature vector. Then, values ​​less than 0 in the obtained intermediate feature vector are set to 0, and values ​​greater than 1 are set to 1. The processed intermediate feature vector is used as the node feature vector of the corresponding node with a length of 16. In the convolutional layer of Figure 2, for each node, its node feature vector of length 16 is divided into 8 consecutive groups, each group containing 2 elements. The average of the 2 elements in each group is used to obtain a node feature vector of length 8.

[0017] Furthermore, step four specifically includes: in the output layer, for each node, the average of the 1st to 3rd elements in the node feature vector of length 8 is used as the delay risk score, the average of the 4th to 6th elements is used as the cost overrun risk score, and the average of the 7th to 8th elements is used as the quality risk score; when any risk score is greater than 0.70, the corresponding component node or process node is marked as a high-risk node; when any risk score is between 0.40 and 0.70, the corresponding node is marked as a medium-risk node; when all risk scores are less than or equal to 0.40, the corresponding node is marked as a low-risk node; and the unified engineering code, risk category, and risk score corresponding to the high-risk node and medium-risk node are output to the project management system interface.

[0018] A data processing system based on engineering projects includes: a knowledge graph construction module for constructing an engineering knowledge graph based on a unified engineering coding system; a graph signal enhancement module for collecting multi-source data during the implementation of engineering projects, normalizing the collected multi-source data, and mapping the normalized data to graph signals attached to nodes and associated edges in the engineering knowledge graph, thus obtaining a graph signal-enhanced engineering knowledge graph; an anomaly region identification module for performing graph signal smoothing and graph signal high-pass filtering on the graph signal-enhanced engineering knowledge graph to identify structured anomaly regions in the graph; and a risk prediction module for generating joint risk prediction results for nodes within the identified structured anomaly regions based on the identified structured anomaly regions and a graph neural network model, and classifying and outputting the risk levels of the corresponding nodes according to the joint risk prediction results.

[0019] By adopting the above technical solutions, this invention achieves the following beneficial effects: First, by unifying components, processes, contract terms, monitoring points, equipment, personnel, and events into a single structured semantic space through an engineering knowledge graph, it enables the topological expression of traditionally decentralized engineering information, ensuring the consistency of various data in time, space, and logic. Subsequently, by constructing a graph signal-enhanced engineering knowledge graph, multi-source dynamic data such as schedule deviations, monitoring values, and cost consumption are mapped into node and edge signals, achieving a quantitative characterization of the engineering operation status. This allows complex engineering systems to be uniformly calculated and analyzed in the form of a graph structure. Furthermore, by performing graph signal smoothing and high-pass filtering, local signal anomaly regions can be automatically identified in the engineering knowledge graph, thereby discovering systemic deviations and potential hazards that are difficult to detect using traditional methods. Finally, by utilizing a graph neural network model to jointly predict delay risks, cost overrun risks, and quality risks within structured anomaly regions, an intelligent closed loop from anomaly detection to risk warning is achieved. This invention can simultaneously analyze engineering behavior and resource status changes at the node and edge levels, possessing high interpretability and traceability. Compared to existing risk analysis methods that rely on single indicators or static graph structures, this invention not only accurately reflects the dynamic correlation characteristics of project progress but also generates real-time risk level distributions at the component and process levels. This facilitates managers in quickly locating the source of problems and developing targeted control measures. This method significantly improves the computability of engineering data, the accuracy of anomaly identification, and the timeliness of risk prediction, offering significant technical advantages and practical application value for the refined management and intelligent decision-making of complex, large-scale engineering projects. Attached Figure Description

[0020] Figure 1 is a schematic diagram of the engineering knowledge graph topology in an embodiment of the present invention;

[0021] Figure 2 is a time-series evolution curve of multi-node risk scoring in an embodiment of the present invention;

[0022] Figure 3 is an experimental curve showing the effect of the high-pass differential threshold on the anomaly recognition performance in an embodiment of the present invention. Detailed Implementation

[0023] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0024] Any feature disclosed in this specification, unless otherwise stated, may be replaced by other equivalent or similar features. That is, unless otherwise stated, each feature is merely one example of a series of equivalent or similar features.

[0025] A method for processing engineering project data includes the following steps executed in sequence:

[0026] Step 1: Construct an engineering knowledge graph based on unified engineering coding.

[0027] The specific implementation of step one can be carried out as follows. The following content is as a directly executable implementation method, while also providing several optional implementation methods. All terms are consistent with "unified engineering coding" and "engineering knowledge graph" in step one above.

[0028] In one implementation, the rules for constructing a unified project code are first determined. The unified project code is used to uniquely identify components, processes, contract terms, monitoring points, equipment, personnel, and events within the scope of an engineering project, ensuring stable and traceable references to the same engineering object during multi-source data access and subsequent diagram structure construction. The unified project code is generated using a fixed-format field concatenation method, including an engineering project number field, a professional code field, a component or process category code field, a serial number field, and a version field. These five fields are directly concatenated in a fixed order to form a single string, without inserting other delimiters. The project number field is 8 digits long and is used to represent the unique number of the project in the enterprise's project master data; the professional code field is 2 digits long and is used to represent professional categories such as civil engineering, mechanical and electrical engineering, decoration, track, locomotive equipment, etc. It can use a pre-configured numerical coding table, for example, civil engineering is "01", mechanical and electrical engineering is "02", and decoration is "03"; the component or process category code field is 4 digits long and is used to mark the specific component type or process type, for example, bridge pier is "0301", box girder is "0302", and tunnel excavation process is "0401"; the serial number field is 4 digits long and is used to distinguish specific component or process instances under the same category, starting from "0001" and increasing sequentially; the version field is 2 digits long and is used to mark the version serial number of the associated information of this project object, starting from "01" and increasing sequentially. By using fixed-length fields and fixed concatenation order, the uniform engineering code can be guaranteed to be deterministic and sortable when comparing strings, which facilitates the rapid retrieval and grouping of nodes in the engineering knowledge graph according to uniform rules, while avoiding confusion between components or processes with the same name in different projects or different versions.

[0029] In the specific implementation process, for each component, a unified engineering code corresponding to the component node is generated item by item according to the component list in the construction drawings and building information model. For example, for a certain civil engineering bridge pier with project number "20251234", a unified engineering code "202512340103000101" can be generated, where "20251234" is the project number field, "01" is the professional code field (representing civil engineering), "0301" is the component category code field (representing the pier category), "0001" is the sequence number of the bridge pier in this project, and "01" is the first version. For construction processes, based on the construction organization design and the breakdown of sub-projects, each construction process is mapped to a corresponding component or construction area, generating a unified engineering code for the process node, such as "202512340401001001". Here, the component or process category code field "0401" marks a specific concrete pouring process, the sequence number field "0010" indicates the 10th process instance in the project, and the version field reflects records after adjustments to process parameters. For contract terms, monitoring points, equipment, personnel, and events, unified engineering codes are generated using the same structure based on contract files, monitoring point lists, equipment ledgers, personnel information systems, and project event records. For example, contract terms may use a fixed combination of professional and category codes to differentiate them from the construction object, but the coding structure remains consistent, ensuring all nodes are managed under the same coding system.

[0030] Referring to Figure 1, as shown in Figure 1, the engineering knowledge graph constructed by this invention contains multiple node types and associated edge types, forming a complete topological structure. Regarding node types, the engineering knowledge graph includes seven basic node types. Component nodes are represented by circles, such as component node A, component node B, and component node C. Each component node is accompanied by a unified engineering code, which is composed of an 8-digit project number field, a 2-digit professional code field, a 4-digit component or process category code field, a 4-digit serial number field, and a 2-digit version field, arranged in a fixed order. For example, the unified engineering code for component node A is "202512340103000101", where "20251234" is the project number, "01" is the professional code representing civil engineering, "0300" is the component category code representing the pier category, and "0101" is the component's serial number and version information within the project. Process nodes are represented by squares, such as process node 1, process node 2, process node 3, and process node 4. They also follow a unified engineering coding rule; process node 1 is coded as "202512340401001001", where "0401" indicates the concrete pouring process category. Monitoring point nodes are represented by rhombuses; for example, monitoring point M1 is used for settlement monitoring, and monitoring point M2 is used for strain monitoring. Contract clause nodes are represented by hexagons; for example, contract clause N1 represents quality standard constraints, and contract clause N2 represents schedule requirement constraints. Equipment and personnel nodes are represented by triangles; for example, equipment E1 represents a concrete pump truck, and personnel P1 represents the project manager. Event nodes are represented by pentagons; for example, event V1 represents a quality problem event. Regarding the types of associated edges, the engineering knowledge graph includes six types of associated edges, each distinguished by a different line type. Sequential edges, represented by solid lines, connect process nodes with sequential relationships. For example, process node 1 is connected to process node 2, process node 2 to process node 3, and process node 3 to process node 4, respectively, reflecting the temporal connection of construction processes. Subordinate edges, represented by long dashed lines, connect process nodes to the component nodes they affect. For example, process node 1 is subordinate to component node A, process node 2 to component node B, and process node 3 to component node C, allowing each process to be precisely located to its affected component. Monitoring edges, represented by dashed-dot lines, connect component nodes to monitoring point nodes. For example, monitoring edges are established between component node B and monitoring point M1, and between component node C and monitoring point M2, indicating that the component is observed by a specific monitoring point. Constraint edges, represented by double dashed-dot lines, connect process nodes or component nodes to contract clause nodes. For example, constraint edges are established between process node 1 and contract clause N1, and between process node 3 and contract clause N2, indicating that the process is constrained by a specific contract clause.Configuration edges, represented by short dashed lines, connect process nodes with equipment or personnel nodes. For example, configuration edges are established between process node 2 and equipment E1, and between process node 3 and personnel P1, indicating the specific equipment and personnel involved in performing the process. Event edges, represented by dashed lines, connect event nodes with affected engineering object nodes. For example, event V1 connects component node C and process node 3 via an event edge, indicating that the quality issue event affects these engineering objects. Regarding topological features, the engineering knowledge graph shown in Figure 1 forms a connected topology, where any component node can access related contract clause nodes and monitoring point nodes through a finite number of associated edges. For example, starting from component node A, one can reach process node 1 along dependent edges, and then reach contract clause N1 along constraint edges; this tracing path is 2 layers long. Starting from component node C, one can reach process node 3 along dependent edges, then reach process node 2 along sequence edges, then reach component node B along dependent edges, and then reach monitoring point M1 along monitoring edges; this tracing path is 4 layers long. According to the limited hop count traceability configuration strategy of this invention, any component node must be able to access at least one contract clause node and at least one monitoring point node within a range of no more than 5 layers of associated edges. This ensures that the engineering knowledge graph meets traceability requirements while controlling topological complexity. The unified engineering code is composed of 5 fields in a fixed format, with a total length of 20 characters. The format is "Project Number (8 digits) + Professional Code (2 digits) + Category Code (4 digits) + Serial Number (4 digits) + Version (2 digits)". By using fixed-length fields and a fixed concatenation order, the unified engineering code can be guaranteed to be deterministic and sortable when comparing strings. This facilitates the rapid retrieval and grouping of nodes in the engineering knowledge graph according to unified rules, while avoiding confusion between components or processes with the same name in different projects or different versions.

[0031] After the unified engineering code is generated, the components, procedures, contract terms, monitoring points, equipment, personnel, and events involved in the project are created as nodes in the engineering knowledge graph, with node types completely corresponding to the aforementioned classifications. Each node stores its unified engineering code and basic attribute information related to the project status. The basic attribute information includes name, location description, construction section, construction unit name, responsible personnel name, planned start time, planned completion time, design parameter reference identifier, contract agreement content reference identifier, monitoring point spatial coordinates, equipment rated capacity, personnel qualification level, event occurrence time, and event classification. For example, for a component node, its pier number, elevation information, contract section number, and unified engineering code list of associated procedures can be recorded; for a monitoring point node, its three-dimensional coordinate numerical information, monitoring type (deformation, settlement, stress), sensor number, and installation time can be recorded; for an event node, the event type (work stoppage, quality problem, safety accident), the unified engineering code of the component node corresponding to the location of occurrence, and the duration of the impact can be recorded.

[0032] After generating nodes, a set of associated edges is created based on the actual engineering relationships, forming a connected topology for the engineering knowledge graph. Sequential edges describe the order of operations; for example, a sequential edge is established between the "foundation excavation operation node" and the "reinforcement binding operation node," and between the "reinforcement binding operation node" and the "concrete pouring operation node." This ensures that the operation nodes associated with any component node clearly reflect the execution order on the topological path, facilitating subsequent tracking of progress anomaly propagation paths along the sequential edges on the graph. Dependent edges connect component nodes to their corresponding operation nodes, allowing each operation node to be precisely located to its affected component node via dependent edges. For example, the concrete pouring operation node is connected to its corresponding pier component node via a dependent edge, enabling subsequent risk prediction at the component level to directly aggregate relevant operation information. Configuration edges connect operation nodes to equipment and personnel nodes, representing the specific equipment and personnel used to execute the operation, allowing subsequent analysis of the relationship between resource configuration changes and risks based on the graph structure. Monitoring edges connect process nodes or component nodes to monitoring point nodes, indicating that the process or component is observed by a specific monitoring point. This allows for direct mapping of monitoring signals to construction activities when identifying structured anomalies. Constraint edges connect process nodes or component nodes to contract clause nodes, indicating that the process or component is constrained by specific contract clauses, such as schedule requirements, quality standards, and payment terms. This allows for direct assessment of the impact of abnormal behavior on contract performance during risk identification. Event edges connect event nodes to component nodes, process nodes, equipment nodes, and personnel nodes, indicating which engineering objects are affected by a particular event. For example, a quality problem event node may be linked to both the relevant component nodes and related process nodes via event edges.

[0033] To ensure the traceability of the engineering knowledge graph while controlling topological complexity, a tracing configuration strategy with a limited number of hops is adopted when establishing associated edges. Specifically, for each component node, it is traversed layer by layer along sequential edges, subordinate edges, constraint edges, and monitoring edges. When the traversal layer count does not exceed 5 layers, if at least one contract clause node and at least one monitoring point node are found, the tracing path of that component node is recorded and marked as a fully traceable component node. When the traversal layer count reaches 5 layers and no contract clause node or monitoring point node is found, the missing constraint edges or monitoring edges are supplemented by checking the original data of the engineering project until the component node can be associated with at least one contract clause node and at least one monitoring point node within the 5-layer range. By limiting the maximum number of traversal layers to 5, any component node can locate contract constraint information and monitoring information within a limited number of steps. This ensures that the traceability path length is controllable, facilitating efficient querying and graph signal processing in the engineering knowledge graph, while avoiding topological redundancy and a sharp increase in computational overhead caused by unconstrained expansion. From an engineering implementation perspective, this ensures operability and stability of resource consumption.

[0034] In one optional implementation, the field length of the unified engineering code can be adjusted according to the scale and complexity of the project. For example, when the number of projects is small but the number of components and processes is large, the project number field can be set to 6 digits, the component or process category code field can be set to 6 digits, and the sequence number field can be set to 6 digits to increase the numbering space for components and processes within a single project. Alternatively, the specific value range of the professional code field and the category code field can be adjusted according to different enterprise or industry standards, but the number of each field remains unchanged, and the unified engineering code still adopts a fixed structure of 5 fields sequentially concatenated to ensure that the downstream processing logic remains consistent. In another optional implementation, additional node types for representing geographical regions, construction segments, and construction area divisions can be introduced into the engineering knowledge graph. Component nodes and process nodes are connected to corresponding region nodes through subordinate edges, so that when identifying structured abnormal regions, multiple adjacent abnormal component nodes can be automatically merged into an abnormal subgraph within the same spatial region, improving the ability to express local concentrated features, while maintaining the overall connected topology of the engineering knowledge graph and ensuring that the traceability path of any component node can access the corresponding contract clause node and monitoring point node within a range of no more than 5 associated edges.

[0035] Step 2: Construct a graph signal-enhanced engineering knowledge graph in the engineering knowledge graph: Collect multi-source data during the implementation of engineering projects, normalize the collected multi-source data, and map the normalized data into graph signals attached to nodes and associated edges in the engineering knowledge graph to obtain the graph signal-enhanced engineering knowledge graph.

[0036] In this step, by systematically collecting multi-source data during the implementation of the engineering project, performing unified numerical processing, and mapping it one-to-one with the engineering knowledge graph, discrete and heterogeneous engineering state information is transformed into graph signals attached to nodes and associated edges. This forms a graph signal-enhanced engineering knowledge graph based on the topological structure of the engineering knowledge graph, providing a direct and computable input basis for subsequent graph signal smoothing, graph signal high-pass filtering, and risk prediction by graph neural network models.

[0037] In one implementation, a pre-set collection cycle of 10 minutes is used for the constructed engineering knowledge graph. Multi-source data generated during the project implementation is collected synchronously according to this cycle. The scope of multi-source data includes at least the planned completion data and actual completion data output by the progress management system, environmental monitoring data and structural monitoring data output by the monitoring system, component cost expenditure data output by the cost management system, and other business data directly related to the project status. During collection, a unified engineering code is used as the basis for data association. All data entering the processing flow must be labeled with the corresponding unified engineering code during the collection phase, ensuring that each piece of data, upon reaching the processing end, establishes a stable mapping relationship with component nodes, process nodes, contract clause nodes, monitoring point nodes, equipment nodes, personnel nodes, and event nodes in the engineering knowledge graph. By adopting a fixed collection cycle of 10 minutes, both data timeliness and data volume can be balanced in most civil engineering, electromechanical, and rail engineering scenarios. On the one hand, monitoring data and progress data can reflect changes in status within a short period; on the other hand, it avoids the data volume from increasing exponentially due to excessively short collection cycles, significantly increasing storage and processing burdens, and maintaining feasibility and maintainability from the perspective of engineering system deployment.

[0038] During data normalization, to enable direct comparison and joint processing of data from different sources, with different dimensions, and different numerical ranges within the same graph structure, the collected progress deviation data, monitoring data, and cost consumption data undergo unit unification and numerical range conversion. In specific implementation, firstly, according to the data standard configuration determined before the project execution, fixed maximum and minimum reference values ​​are established for various types of data. For example, for the progress completion percentage, 0 is set as the minimum reference value and 1 as the maximum reference value; for construction site settlement monitoring data, 0 is set as the minimum reference value and 50 as the maximum reference value, in millimeters; for concrete structure strain monitoring data, 0 is set as the minimum reference value and 2000 as the maximum reference value, in microstrain; for the cost consumption percentage of sub-items, 0 is set as the unconsumed state and 1 as the target cost ceiling has been reached. During normalization, for each piece of collected data, a preset minimum and maximum reference value is used based on its data type to linearly map the original value to the range of 0 to 1. If the original value is lower than the minimum reference value, it is directly set to 0; if the original value is higher than the maximum reference value, it is directly set to 1. By using a pre-defined reference range and linear mapping method, on-site personnel do not need to derive complex mathematical formulas. They only need to look up and transform the input data according to the configuration table to obtain normalized data that meets the uniform range requirements. This processing method makes the values ​​from the progress management system, monitoring system, and cost management system comparable at the graph signal level, thereby avoiding numerical bias caused by differences in units. This facilitates subsequent graph signal smoothing and high-pass filtering to identify abnormal behavior on the same scale.

[0039] After normalization, the normalized data is mapped to graph signals attached to nodes and associated edges in the engineering knowledge graph, thereby constructing a graph signal-enhanced engineering knowledge graph based on the engineering knowledge graph. In one specific implementation, a snapshot of the graph signal-enhanced engineering knowledge graph is constructed on a per-collection-cycle basis. For each collection cycle, the normalized data is grouped according to a unified engineering code. Data related to component nodes and process nodes is mapped to node graph signals, and data directly related to the engineering relationships between nodes is mapped to associated edge graph signals. For component nodes and process nodes, for each collection cycle, schedule deviation information, monitoring information, and cost consumption information are selected from the normalized data and combined in a fixed order. A total of 12 values, including multiple indicators of schedule deviation information, multiple indicators of monitoring information, and multiple indicators of cost consumption information, are arranged into a node signal vector of length 12, and this node signal vector is attached to the corresponding component node or process node. For example, within a data acquisition cycle, for a specific component node, 12 normalized values ​​can be selected, including the difference between the planned and actual completion rates, the settlement changes over three consecutive acquisition cycles, the strain changes over three consecutive acquisition cycles, the current cost consumption rate, the cumulative cost consumption rate, and the quality inspection pass rate related to the component. These values ​​are then arranged in a preset order to form a node signal vector and recorded in the graph signal field of the component node. This setup ensures that nodes of the same type have graph signals with fixed dimensions and a fixed order in each acquisition cycle. Engineering technicians can directly encode and read the signals based on the predefined field descriptions, reducing ambiguity during implementation.

[0040] For associated edges, behavioral or state data directly related to the engineering relationship between nodes are selected and organized into edge signal vectors of length 6. For example, for sequential edges, the normalized value of the planned interval between preceding and following processes relative to the baseline interval, the completion status value of the preceding process, and the start status value of the subsequent process can be recorded; for subordinate edges, the contribution ratio of the process completion progress to the overall completion of the component and the current completion status of the component can be recorded; for monitoring edges, whether the monitoring point generates a status flag close to the warning threshold within the current acquisition cycle and the spatial distance ratio with the monitored component can be recorded. By configuring fixed structural edge signal vectors of length 6 for associated edges, each associated edge in the engineering knowledge graph not only expresses the topological relationship but also carries quantitative information related to that relationship. This is beneficial for considering both node status and relationship status simultaneously in subsequent graph signal smoothing and graph signal high-pass filtering, enabling a holistic analysis of the engineering status on the topological structure.

[0041] When generating the graph signal enhancement engineering knowledge graph, for each acquisition cycle, a complete snapshot of the graph signal enhancement engineering knowledge graph is generated based on the currently acquired normalized data and the existing engineering knowledge graph structure, and this snapshot is marked as the current version. Each component node, process node, contract clause node, monitoring point node, equipment node, personnel node, and event node in the current version maintains the unified engineering code and association relationships determined in step one, while supplementing with attached node signal vectors and edge signal vectors. When entering the next acquisition cycle, a new snapshot of the graph signal enhancement engineering knowledge graph is generated based on the new normalized data, and this snapshot is set as the new current version, while retaining several recent versions for comparative analysis. In actual deployment, the most recent 144 versions can be retained to cover graph signal changes within a 24-hour timeframe. This time-progressive version generation method ensures that the engineering knowledge graph not only reflects the static topological relationships between engineering objects but also continuously records the state evolution trajectory during project implementation, providing direct data support for subsequent time-related anomaly identification and risk prediction on the same topological structure.

[0042] In one optional implementation, when the monitoring data update frequency of the engineering project exceeds 10 minutes, the acquisition cycle can be shortened to 5 minutes or 1 minute while maintaining the engineering knowledge graph structure. This improves the responsiveness of the graph-enhanced engineering knowledge graph to rapidly changing states. When the engineering project is large-scale and communication resources are limited, the acquisition cycle can be adjusted to 15 minutes or 30 minutes to reduce data transmission pressure. The acquisition cycle adjustment is set through a configuration file before implementation and fixed during on-site deployment. Subsequent normalization processing rules, node signal vector lengths, and edge signal vector lengths remain unchanged, ensuring that downstream graph signal smoothing and graph signal high-pass filtering can run directly without modifying the algorithm structure. In another alternative implementation, the graph signal dimension can be increased according to the professional characteristics of the engineering project. For example, in projects with a large number of high-pier bridges or deep foundation pits, additional monitoring data such as wind speed, temperature, and groundwater level related to component nodes can be introduced to extend the node signal vector length to 16 or 20. However, the extended length is fixed in the system design stage and the same length is uniformly adopted in the graph signal enhancement engineering knowledge graph, so that the implementation of this step still maintains clear field definitions and a directly coded execution method.

[0043] Step 3: Perform graph signal smoothing and graph signal high-pass filtering on the graph signal enhancement engineering knowledge graph to identify structured anomalous regions in the graph.

[0044] In one implementation, the scope of nodes and associated edges participating in graph signal smoothing and high-pass filtering are first determined for the current version of the graph signal enhancement engineering knowledge graph. The scope of nodes participating in the processing includes all component nodes and process nodes. The scope of associated edges participating in the processing includes sequential edges, subordinate edges, and monitoring edges connecting component nodes and process nodes. Sequential edges, subordinate edges, and monitoring edges are chosen as the basis for smoothing adjacency relationships because sequential edges reflect the temporal connection between processes, subordinate edges reflect the direct interaction between processes and components, and monitoring edges reflect the observation relationship between components / processes and monitoring points. Using these three types of edges as signal propagation paths allows for information fusion concentrated in localized areas where there is direct impact or close coupling in the actual engineering process. This avoids noise introduced by irrelevant associated edges, enabling the smoothing process to highlight the real collaborative change characteristics under the same construction unit, the same technical link, and the same monitoring association.

[0045] When performing graph signal smoothing, for each component node or process node, based on the adjacency relationships defined in the graph signal enhancement engineering knowledge graph, the smoothed signal vector is calculated as follows: For the current target node, firstly, traverse all adjacent nodes directly connected to it via sequential edges, subordinate edges, and monitoring edges to form an adjacent node set; when the adjacent node set contains two or more adjacent nodes, sum the node signal vector corresponding to the current acquisition cycle of the target node with the node signal vector of each adjacent node in the adjacent node set at the same position to obtain an intermediate vector, then divide the value at each position of the intermediate vector by the number of adjacent nodes and add 1 to obtain a smoothed signal vector with the same length as the original node signal vector; when the adjacent node set contains only one adjacent node, sum the node signal vector of the target node with the node signal vector of the adjacent node at each position and divide by 2 to obtain the smoothed signal vector; when the current target node does not have any adjacent nodes under the above three types of associated edges, directly use the original node signal vector of the target node as the smoothed signal vector of the target node. By using this fixed averaging method, the smoothed signal vector of each component node and process node can numerically reflect its collective behavioral characteristics in the local topological environment of the engineering knowledge graph: when the target node and its adjacent nodes are in a coordinated state, the smoothed signal vector is close to the original node signal vector, reflecting that the engineering state at this location has continuity in space and process links; when the state of the target node is significantly different from that of the adjacent nodes, the smoothed signal vector will converge to the normal state of the adjacent nodes, thereby amplifying this local inconsistency in the subsequent high-pass differential processing, which is used to identify local areas where there may be engineering anomalies.

[0046] After smoothing the graph signal, high-pass filtering is performed on each component node and process node to extract significant differences between the target node and its smoothed environment. Specifically, the difference between the original node signal vector and the obtained smoothed signal vector at each corresponding position of the target node in the current acquisition cycle is calculated, and the absolute value of the difference is taken to form a high-pass differential signal vector with the same length as the original node signal vector. Each value in the high-pass differential signal vector represents the degree of deviation of the target node from its local adjacent environment in terms of the corresponding engineering indicator. When the value at any position in the high-pass differential signal vector is greater than 0.20, the component node or process node is marked as an abnormal node. Setting 0.20 as a fixed threshold is necessary for clear differentiation within a unified normalized numerical space ranging from 0 to 1. When the deviation of a certain indicator reaches or exceeds 0.20, it indicates that the relative deviation of that indicator relative to surrounding nodes reaches 20% or more. In scenarios such as project schedule deviation, monitoring offset, or cost deviation, this usually means that the deviation is sufficient to affect local construction behavior or project safety status. Using a fixed threshold allows implementers to directly configure and execute the judgment logic in the system without complex parameter tuning, which is beneficial for obtaining stable and consistent anomaly identification behavior in different projects. When the values ​​at all positions in the high-pass differential signal vector are less than or equal to 0.20, the node is kept as a non-abnormal node, indicating that the node's various states are basically coordinated with its local environment.

[0047] After marking all component nodes and process nodes as anomalous nodes, structured anomalous regions that meet the continuity and influence range conditions are identified based on the marked set of anomalous nodes and the sequential and dependent edges in the engineering knowledge graph. Specifically, in the graph signal enhancement engineering knowledge graph, the component nodes and process nodes marked as anomalous nodes are continuously traversed along sequential and dependent edges. When multiple anomalous nodes are sequentially connected by sequential or dependent edges without interruption by non-anomalous nodes, this set of anomalous nodes continuously connected along the engineering topology path is identified as a candidate anomalous chain. For each candidate anomalous chain, the number of anomalous nodes it contains and the total planned process time corresponding to the candidate anomalous chain are counted. The total planned process time is accumulated based on the difference between the planned start time and planned completion time recorded in the process node attributes in step one. When the number of anomalous nodes in a candidate anomalous chain is greater than or equal to 3, and the total planned process time corresponding to the candidate anomalous chain is greater than or equal to 24 hours, the candidate anomalous chain is confirmed as a critical anomalous chain that meets the continuity and scale requirements. The threshold for the number of nodes is set to 3 because a single abnormal node or only two consecutive abnormal nodes may be caused by factors such as local measurement errors or short-term construction adjustments in engineering practice. However, three or more consecutive abnormal nodes usually reflect a systemic abnormal state in a certain work surface, a certain process section, or a certain component group. The threshold for the total planned time of the process is set to 24 hours because 24 hours corresponds to a day in conventional construction organization. When a continuous abnormal chain covers a planned construction time of one day or more, the abnormality has a real impact on the overall project duration and resource allocation. By simultaneously introducing the abnormal node number condition and the planned time condition of the process, it is possible to distinguish between scattered abnormal points and structured abnormal chains that truly have engineering impact, avoiding the generation of too many invalid warnings for single-point fluctuations.

[0048] After identifying the critical anomaly chain, all anomalous component nodes and anomalous process nodes contained in the critical anomaly chain, as well as monitoring point nodes directly connected to any node in the critical anomaly chain via monitoring edges, are collectively grouped to form a structured anomaly region. Within this structured anomaly region, component nodes, process nodes, and monitoring point nodes form closely interconnected local subgraphs through sequential edges, subordinate edges, and monitoring edges. These subgraphs contain both consecutive anomalous construction behavior nodes and monitoring observation points directly related to these nodes. By including monitoring point nodes within the structured anomaly region, the graph signals from these monitoring points can be directly used in subsequent risk prediction to cross-validate the physical entity states corresponding to the anomaly chain, significantly reducing the false alarm probability caused by relying on a single data source. Simultaneously, the structured anomaly region is confined within the actual association paths of the engineering knowledge graph, covering both the key dependencies in the actual construction process and avoiding disordered expansion across unrelated nodes. This allows the subsequent graph neural network model to learn risk characteristics within a defined local context, improving the relevance and stability of the prediction results.

[0049] In one optional implementation, the anomaly threshold of the high-pass differential signal vector can be configured hierarchically according to the project type. For example, for high-risk bridge projects, the threshold for marking nodes as anomalous nodes is set to 0.15 to improve sensitivity to minor adverse changes; for general building construction projects, the threshold is kept at 0.20 to balance sensitivity and false alarm rate. In another optional implementation, when identifying candidate anomaly chains, constraints on spatial continuity can be added. That is, the candidate anomaly chain is only included in the structured anomaly region if the spatial distance between corresponding component nodes in the candidate anomaly chain is continuously distributed within the same construction area or adjacent construction areas, to prevent the mismerging of anomalous nodes scattered over long distances into a single anomaly region.

[0050] Step four: Based on the identified structured anomaly regions and graph neural network models, generate joint risk prediction results for nodes within the structured anomaly regions, and classify and output the risk levels of the corresponding nodes according to the joint risk prediction results.

[0051] In one implementation, a subgraph to be evaluated is first extracted from the graph signal enhancement engineering knowledge graph based on structured anomalous regions. For each structured anomalous region, the component nodes and process nodes contained within it, as well as the monitoring point nodes directly connected to these nodes via sequential edges, dependent edges, and monitoring edges, are selected together. The sequential edges, dependent edges, and monitoring edges, as well as the edge signal vectors attached to these edges, are retained to form a subgraph for risk prediction. The reason for selecting structured anomalous regions as the extraction target is that these regions already meet the conditions for the number of anomalous nodes and the total planned time of processes, indicating that the signal deviation in these regions has continuity and engineering impact. Limiting the computational scope of the graph neural network model to such regions allows for concentrated computational resources to perform detailed analysis of high-value regions, reducing the resource waste caused by repeated calculations in a large number of normal regions. At the same time, it reduces the probability of the graph neural network model learning invalid patterns in regions without anomalies, making the output joint risk prediction results more focused on objects with practical value for engineering decisions.

[0052] After selecting the subgraph, to ensure the consistency and direct processing of the input data for the graph neural network model, following the rules defined in step two above, a node signal vector of length 12 for each component node and process node, as well as an edge signal vector of length 6 associated with sequential and dependent edges, are read from the graph signal enhancement engineering knowledge graph. In this embodiment, when the graph neural network model predicts node-level risks, it uses the node signal vector as the main input and utilizes the topological relationships reflected by sequential and dependent edges to determine the information propagation path between nodes. The state of the monitoring point node is reflected in the node signal vectors of the connected component nodes and process nodes through the monitoring edges. Therefore, when executing the graph neural network model, the monitoring point node can be regarded as an indirect information source that has been integrated into the node signal through the graph signal construction process, thereby maintaining the stability of the length and meaning of the node input vector in structure.

[0053] In this embodiment, the graph neural network model adopts a three-layer sequential processing structure, including a first graph convolutional layer, a second graph convolutional layer, and an output layer. The function of the first graph convolutional layer is to convert the node signal vector of length 12 for each component node and process node into a node feature vector of length 16 while preserving the graph structure constraints, so that the node features can integrate information about their own state and direct adjacency relationships. To achieve this, for all component and process nodes directly connected to any component or process node within the structured anomaly region via sequential and subordinate edges, the following processing flow is executed: For each target node, its node signal vector of length 12 is read, and simultaneously, the node signal vectors of all adjacent nodes connected to the target node via sequential and subordinate edges of length 12 are read. The node signal vector of the target node and the node signal vectors of all adjacent nodes are summed at each position and divided by the number of nodes to obtain an adjacency fusion vector of length 12. Subsequently, the 12 values ​​of this adjacency fusion vector are mapped to 16 values ​​according to a preset expansion rule. For example, the first 6 values ​​are copied sequentially and appended to the end to form a node feature vector of length 16. Through this fixed-rule adjacency fusion processing, the output of the first graph convolutional layer of each node simultaneously contains the local state of the node and the states of adjacent nodes with direct process sequence or component subordinate relationships. Engineering implementers only need to perform summation, averaging, and sequential copying according to the given steps to complete the calculation, without configuring weights or writing complex training processes.

[0054] In the second convolutional layer, the 16-length node feature vector output from the first convolutional layer is compressed to a length of 8, reducing subsequent computation while preserving key differences. Specifically, for each node's 16-length feature vector, the 16 values ​​are sequentially divided into 8 groups, each containing two adjacent values. The sum of the two values ​​in each group is divided by 2 to obtain the corresponding output value, forming the 8-length node feature vectors. This deterministic grouping and averaging method effectively merges and compresses the output features of the first convolutional layer, combining repetitive or highly similar features into more stable representative values, reducing the impact of noise fluctuations on the final risk score, and maintaining consistency in feature calculation rules across nodes. Engineers can implement this process entirely based on the field order without introducing uninterpretable training parameters.

[0055] In the output layer, three output values—delay risk score, cost overrun risk score, and quality risk score—are generated based on node feature vectors of length 8. These scores are then used to classify the risk level of each component node and process node. In this implementation, for each node, the first to third values ​​in the 8-length node feature vector are summed and divided by 3 to obtain the delay risk score; the fourth to sixth values ​​are summed and divided by 3 to obtain the cost overrun risk score; and the seventh and eighth values ​​are summed and divided by 2 to obtain the quality risk score. This grouping method prioritizes features related to time schedule in the first three positions, features related to cost consumption in the middle three positions, and features related to quality performance and monitoring feedback in the last two positions. This ensures that each risk score is determined by a set of pre-defined feature values, resulting in a stable score through simple averaging. For each score, the numerical range remains between 0 and 1, allowing implementers to directly determine the risk level based on the threshold.

[0056] During the risk level classification process, a fixed threshold strategy is used to classify and label nodes. When any one of the delay risk score, cost overrun risk score, or quality risk score of a component node or process node is greater than 0.70, the node is marked as a high-risk node, indicating that the component or process corresponding to the node shows a significantly high risk tendency in at least one risk dimension. Such nodes are given priority for inclusion in the key investigation scope of engineering management personnel. When all scores are not greater than 0.70 and at least one score is greater than 0.40, the node is marked as a medium-risk node, indicating that the node has a certain degree of deviation or adverse trend, and it is recommended to pay more attention to and review it during subsequent construction and monitoring. When the delay risk score, cost overrun risk score, and quality risk score are all not greater than 0.40, the node is marked as a low-risk node, indicating that the node is at an acceptable risk level under the current state of the graph signal enhancement engineering knowledge graph. Using 0.40 and 0.70 as unified thresholds facilitates direct comparison of risk levels across projects and sections. Furthermore, by providing each node with a clear three-stage level, it enables direct integration with early warning rules and handling procedures in the project management system, reducing the need for manual interpretation and ensuring the system output is executable.

[0057] Referring to Figure 2, this invention continuously monitors the risk scores of multiple component nodes and process nodes identified within a structured anomaly area, and displays the risk score change trend of each node within a continuous acquisition cycle through a time-series evolution curve. Regarding the coordinate system setup, Figure 2 uses a standard two-dimensional rectangular coordinate system. The horizontal axis represents the acquisition cycle number, with a scale range from 0 to 24, corresponding to 24 consecutive acquisition cycles. Each acquisition cycle is 10 minutes, therefore the horizontal axis covers a time range of 240 minutes, or 4 hours. The horizontal axis scale is represented by integer numbers, with a scale value marked every 4 acquisition cycles, i.e., 0, 4, 8, 12, 16, 20, 24. The corresponding actual time is indicated in parentheses below the scale value; for example, acquisition cycle number 4 corresponds to "(40 minutes)" and acquisition cycle number 12 corresponds to "(120 minutes)". The unit "piece" is marked at the end of the horizontal axis, indicating that the acquisition cycle number is a dimensionless counting unit. The vertical axis represents the risk score, ranging from 0.0 to 1.0, divided into equal increments of 0.1. The values ​​are labeled 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0. The unit of the vertical axis is a normalized dimensionless score; a higher value indicates a higher level of risk. The words "Risk Score" are vertically labeled to the left of the vertical axis to explain its physical meaning. The axis is drawn with a solid black line, 3 pixels wide, and an arrow at the end indicates the positive direction.

[0058] Figure 2 shows two horizontal threshold lines along the vertical axis as reference benchmarks for risk level classification. The first threshold line is located at the 0.70 position on the vertical axis, using a long dashed line style (20 pixels long, 10 pixels apart), with a line width of 2 pixels. The right side of this threshold line is labeled "High Risk Threshold (0.70)," indicating that when the risk score of any node exceeds 0.70, that node is marked as a high-risk node. The second threshold line is located at the 0.40 position on the vertical axis, using a short dashed line style (10 pixels long, 5 pixels apart), with a line width of 2 pixels. The right side of this threshold line is labeled "Medium Risk Threshold (0.40)," indicating that when the risk score of a node is between 0.40 and 0.70, that node is marked as a medium-risk node; when all risk scores do not exceed 0.40, that node is marked as a low-risk node. Both threshold lines extend horizontally throughout the entire horizontal axis, providing a unified judgment standard for risk scores across 24 consecutive data collection periods. Figure 2 shows four time-series evolution curves, corresponding to different risk type scores for four different nodes within the structured anomaly region. The first curve represents the delay risk score of node 1, drawn as a solid black line with a line width of 3 pixels. The risk score of this curve is approximately 0.35 at collection cycle number 0, and shows a slow upward trend as the collection cycle progresses. At collection cycle number 24, the risk score rises to approximately 0.50. Throughout the evolution process, the risk score remains below or slightly above the medium-risk threshold of 0.40, indicating that the delay risk of this node is in a transitional state from low to medium risk. The second curve represents the delay risk score of node 2, drawn using a long black dashed line (15 pixels long, 5 pixels apart), with a line width of 3 pixels. The risk score of this curve is approximately 0.45 at collection period number 0, and then shows a clear upward trend. Around collection period number 16, the risk score exceeds the high-risk threshold of 0.70, and reaches approximately 0.70 at collection period number 24. This indicates that the delay risk of this node evolved from a medium-risk state to a high-risk state during the monitoring period, requiring close attention and intervention from engineering management personnel. The third curve represents the cost overrun risk score for node 3, drawn using a black dashed line (25 pixels long, 10 pixels apart), with a line width of 3 pixels. The risk score of this curve is approximately 0.38 at collection period number 0, and then shows a stable upward trend. Around collection period number 20, the risk score is close to the medium risk threshold of 0.40 but does not significantly exceed it. At collection period number 24, the risk score is approximately 0.58, indicating that the cost overrun risk of this node gradually evolved from a low-risk state to a medium-risk state during the monitoring period.The fourth curve represents the quality risk score of node 4, drawn using a black dotted line (solid line segment length 10 pixels, interval 5 pixels, dot length 2 pixels, interval 5 pixels), with a line width of 3 pixels. The risk score of this curve is approximately 0.55 at collection period number 0, and shows significant fluctuations between collection periods number 8 and 16, reaching a peak of nearly 0.80. Subsequently, the risk score drops, reaching approximately 0.75 at collection period number 24. Throughout the evolution process, this curve crosses the high-risk threshold line multiple times, indicating that the quality risk of this node remains at a high risk level and fluctuates significantly during the monitoring period, making it a risk node requiring priority handling.

[0059] Figure 2 shows a legend in the upper left corner, illustrating the nodes and risk types represented by each curve. The legend uses a combination of text labels and line styles. Each legend item includes a line style example (approximately 100 pixels long) matching the corresponding curve, along with a text description. The four legend items, from top to bottom, are: "Node 1 - Delay Risk" (solid line), "Node 2 - Delay Risk" (long dashed line), "Node 3 - Cost Overrun Risk" (medium dashed line), and "Node 4 - Quality Risk" (dotted line). The legend items are vertically spaced 50 pixels apart, allowing readers to quickly identify the specific node and risk type corresponding to each curve. The four time-series evolution curves shown in Figure 2 demonstrate the differences in risk evolution at different nodes within the same structured anomaly region. The risk score curves for nodes 1 and 3 show a relatively smooth upward trend with a stable rate of increase, indicating that the risk accumulation process at these two nodes is continuous and predictable. Timely engineering interventions may prevent further escalation of the risk. The risk score curve for node 2 shows an accelerated increase after collection period number 12, indicating that the node was affected by external factors at a certain point in time, leading to a rapid accumulation of risk. The engineering events and changes in construction conditions corresponding to this point in time should be included in the subsequent causal analysis. The risk score curve for node 4 exhibits significant fluctuations, peaking between collection periods 8 and 16 before declining, but remaining generally within the high-risk range. This indicates that the quality risk of this node is significantly affected by short-term construction activities or environmental conditions, requiring continuous monitoring and comprehensive judgment based on actual site conditions. The multi-node risk score time-series evolution curves shown in Figure 2 enable dynamic tracking and visualization of the risk status of each node within the structured anomaly area. By plotting multiple risk score curves in the same coordinate system, engineering managers can intuitively compare the risk evolution trends of different nodes, identify nodes with rapid risk escalation rates and nodes with large risk fluctuations, thereby rationally allocating limited management resources and intervention measures. By setting unified high-risk and medium-risk thresholds, the risk level judgment criteria remain consistent throughout the monitoring period, avoiding inconsistencies caused by subjective human judgment. By employing a fixed 10-minute data collection cycle, the temporal resolution of risk scoring can capture short-term changes in the project status, while avoiding data spikes and noise interference caused by excessively short collection cycles. The temporal evolution analysis method shown in Figure 2 provides a complete technical chain for the project management system, from anomaly identification to risk prediction and dynamic monitoring, significantly improving the foresight and relevance of project risk management.

[0060] After the risk levels are classified, the unified engineering codes, corresponding risk categories, and risk scores of all nodes marked as high-risk and medium-risk are written into the data interface of the project management system. These are then mapped one-to-one with component nodes and process nodes in the engineering knowledge graph. This allows managers to directly view the risk level distribution of specific components and processes in the diagram structure view or project plan view within the visual interface. For example, in the project management interface, component nodes with delay risk scores greater than 0.70 are highlighted in red, while process nodes with cost overrun risk scores between 0.40 and 0.70 are highlighted in yellow. When a node is selected, its unified engineering code, the key indicator values ​​corresponding to the node's signal vector, and the specific values ​​of its delay risk score, cost overrun risk score, and quality risk score are displayed. Through this direct risk result output method based on unified engineering codes, engineering technicians can match risk data with construction drawings, contract terms, and monitoring records without additional matching steps, achieving a complete closed loop from identifying structured anomaly areas to presenting risk decision information.

[0061] In one optional implementation, the computational scope of the graph neural network model can be expanded from the structured anomaly region to adjacent nodes one layer away from the structured anomaly region on sequential edges, subordinate edges, or monitoring edges. This allows the graph neural network model to consider the states of adjacent nodes with direct construction or monitoring relationships when calculating the risk score of a node, thereby capturing anomaly propagation trends in space and process links. When it is necessary to improve the sensitivity to local risk propagation, the adjacent range participating in the graph convolution calculation can be expanded from one layer to two layers in the configuration. In implementation, a fixed number of layers is traversed without enabling adaptive layer adjustment, maintaining a fixed execution step. In another optional implementation, the number of group positions in the output layer can be adjusted according to the project type. For example, in a water conservancy project with quality and safety as its core, the number of quality-related numerical positions in the 8-length node feature vector can be expanded from two to four. By having more positions participate in the quality risk score calculation, the quality risk score becomes more sensitive to small adverse changes. The above adjustments are fixed in the system design phase and the corresponding relationships are clearly defined in the implementation documents to avoid inconsistencies caused by dynamic changes during operation.

[0062] Figure 3 shows the experimental curves illustrating the effect of the high-pass differential threshold on anomaly detection performance. In Figure 3, the horizontal axis represents the high-pass differential threshold θ, ranging from 0.00 to 0.35, and the vertical axis represents the detection rate, expressed as a percentage (%), ranging from 0 to 100. Figure 3 contains three curves to characterize the anomaly detection performance at different thresholds. The first curve, a thick solid line, represents the true anomaly detection rate (i.e., recall rate). This curve extends from the lower left to the upper right, indicating that the true anomaly detection rate decreases as the high-pass differential threshold θ increases. When the threshold θ is 0.00, the true anomaly detection rate is approximately 90%. When the threshold θ increases to 0.20, the true anomaly detection rate drops to 54.6%. When the threshold θ further increases to 0.35, the true anomaly detection rate drops to approximately 43%. The second curve, a dashed line, represents the false alarm rate. This curve extends from the upper left to the lower right, indicating that the false alarm rate decreases as the high-pass differential threshold θ increases. When the threshold θ is 0.00, the false alarm rate is close to 100%. When the threshold θ increases to 0.20, the false alarm rate drops to 8.3%. When the threshold θ further increases to 0.35, the false alarm rate drops to approximately 22%. The third curve, a dashed line, represents the F1 score, used to comprehensively evaluate the balance between the true anomaly detection rate and the false alarm rate. This curve shows a trend of first rising and then flattening out. When the threshold θ is 0.20, the F1 score reaches its peak of 69.2%, corresponding to a true anomaly detection rate of 54.6% and a false alarm rate of 8.3%, achieving the optimal balance between the two. In Figure 3, the position of threshold θ = 0.20 is marked with a vertical dashed line. The intersection of the three curves at this position is marked with a bold dot and the corresponding value. Figure 3 shows the experimental conditions at the bottom, which are for a large bridge project. The data collection period is 10 minutes. The anomaly identification effect of 4320 node samples within 24 hours is statistically analyzed. It shows that the threshold θ=0.20 is the optimal threshold parameter selected in this invention. This parameter setting can effectively balance the identification sensitivity and false alarm control capability of abnormal nodes in engineering implementation.

[0063] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for processing engineering project data, characterized in that, The process includes the following steps executed in sequence: Step 1, constructing an engineering knowledge graph based on unified engineering coding; Step 2, constructing a graph signal-enhanced engineering knowledge graph within the engineering knowledge graph: collecting multi-source data during the engineering project implementation process, normalizing the collected multi-source data, and mapping the normalized data to graph signals attached to nodes and associated edges in the engineering knowledge graph, thus obtaining a graph signal-enhanced engineering knowledge graph; Step 3, performing graph signal smoothing and graph signal high-pass filtering on the graph signal-enhanced engineering knowledge graph to identify structured anomalous regions in the graph; Step 4, based on the identified structured anomalous regions and the graph neural network model, generating joint risk prediction results for nodes within the structured anomalous regions, and classifying and outputting the risk levels of the corresponding nodes according to the joint risk prediction results; Step 3 specifically includes: for the current version of the graph signal-enhanced engineering knowledge graph... For each component node and process node, the following signal smoothing process is performed: From the set of direct adjacent nodes of the target node, select adjacent nodes connected by sequential edges, subordinate edges, and monitoring edges. If at least two adjacent nodes are selected, sum the signal vector of the target node element-wise with the signal vectors of all selected adjacent nodes, divide by the number of selected adjacent nodes plus 1, and obtain the smoothed signal vector. If only one adjacent node is selected, average the signal vectors of the target node and the selected adjacent node element-wise to obtain the smoothed signal vector. If no adjacent nodes are selected, retain the original signal vector as the smoothed signal vector. Subsequently, for each component node and process node, subtract the original signal vector from the corresponding smoothed signal vector element-wise and take the absolute value to obtain the high-pass differential signal vector. A high-pass differential signal vector is obtained when any element in the high-pass differential signal vector is greater than 0.At 20:00, this node is marked as an anomalous node; Step four specifically includes: establishing a graph neural network model, which includes a first graph convolutional layer, a second graph convolutional layer, and an output layer connected in sequence. The first graph convolutional layer expands the input node signal vector of length 12 into a node feature vector of length 16. The second graph convolutional layer compresses the node feature vector of length 16 into a node feature vector of length 8. The output layer maps the node feature vector of length 8 into three output values: delay risk score, cost overrun risk score, and quality risk score. In the first graph convolutional layer, for structured anomalous regions and those related to the structure... For structurally abnormal regions, adjacent component nodes and process nodes are processed by averaging the input signal vector of each node with the input signal vectors of its adjacent nodes connected by sequential and subordinate edges, resulting in an intermediate feature vector. Values ​​less than 0 in this intermediate feature vector are set to 0, and values ​​greater than 1 are set to 1. This processed intermediate feature vector is then used as the node feature vector of length 16 for the corresponding node. In the convolutional layer of Figure 2, for each node, its 16-length node feature vector is divided into 8 consecutive groups, each containing 2 elements. The average of the 2 elements in each group is then used to obtain a node feature vector of length 8.

2. The method as described in claim 1, characterized in that, Step one specifically includes: establishing a unique and unified engineering code for each component, process, contract clause, monitoring point, equipment, personnel, and event involved in the project. This unified engineering code is obtained by concatenating the project number, professional code, component or process category code, serial number field, and version field in a preset order. The field lengths are set as follows: 8 digits for the project number, 2 digits for the professional code, 4 digits for the component or process category code, 4 digits for the serial number field, and 2 digits for the version field. Based on the established unified engineering codes, a set of nodes in the project knowledge graph is generated. This set of nodes includes at least component nodes, process nodes, contract clause nodes, monitoring point nodes, equipment nodes, personnel nodes, and... Event nodes, each storing a corresponding unified project code and attribute fields related to the project status; establishing a set of associated edges between nodes based on the actual project relationships, this set of associated edges includes at least sequence edges representing process sequence constraints, subordinate edges representing the subordinate relationship between components and processes, configuration edges representing the relationship between processes and equipment / personnel configuration, monitoring edges representing the correspondence between processes and monitoring points, constraint edges representing the correspondence between processes and contract terms, and event edges representing the scope of event impact; enabling the project knowledge graph to form a connected topology, and any component node can be traced back to at least one contract term node and at least one monitoring point node through at most 5 associated edges.

3. The method as described in claim 1, characterized in that, Step two specifically includes: during the implementation of the project, with a fixed sampling period of 10 minutes, synchronously collect schedule deviation data, monitoring data and cost consumption data within the current period from the schedule management system, monitoring system and cost management system; perform unit unification and numerical normalization processing on the collected schedule deviation data, monitoring data and cost consumption data, so that all data are converted to a closed interval range of 0 to 1.

4. The method as described in claim 3, characterized in that, Step two further includes: for each sampling period, mapping the difference between the actual completion ratio and the planned completion ratio corresponding to the component node and the process node to a schedule deviation signal; mapping the monitoring measurement results corresponding to the monitoring point node to a monitoring signal; and mapping the comparison results of the current cost expenditure and the target cost corresponding to the process node and the component node to a cost consumption signal; forming a signal vector of length 12 by the schedule deviation signal, the monitoring signal, and the cost consumption signal in a fixed order and attaching it to the corresponding component node and process node; generating a signal vector of length 6 by the data directly related to the engineering relationship between nodes according to the same normalization rule and attaching it to the corresponding associated edge to obtain a graph signal enhancement engineering knowledge graph; and generating a complete snapshot of the graph signal enhancement engineering knowledge graph for each sampling period and updating the current version in real time.

5. The method as described in claim 4, characterized in that, Step 3 further includes: after completing the anomaly marking of all component nodes and process nodes, a continuous set of nodes in the graph that are connected by sequential edges and subordinate edges and are all marked as anomaly nodes will be identified as a candidate anomaly chain; when the number of anomaly nodes in the identified candidate anomaly chain is greater than or equal to 3 and the total planned time of the process corresponding to the span of the identified candidate anomaly chain is greater than or equal to 24 hours, the identified candidate anomaly chain and the monitoring point nodes directly connected to any node in the identified candidate anomaly chain through monitoring edges will be defined together as a structured anomaly region.

6. The method as described in claim 5, characterized in that, Step four further includes: In the output layer, for each node, the average of the 1st to 3rd elements in the node feature vector of length 8 is used as the delay risk score, the average of the 4th to 6th elements is used as the cost overrun risk score, and the average of the 7th to 8th elements is used as the quality risk score; when any risk score is greater than 0.70, the corresponding component node or process node is marked as a high-risk node; when any risk score is between 0.40 and 0.70, the corresponding node is marked as a medium-risk node; when all risk scores are less than or equal to 0.40, the corresponding node is marked as a low-risk node; and the unified engineering code, risk category, and risk score corresponding to the high-risk node and medium-risk node are output to the project management system interface.

7. A system for processing engineering project data to implement the method according to any one of claims 1 to 6, characterized in that, include: The knowledge graph construction module is used to build engineering knowledge graphs based on unified engineering coding. The graph signal enhancement module is used to collect multi-source data during the implementation of engineering projects, normalize the collected multi-source data, and map the normalized data into graph signals attached to nodes and associated edges in the engineering knowledge graph, thus obtaining a graph signal enhanced engineering knowledge graph. The abnormal region identification module is used to perform graph signal smoothing and graph signal high-pass filtering on the graph signal enhancement engineering knowledge graph to identify structured abnormal regions in the graph; the risk prediction module is used to generate joint risk prediction results for nodes within the structured abnormal regions based on the identified structured abnormal regions and the graph neural network model, and to classify and output the risk level of the corresponding nodes according to the joint risk prediction results.

Images