Construction engineering safety risk knowledge graph construction method based on multi-source data fusion

Abstract

This invention discloses a method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion. The method includes: constructing a multi-level ontology architecture covering risk sources, construction procedures, environmental elements, monitoring indicators, protective measures, and risk events; collecting and preprocessing safety specification texts, structured sensor monitoring data, and on-site images, and adding spatiotemporal tags; extracting entities and relationships under ontology constraints, constructing a risk evidence chain based on the sequential relationship of construction procedures and the spatial topological relationship between construction areas, work objects, and risk sources, enhancing the confidence of consistent evidence, and comparing the credibility and temporal proximity of conflicting evidence; constructing an initial knowledge graph and establishing spatiotemporal co-occurrence associations, locating local subgraphs based on newly added data and incrementally updating them to obtain a dynamically evolving target knowledge graph. This invention can improve the accuracy and timeliness of risk tracing and early warning in construction projects.

Description

Technical Field

[0001] This disclosure relates to generative artificial intelligence, more specifically to the fields of image analysis and visual recognition technology, and even more specifically, to a method for constructing a knowledge graph of construction engineering safety risks based on multi-source data fusion. Background Technology

[0002] Construction sites typically involve various safety risks during construction processes such as foundation pit excavation, formwork support, hoisting, working at heights, and temporary power supply. Existing safety risk management methods primarily identify, record, and analyze risks at construction sites using safety regulations, sensor monitoring data, and on-site video footage. Some methods also incorporate image recognition, sensor monitoring, or knowledge graph technologies to assist in safety risk assessment.

[0003] However, existing methods often process safety specifications, sensor monitoring data, and on-site images independently, lacking a unified semantic framework, making it difficult to integrate and correlate risk information from different sources. Furthermore, there is a lack of effective cross-modal correspondence between sensor anomalies, image recognition results, log records, and regulatory constraints, making it difficult to accurately depict the correlation links between risk sources, construction procedures, monitoring indicators, protective measures, and risk events.

[0004] Therefore, it is necessary to provide a method for constructing a knowledge graph of construction safety risks that can uniformly organize multi-source heterogeneous data, realize cross-modal risk association, and support local incremental updates. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for constructing a knowledge graph of construction engineering safety risks based on multi-source data fusion.

[0006] In a first aspect of the present invention, a method for constructing a knowledge graph of construction engineering safety risks based on multi-source data fusion is provided, the method comprising:

[0007] A multi-level ontology architecture for construction project safety risks is constructed. The multi-level ontology architecture includes risk sources, construction procedures, environmental elements, monitoring indicators, protective measures, and entity categories of risk events. The association relationships between each entity category, modal evidence binding rules, and spatiotemporal attribute slots are defined.

[0008] Collect multi-source heterogeneous data from the construction site and preprocess it. Add spatiotemporal tags containing the collection time and spatial location to the preprocessed data. The multi-source heterogeneous data includes safety specification text, structured sensor monitoring data, and on-site images.

[0009] Under the constraints of the multi-level ontology architecture, entities and relationships are extracted from multi-source heterogeneous data after spatiotemporal tags are added. Based on the sequential relationship between construction procedures and the spatial topological relationship between construction areas, work objects and risk sources, entities and relationships pointing to the same risk event are organized into risk evidence chains. The confidence of consistent evidence in the risk evidence chains is enhanced, and conflicting evidence is compared in terms of source credibility and temporal proximity to obtain a set of entities and relationships carrying modal source, association strength, confidence, and spatiotemporal tags.

[0010] The multi-level ontology architecture is instantiated based on the entity and relation set to construct an initial knowledge graph of construction project safety risks. Then, risk knowledge nodes in the initial knowledge graph are bound to their corresponding spatial locations and time sequences according to the spatiotemporal tags to establish spatiotemporal co-occurrence associations.

[0011] New construction site data is continuously added and processed as multi-source heterogeneous data to obtain new entities and relationships. The semantic state corresponding to the new entities and relationships is compared with the existing state in the initial knowledge graph. The affected local subgraphs are located according to the spatiotemporal tags, and the local subgraphs are incrementally updated to obtain the dynamically evolved target knowledge graph of construction safety risks.

[0012] Furthermore, the modal evidence binding rules are used to determine the corresponding data corroboration sources for each entity category; wherein, the risk source and the risk event are corroborated by at least one of the safety specification text, the structured sensor monitoring data, and the on-site images, the monitoring indicators are corroborated by the structured sensor monitoring data, and the protective measures are corroborated by the safety specification text; the spatiotemporal attribute slots are used to determine whether each entity category needs to be bound to the acquisition time, work area code, measurement point number, or component number.

[0013] Furthermore, the collection and preprocessing of multi-source heterogeneous data from the construction site includes:

[0014] The safety specification text is cleaned, segmented, and denoised; the structured sensor monitoring data is clock-aligned, missing values ​​are processed, and normalized; and the on-site images are subjected to frame extraction, deduplication, and distortion correction.

[0015] Furthermore, the extraction of entities and relations from the multi-source heterogeneous data after the addition of spatiotemporal labels includes:

[0016] Extract risk sources, construction procedures, protective measures, and regulatory constraints from the aforementioned safety specification text;

[0017] Based on the correspondence between monitoring indicators and risk sources in the multi-level ontology architecture, the structured sensing monitoring data is transformed into risk status semantic tags.

[0018] Identify defect-related risk sources, protection-deficiency-related risk sources, or non-compliance-related risk events from the on-site images.

[0019] Furthermore, based on the sequential relationship between construction procedures and the spatial topological relationship between the construction area, the work object, and the risk source, entities and relationships pointing to the same risk event are organized into a risk evidence chain, including:

[0020] Based on the spatiotemporal tags carried by each entity and relationship, candidate entities and relationships that occur in the same construction area and whose time intervals meet preset conditions are filtered.

[0021] Based on the multi-level ontology architecture, the construction procedures, work objects, environmental elements, monitoring indicators, risk sources, and protective measures corresponding to the candidate entities are determined respectively.

[0022] Based on the sequential relationship between construction procedures and the spatial topological relationship between work objects, environmental elements, monitoring indicators and risk sources, the pointing relationship of the candidate entities and candidate relationships to the same risk event is determined.

[0023] Based on the aforementioned pointing relationship, candidate entities and candidate relationships that respectively represent risk causes, risk states, and risk outcomes are connected in the order of risk formation to obtain the risk evidence chain.

[0024] Furthermore, the confidence enhancement of consistent evidence in the risk evidence chain and the comparison of source credibility and temporal proximity of conflicting evidence include:

[0025] Different modal data in the risk evidence chain that point to the same entity attribute or the same relationship and have the same representation result are identified as consistent evidence. Based on the number of modalities of the consistent evidence and the credibility of the modal source, the confidence of the corresponding entity or relationship is increased.

[0026] Different modal data in the risk evidence chain that point to the same entity attribute or the same relationship but have different representation results are identified as conflicting evidence.

[0027] The modal source credibility of each conflicting piece of evidence and the temporal proximity between the time of collection of each conflicting piece of evidence and the time of occurrence of the risk event are determined respectively.

[0028] Based on the modality source credibility and the temporal proximity, retained evidence is determined from the conflicting evidence, and the unretained conflicting evidence and their conflict states are associated with the retained evidence.

[0029] Furthermore, the initial knowledge graph for constructing safety risks in construction projects includes:

[0030] The entities in the entity and relation set are used as graph nodes, the relations in the entity and relation set are used as graph edges, and the entity category, modality source, association strength, confidence level, risk status, and spatiotemporal label are used as attribute information of the corresponding graph node or graph edge. All of these are stored in the graph database to construct the initial knowledge graph.

[0031] Furthermore, the establishment of spatiotemporal co-occurrence correlation includes:

[0032] Based on the spatiotemporal tags, determine whether different risk knowledge nodes are in the same or adjacent spatial ranges, and determine whether there are overlapping time windows between the time series corresponding to different risk knowledge nodes.

[0033] When different risk knowledge nodes simultaneously satisfy spatial range association and time window overlap, a corresponding spatiotemporal co-occurrence association is established in the initial knowledge graph.

[0034] Furthermore, after obtaining the dynamically evolved target knowledge graph of construction project safety risks, it also includes:

[0035] Based on the target knowledge graph, the safety risks of construction projects are assessed, and the risk chain is traced along the relationship between environmental elements, construction procedures, risk sources and risk events.

[0036] Based on the construction procedures, query the associated risk sources, monitoring indicators, and protective measures;

[0037] When the risk source or risk event reaches a preset risk state after cross-verification with multimodal evidence, dynamic early warning information is output.

[0038] Furthermore, the incremental update of the local subgraph includes:

[0039] The update range is determined based on the spatiotemporal tags carried by the newly added entities and relationships;

[0040] Within the scope of the update, the risk status, confidence level, and association strength corresponding to the newly added entities and relationships are matched with the existing nodes and existing relationships in the local subgraph;

[0041] When the matching result indicates a new risk, add the corresponding node or relationship to the local subgraph.

[0042] When the matching result indicates a change in risk status, update the attribute information of the corresponding node or relationship;

[0043] When the matching result indicates that the original risk has failed, add historical status tags to the corresponding node or relationship and retain its spatiotemporal traceability information.

[0044] Therefore, this invention constructs a multi-level ontology architecture for construction project safety risks, uniformly defining entity categories such as risk sources, construction procedures, environmental elements, monitoring indicators, protective measures, and risk events, and setting the relationships between entities, modal evidence binding rules, and spatiotemporal attribute slots. Based on this, it collects multi-source heterogeneous data such as safety specification texts, structured sensor monitoring data, and on-site images, preprocesses them, and adds spatiotemporal tags composed of the collection time and spatial location. Under the constraints of the multi-level ontology architecture, it extracts entities and relationships from the multi-source heterogeneous data after adding spatiotemporal tags, maps the risk information to a unified semantic space, and further processes it. Cross-modal association and multimodal evidence cross-validation are performed to obtain a set of entities and relations carrying modality source, association strength, confidence level, and spatiotemporal labels. Subsequently, the multi-level ontology architecture is instantiated based on the set of entities and relations to construct an initial knowledge graph of construction safety risks. Risk knowledge nodes in the graph are bound to spatial location and time series to establish spatiotemporal co-occurrence associations. At the same time, newly added construction site data is continuously added, and the semantic state corresponding to the newly added entities and relations is compared with the existing state in the initial knowledge graph. The affected local subgraphs are located according to the spatiotemporal labels and incrementally updated to obtain a dynamically evolved target knowledge graph of construction safety risks.

[0045] Through the above methods, this invention can unify scattered data such as safety specification texts, sensor monitoring data, and on-site images into a single semantic framework, avoiding the problems of semantic inconsistency and difficulty in associating risk information from different sources; it can perform cross-modal fusion and evidence verification of sensor anomalies, image recognition results, log records, and specification constraints around the same risk source, improving the accuracy of risk identification and risk relationship construction, and reducing false alarms and missed alarms caused by a single data source; it can establish the correspondence between risk knowledge nodes and specific work areas, components, measuring points, and construction stages based on spatiotemporal tags, forming a queryable and traceable risk link; at the same time, after the access of new on-site data, only the affected local subgraphs are incrementally updated, avoiding the computational overhead of full reconstruction, so that the construction project safety risk knowledge graph can dynamically evolve with the changes in the construction site status, thereby improving the timeliness and adaptability of construction safety risk assessment, source tracing analysis, and dynamic early warning. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the construction method for building a knowledge graph of construction safety risks based on multi-source data fusion, as described in this invention.

[0047] Figure 2This is a schematic diagram of the node architecture of the target knowledge graph of construction safety risks in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the electronic device structure for implementing the method of the present invention in an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please see Figure 1 In a first aspect, the present invention provides a method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion, the method comprising:

[0051] S101. Construct a multi-level ontology architecture for construction project safety risks. The multi-level ontology architecture includes risk sources, construction procedures, environmental elements, monitoring indicators, protective measures, and entity categories of risk events. Define the relationships between entity categories, modal evidence binding rules, and spatiotemporal attribute slots.

[0052] In this embodiment, the multi-level ontology architecture refers to a semantic organization framework used to uniformly describe objects, object attributes, and relationships between objects related to construction project safety risks. This multi-level ontology architecture is not a simple listing of risk names, but rather, according to the inherent logic of construction project safety management, it divides information related to safety risks at the construction site into entity categories at different levels, and further constrains the possible relationship types between these entity categories. This ensures that subsequent multi-source heterogeneous data can be expressed according to a unified semantic standard during entity extraction, relationship extraction, and knowledge graph construction.

[0053] Specifically, the multi-level ontology architecture includes entity categories such as risk sources, construction procedures, environmental factors, monitoring indicators, protective measures, and risk events. Risk sources represent objects or states on the construction site that may cause safety accidents or risks, such as foundation pit deformation, formwork support instability, lack of edge protection, and abnormal tower crane posture. Construction procedures represent various operational stages in the construction process, such as earthwork excavation, formwork erection, concrete pouring, hoisting, and working at heights. Environmental factors represent external environmental conditions that affect construction safety, such as wind speed, rainfall, temperature, and humidity. Monitoring indicators represent quantitative or state-based indicators that reflect the risk status, such as settlement values, displacement values, verticality deviation, and electrical load. Protective measures represent safety control methods adopted for risk sources, such as support reinforcement, installation of guardrails, suspension of work, and recalculation. Risk events represent potential safety events that may occur after the risk develops further, such as collapse, electric shock, falls from height, and mechanical injuries.

[0054] In this embodiment, the association relationship refers to the semantic connection relationship formed between different entity categories in construction safety risk management, used to represent the influence, characterization, control, or evolution of one entity on another. Association relationships include, but are not limited to, relationship types such as "construction procedure—accompanying—risk source," "environmental element—inducing—risk source," "monitoring indicator—characterizing—risk source," "protective measures—control—risk source," and "risk source—evolution into—risk event." By defining the above association relationships, the knowledge graph subsequently constructed can not only store individual risk nodes but also express the hierarchical chain-like association between construction procedures, risk inducements, monitoring evidence, protective measures, and risk consequences.

[0055] In this embodiment, modal evidence binding rules refer to pre-configured rules for different entity categories that specify which data modalities can be used to identify, corroborate, or verify them. For example, monitoring indicators are mainly corroborated by structured sensor monitoring data, defect-type risk sources can be corroborated by on-site images, and regulatory constraints and protective measures can be corroborated by safety specification texts. Spatiotemporal attribute slots refer to attribute fields reserved for entity categories that need to be bound to on-site location and occurrence time. Their content may include acquisition time, work area code, measuring point number, component number, spatial coordinates, or construction stage, etc.

[0056] In some embodiments, safety specification texts may include construction logs, compilation basis, construction organization design, safety-specific construction plans, safety technical briefing records, hazard identification and risk classification and control data, hidden danger investigation and rectification records, supervision notices, emergency plans, and applicable laws, regulations, standards, and technical procedures related to safety production in construction projects. Specifically, safety specification texts can be used to record or reflect information such as construction procedures, operating conditions, types of risk sources, protective measures, monitoring requirements, limit thresholds, abnormal handling methods, and responsible parties, thereby providing textual basis for extracting entities and relationships of risk sources, construction procedures, protective measures, and regulatory constraints.

[0057] When constructing a multi-level ontology architecture, the following steps are taken: First, the set of entity categories is determined based on the construction project safety management objects, and each entity category is hierarchically divided. Second, based on the formation mechanism of construction site risks and safety management business rules, the permissible associations and their directions between entity categories are determined. Third, based on the specific manifestations of each entity category in different data sources, corresponding modal evidence binding rules are configured. Finally, based on whether the entity needs to participate in spatiotemporal positioning, risk tracing, or dynamic updates, corresponding spatiotemporal attribute slots are configured. The resulting multi-level ontology architecture can serve as a unified constraint basis for subsequent semantic extraction from multi-source heterogeneous data, cross-modal association, knowledge graph instantiation, and incremental graph updates.

[0058] In some embodiments, modal evidence binding rules are used to determine the corresponding data corroboration sources for each entity category; wherein, risk sources and risk events are corroborated by at least one of the following data: safety specification text, structured sensor monitoring data, and on-site images; monitoring indicators are corroborated by structured sensor monitoring data; and protective measures are corroborated by safety specification text; the spatiotemporal attribute slots are used to determine whether each entity category needs to be bound to the acquisition time, work area code, measurement point number, or component number.

[0059] In some embodiments, modal evidence binding rules are used to determine the corresponding data corroboration sources for each entity category. Data corroboration sources refer to the data sources used to prove the existence, abnormality, and risk status of a certain entity category during subsequent entity identification, risk confirmation, relationship extraction, or risk status updates. Since different safety risk objects at construction sites often exhibit different forms in different data, this embodiment pre-sets modal evidence binding rules for each entity category. This allows subsequent multi-source heterogeneous data to automatically match the corresponding data source according to the entity category when entering the semantic extraction and knowledge graph construction process, thereby avoiding inconsistencies in risk identification results caused by the unregulated mixing of different data.

[0060] Specifically, risk sources and risk events can be corroborated by at least one of the following data sources: safety specification texts, structured sensor monitoring data, and on-site video footage. Safety specification texts can provide normative descriptions and judgment criteria for risk sources or risk events, such as the types of risks that should be prioritized for control under a certain type of construction procedure. Structured sensor monitoring data can provide quantitative evidence of whether a risk source is in an abnormal state, such as excessive foundation pit displacement, abnormal tower crane posture, or abnormal temporary power load. On-site video footage can provide visual evidence of risk sources or risk events, such as inadequate protection, structural cracks, abnormal material stacking, or violations of operational regulations. Therefore, risk sources and risk events no longer rely on a single data source for judgment, but can be comprehensively corroborated based on the actual availability of different modal data.

[0061] In some embodiments, monitoring indicators are supported by structured sensor monitoring data. These monitoring indicators characterize safety status parameters at the construction site that can be continuously or periodically collected, such as settlement values, displacement values, verticality deviation, wind speed, temperature and humidity, current, voltage, or equipment attitude data. Since these indicators typically have clearly defined data acquisition equipment, acquisition time, and numerical records, using structured sensor monitoring data as the primary source of support ensures the stability of the monitoring indicator source and the quantifiability of the status assessment.

[0062] In some embodiments, protective measures are supported by safety specification texts. Protective measures represent preventative, control, or remedial actions taken against a risk source, such as setting up edge protection, reinforcement, suspending operations, recalculating, and establishing warning zones. Safety specification texts can be used to determine the standard protective requirements corresponding to a particular type of risk source. By linking protective measures with safety specification texts, a relationship between "protective measures—control—risk source" can be established in the subsequent knowledge graph.

[0063] In some embodiments, spatiotemporal attribute slots are used to determine whether each entity category needs to be bound to a data acquisition time, work area code, measuring point number, or component number. The data acquisition time indicates the time when data is generated or a risk status is identified; the work area code indicates the construction area corresponding to the risk information; the measuring point number indicates the monitoring point corresponding to the structured sensor monitoring data; and the component number indicates the specific engineering component corresponding to the risk information. For entity categories requiring on-site location, time tracking, or dynamic updates, such as risk sources, monitoring indicators, risk events, and construction procedures, this embodiment configures corresponding spatiotemporal attribute slots. For entity categories primarily serving a regulatory or general control function, such as certain protective measures or regulatory requirements, spatiotemporal attribute slots can be selectively configured based on whether they correspond to a specific work area, measuring point, or component.

[0064] When configuring modal evidence binding rules and spatiotemporal attribute slots, the system first determines which data sources can identify or corroborate an entity based on its risk expression method. Then, based on whether the entity needs to participate in subsequent cross-modal association, spatiotemporal co-occurrence analysis, and local subgraph incremental updates, it determines whether it needs to be bound to the acquisition time, work area code, measurement point number, or component number. Thus, when each entity enters the subsequent data extraction and graph instantiation process, it has both clear data source evidence and carries the necessary temporal and spatial attributes, providing a foundation for subsequent risk association analysis, risk link tracing, and dynamic knowledge graph updates.

[0065] S102. Collect multi-source heterogeneous data from the construction site and preprocess it, then attach spatiotemporal labels containing the collection time and spatial location to the preprocessed data.

[0066] The multi-source heterogeneous data includes safety specification texts, structured sensor monitoring data, and on-site images.

[0067] In some embodiments, preprocessing multi-source heterogeneous data may include:

[0068] The safety specification text is cleaned, segmented, and denoised; the structured sensor monitoring data is clock-aligned, missing values ​​are handled, and normalized; and the on-site images are processed by frame extraction, deduplication, and distortion correction.

[0069] In some embodiments, modal evidence binding rules are used to determine the corresponding data corroboration sources for each entity category. Data corroboration sources refer to the data sources used to prove the existence, abnormality, and risk status of a certain entity category during subsequent entity identification, risk confirmation, relationship extraction, or risk status updates. Since different safety risk objects at construction sites often exhibit different forms in different data, this embodiment pre-sets modal evidence binding rules for each entity category, enabling subsequent multi-source heterogeneous data to automatically match the corresponding data source according to the entity category when entering the semantic extraction and knowledge graph construction process. This avoids inconsistencies in risk identification results caused by the unregulated mixing of different data.

[0070] Specifically, risk sources and risk events can be corroborated by at least one of the following data sources: safety specification texts, structured sensor monitoring data, and on-site video footage. Safety specification texts can provide normative descriptions and judgment criteria for risk sources or risk events, such as the types of risks that should be prioritized for control under a certain type of construction procedure. Structured sensor monitoring data can provide quantitative evidence of whether a risk source is in an abnormal state, such as excessive foundation pit displacement, abnormal tower crane posture, or abnormal temporary power load. On-site video footage can provide visual evidence of risk sources or risk events, such as inadequate protection, structural cracks, abnormal material stacking, or violations of operational regulations. Therefore, risk sources and risk events no longer rely on a single data source for judgment, but can be comprehensively corroborated based on the actual availability of different modal data.

[0071] In some embodiments, monitoring indicators are supported by structured sensor monitoring data. These monitoring indicators characterize safety status parameters at the construction site that can be continuously or periodically collected, such as settlement values, displacement values, verticality deviation, wind speed, temperature and humidity, current, voltage, or equipment attitude data. Since these indicators typically have clearly defined data acquisition equipment, acquisition time, and numerical records, using structured sensor monitoring data as their primary source of support ensures the stability of the monitoring indicator source and the quantifiability of the status assessment.

[0072] In some embodiments, protective measures are supported by safety specification texts. Protective measures represent preventative, control, or remedial actions taken against a risk source, such as setting up edge protection, reinforcement, suspending operations, recalculating, and establishing warning zones. Safety specification texts can be used to determine the standard protective requirements corresponding to a particular type of risk source. By linking protective measures with safety specification texts, a relationship between "protective measures—control—risk source" can be established in the subsequent knowledge graph.

[0073] In some embodiments, spatiotemporal attribute slots are used to determine whether each entity category needs to be bound to a data acquisition time, work area code, measuring point number, or component number. The data acquisition time indicates the time when data is generated or a risk status is identified; the work area code indicates the construction area corresponding to the risk information; the measuring point number indicates the monitoring point corresponding to the structured sensor monitoring data; and the component number indicates the specific engineering component corresponding to the risk information. For entity categories requiring on-site location, time tracking, or dynamic updates, such as risk sources, monitoring indicators, risk events, and construction procedures, this embodiment configures corresponding spatiotemporal attribute slots. For entity categories primarily serving a regulatory or general control function, such as certain protective measures or regulatory requirements, spatiotemporal attribute slots can be selectively configured based on whether they correspond to a specific work area, measuring point, or component.

[0074] When configuring modal evidence binding rules and spatiotemporal attribute slots, we can first determine which data sources can identify or corroborate an entity category based on its risk expression method. Then, we can determine whether the entity category needs to be bound to the acquisition time, work area code, measurement point number, or component number, depending on whether it needs to participate in subsequent cross-modal association, spatiotemporal co-occurrence analysis, and local subgraph incremental updates. Thus, when each entity category enters the subsequent data extraction and graph instantiation process, it has both clear data source evidence and carries the necessary time and space attributes, providing a foundation for subsequent risk association analysis, risk link tracing, and dynamic updates of the knowledge graph.

[0075] S103. Under the constraints of a multi-level ontology architecture, entities and relationships are extracted from multi-source heterogeneous data after the addition of spatiotemporal labels. Based on the sequential relationship between construction procedures and the spatial topological relationship between construction areas, work objects and risk sources, entities and relationships pointing to the same risk event are organized into a risk evidence chain. The confidence of consistent evidence in the risk evidence chain is enhanced, and the source credibility and temporal proximity of conflicting evidence are compared to obtain a set of entities and relationships carrying modal source, association strength, confidence and spatiotemporal labels.

[0076] In some embodiments, entity and relation extraction from multi-source heterogeneous data with added spatiotemporal labels may include:

[0077] Extract risk sources, construction procedures, protective measures, and regulatory constraints from safety specification texts;

[0078] Based on the correspondence between monitoring indicators and risk sources in the multi-level ontology architecture, the structured sensing monitoring data is transformed into risk status semantic labels.

[0079] Identify risk sources such as defects, lack of protection, or violations of regulations from on-site images.

[0080] In some embodiments, entity and relation extraction from multi-source heterogeneous data with added spatiotemporal labels refers to, under the constraints of the aforementioned multi-level ontology architecture, converting information related to construction safety risks from safety specification texts, structured sensor monitoring data, and on-site images into entities, relations, and their attribute information that can be identified and stored by a knowledge graph. Entities may include objects such as risk sources, construction procedures, environmental elements, monitoring indicators, protective measures, and risk events; relations may include the accompanying relationship between construction procedures and risk sources, the representational relationship between monitoring indicators and risk sources, the inducing relationship between environmental elements and risk sources, the control relationship between protective measures and risk sources, and the evolutionary relationship between risk sources and risk events.

[0081] In this embodiment, the multi-source heterogeneous data with added spatiotemporal tags refers to data that has been preprocessed and bound with the collection time and spatial location. Since each data item carries corresponding time and spatial information, the extracted entities and relationships can inherit the spatiotemporal tags of the corresponding data during the entity and relationship extraction process, thereby enabling the subsequently formed risk knowledge nodes to correspond one-to-one with specific construction areas, measuring points, components, or construction stages.

[0082] In this embodiment, risk state semantic labels refer to the label information formed after converting the numerical states in structured sensor monitoring data into risk semantics. Structured sensor monitoring data itself is usually represented by values ​​such as displacement, settlement, verticality deviation, wind speed, current, or voltage, which are difficult to use directly as risk representations in a knowledge graph. Therefore, it is necessary to convert the structured sensor monitoring data into risk state semantic labels such as "increased risk of foundation pit deformation," "structural instability risk," and "abnormal risk of temporary power supply," which can be recognized by the knowledge graph, based on the correspondence between monitoring indicators and risk sources in the multi-level ontology architecture.

[0083] In this embodiment, defect-related risk sources refer to structural defects, material defects, or abnormal construction quality that can be identified through on-site images, such as cracks, deformation, water accumulation, and abnormal material stacking; protection-related risk sources refer to risk sources caused by inadequate on-site safety protection measures, such as lack of edge protection, unsealed openings, and lack of safety warnings; and non-compliance risk events refer to personnel or equipment violations that can be identified through on-site images, such as not wearing safety helmets, not wearing safety belts when working at heights, smoking in violation of regulations, and personnel remaining in hoisting areas.

[0084] Specifically, when extracting entities and relationships from safety specification texts, the text can first be segmented and semantically identified to recognize the risk sources, construction procedures, protective measures, and regulatory constraints involved. Regulatory constraints represent the requirements imposed on a specific construction procedure, risk source, or protective measure within the safety specification text. For example, when a safety specification text contains statements such as "a certain type of construction procedure should have corresponding protective measures," "a certain type of risk source should be monitored," or "a certain type of abnormal state should require work to cease," the corresponding construction procedure entities, risk source entities, protective measure entities, and regulatory constraints can be extracted respectively.

[0085] When extracting entities and relationships from structured sensor monitoring data, the types of monitoring indicators corresponding to the structured sensor monitoring data can be identified first. Then, based on the predefined correspondence between monitoring indicators and risk sources in the multi-level ontology architecture, the numerical state of the monitoring indicators can be converted into corresponding risk state semantic tags. For example, when monitoring indicators such as foundation pit displacement, settlement, or support structure deformation reach abnormal states, they can be converted into risk state semantic tags related to foundation pit deformation, slope instability, or support structure risks; when tower crane posture, wind speed, or equipment operating parameters are abnormal, they can be converted into risk state semantic tags related to lifting or equipment operation safety. This allows structured numerical data to participate in the semantic association of subsequent knowledge graphs.

[0086] When extracting entities and relationships from on-site images, target and scene recognition can be performed on images or video frames captured by fixed cameras, drones, or surveillance cameras to identify defect-type risk sources, protection-deficiency-type risk sources, or non-compliant operation-type risk events. The recognition results can be categorized according to entity types in a multi-level ontology architecture and bound to the acquisition time and spatial location carried by the image data. For example, when structural cracks, missing guardrails, or personnel violations are identified in on-site images of the same work area, corresponding risk source entities or risk event entities can be generated respectively, and their image source information can be retained.

[0087] Therefore, by extracting entities and relationships from safety specification texts, structured sensor monitoring data, and on-site images, risk information that was originally scattered across different data sources can be transformed into unified graph construction data. At the same time, since the extraction process is always constrained by a multi-level ontology architecture, the resulting entity categories, relationship types, and risk state expressions are consistent, thus providing a foundation for subsequent unified semantic space mapping, cross-modal association, multimodal evidence cross-validation, and knowledge graph instantiation.

[0088] In some embodiments, organizing entities and relationships pointing to the same risk event into a risk evidence chain based on the sequential relationship between construction procedures and the spatial topological relationship between construction areas, work objects, and risk sources may include the following steps:

[0089] First, based on the spatiotemporal tags carried by each entity and relationship, candidate entities and relationships that occur in the same construction area and whose time intervals meet preset conditions are selected.

[0090] Specifically, the spatiotemporal label may include at least one of the following: data acquisition time, corresponding time interval, work area code, floor number, work surface number, component number, measuring point number, and spatial coordinates. Entities and relationships with the same work area code, the same work surface number, or overlapping corresponding spatial ranges are identified as being in the same construction area; alternatively, when two entities correspond to adjacent construction areas, and the work activities in one construction area can affect the other construction area, the two entities can also be identified as meeting the spatial screening criteria.

[0091] Preset conditions are used to limit the temporal correlation range of different risk information. These preset conditions can be determined based on the duration of the construction process, the sampling period of the monitoring data, and the time it takes for a risk factor to take effect from its occurrence to the formation of a risk outcome. For example, for tower crane operation risks, the same hoisting operation cycle can be used as the time screening range; for foundation pit deformation risks, multiple consecutive monitoring cycles can be used as the time screening range; and for concrete pouring quality risks, the time interval from the start of pouring to the end of the curing stage can be used as the time screening range.

[0092] Based on the aforementioned spatial and temporal scope, entities and relationships that may jointly describe the same risk event are selected from the set of entities and relationships. For example, within the same foundation pit construction area, candidate relationships such as "continuous rainfall," "earthwork excavation," "support structure," "increased horizontal displacement," "cracks in the support structure," "foundation pit instability risk," and "inducing," "acting on," "characterizing," and "causing" among the aforementioned entities can be selected.

[0093] Secondly, based on the multi-level ontology architecture, the construction procedures, work objects, environmental elements, monitoring indicators, risk sources, and protective measures corresponding to the candidate entities are determined.

[0094] Specifically, the entity names, attribute information, and semantic features of candidate entities are matched with entity categories in a multi-level ontology architecture to determine the semantic role of each candidate entity in the risk formation process. Here, construction procedures represent ongoing or completed construction activities, such as earthwork excavation, formwork erection, concrete pouring, and hoisting; work objects represent personnel, equipment, materials, components, or temporary facilities directly affected by the construction procedures; environmental elements represent wind speed, rainfall, temperature, illuminance, dust concentration, or geological conditions that may affect construction safety; monitoring indicators represent displacement, tilt angle, stress, settlement, vibration, or load reflecting the state of the work objects; risk sources represent unsafe conditions or unsafe behaviors that may lead to safety accidents; and protective measures represent monitoring, isolation, reinforcement, limiting, or personnel protection measures implemented for risk sources.

[0095] When multiple categories match for the same candidate entity, disambiguation can be performed by combining the data source and contextual relationships of the candidate entity. For example, "support" in safety specification text may refer to support installation in construction procedures, support components in on-site images, and protective measures in safety specification text. In this case, its corresponding ontology category can be determined based on its associated action entities, spatial location, and upstream and downstream relationships.

[0096] Next, based on the sequential relationship between construction procedures and the spatial topological relationship between the work object, environmental elements, monitoring indicators and risk sources, the pointing relationship of the candidate entities and candidate relationships to the same risk event is determined.

[0097] Specifically, the sequential relationship between construction procedures can be determined based on the actual start and end times in the construction schedule and procedure flow template, including preceding, current, and subsequent procedures. For example, formwork erection occurs before concrete pouring, and foundation pit excavation occurs before the stress changes in the support structure. If the construction procedures corresponding to the candidate entities are continuous or have a direct impact on the construction process, it indicates that the corresponding risk information has process-related characteristics.

[0098] Spatial topological relationships are used to determine whether work objects, environmental factors, and monitoring indicators can spatially contribute to the same risk source. These spatial topological relationships can include being located at, containing, adjacent to, intersecting with, covering, near, and having overlapping areas of influence. For example, if a wind speed monitoring point is located within the operating area of ​​a tower crane, and the slewing range of the tower crane's boom intersects with the personnel passage area, then the wind speed monitoring results, the tower crane's operating status, and the personnel's location can all point to a crane collision risk. Similarly, if a foundation pit displacement monitoring point is located on the support structure, and the cracks identified in the image are also located on the same support component, then the increased displacement entity and the crack entity can both point to a risk of support structure instability.

[0099] When determining the pointing relationship, we can first determine whether the candidate entities correspond to the same or related work objects, then determine whether environmental factors and monitoring indicators can affect or characterize the risk status of the work object, and finally determine whether the candidate relationship can form a continuous semantic path from construction activities and risk factors to risk events. When the candidate entities meet the association conditions in terms of process sequence, spatial scope, and semantic relationship, it is determined that the candidate entities and candidate relationships point to the same risk event; otherwise, they are classified into different risk events or marked as information to be verified.

[0100] Finally, based on the aforementioned pointing relationship, the candidate entities and candidate relationships that respectively represent the risk cause, risk state, and risk outcome are connected in the order of risk formation to obtain the risk evidence chain.

[0101] Among them, risk inducement is used to represent construction activities, environmental changes, unsafe behaviors, or lack of protective measures that cause risk to occur; risk state is used to represent the intermediate state exhibited by the risk source during its formation or development; and risk outcome is used to represent the possible risk event. Following the chronological and logical order of risk formation, the risk inducement entity, risk state entity, and risk outcome entity are connected sequentially, and candidate relationships are used as directed connections between adjacent entities.

[0102] For example, in the scenario of foundation pit construction, the following chain of evidence for risk can be formed:

[0103] "Continuous rainfall" is connected to "increased soil moisture content" through the "change" relationship; "increased soil moisture content" is connected to "decreased slope stability" through the "decrease" relationship; "decreased slope stability" is connected to "increased support structure displacement" through the "cause" relationship; and "increased support structure displacement" is connected to "risk of foundation pit instability" through the "characterize" relationship.

[0104] Among them, "continuous rainfall" and "earthwork excavation" can be considered as risk triggers, "increased soil moisture content," "decreased slope stability," and "increased displacement of the support structure" can be considered as risk states, and "risk of foundation pit instability" can be considered as a risk outcome. Structured sensor monitoring data provides numerical evidence of "increased displacement," while on-site images provide visual evidence of "cracks in the support structure."

[0105] Each risk evidence chain can be associated with a risk event identifier and record the modal origin, spatiotemporal label, evidence chain position, confidence level, and conflict state of each entity and relationship in the chain. The evidence chain position is used to indicate whether the corresponding entity or relationship belongs to the risk initiation stage, risk state stage, or risk outcome stage, so as to enhance the confidence of consistent evidence, adjudicate conflicting evidence, and trace and warn of the risk formation process in the knowledge graph.

[0106] In some embodiments, enhancing the confidence of consistent evidence in a risk evidence chain and comparing the source credibility and temporal proximity of conflicting evidence may include the following steps:

[0107] First, different modal data in the risk evidence chain that point to the same entity attribute or the same relationship and have the same representation result are identified as consistent evidence. Then, based on the number of modalities of consistent evidence and the credibility of the modal source, the confidence of the corresponding entity or relationship is increased.

[0108] Specifically, based on the entity identifier, attribute identifier, relationship type, and spatiotemporal label corresponding to each piece of evidence, it can be determined whether different modal data pertain to the same object to be verified. For example, structured sensor monitoring data indicates that the horizontal displacement of a foundation pit support component exceeds a preset threshold, on-site imagery identifies a significant tilting of the same support component, and safety regulations record abnormal deformation in the area where the support component is located. When the above evidence corresponds to each other in terms of construction area, component number, and time of occurrence, and all indicate that the support component is in an abnormal deformation state, they are identified as different modal evidence pointing to the same entity attribute.

[0109] For example, safety regulations may state that "continuous rainfall may induce slope instability," while structured sensor monitoring data shows that soil moisture content continuously increases during rainfall, and on-site images identify localized slippage on the slope surface. When the above evidence collectively supports the correlation that "continuous rainfall induces a decrease in slope stability," it can be identified as consistent evidence pointing to the same correlation.

[0110] Modal source reliability characterizes the reliability of a data source in reflecting the attributes or relationships of a target entity. Different modal source reliability can be determined based on the operating status of the data acquisition equipment, detection accuracy, data completeness, data generation method, and historical verification results. For example, data generated by a calibrated and normally functioning displacement sensor can have high source reliability; on-site images acquired by a fixed camera with no obstructions in the target area can have high source reliability; safety specification texts confirmed by the on-site supervisor can have high source reliability; unconfirmed free text records or on-site images with obstructions can have relatively low source reliability.

[0111] After establishing consistent evidence, the initial confidence level of an entity or relationship within a single modality can be used as a foundation. This confidence level can then be enhanced based on the number of valid modalities supporting the entity or relationship and the credibility of each modal source. The number of valid modalities can be the number of independent data modalities, rather than the number of duplicate records from the same data source. For example, multiple similar frames continuously captured by the same camera can be considered evidence of the same modality, while sensor data and on-site images can be considered evidence of different modalities.

[0112] When the number of valid modalities supporting the same entity or relationship increases, and the credibility of the corresponding modal source is high, the confidence level of that entity or relationship can be improved. For data with low source credibility or that highly overlaps with existing evidence, their impact on the confidence enhancement results can be reduced, thus avoiding an unreasonable amplification of the credibility of an entity or relationship by a large amount of duplicate data.

[0113] For example, when a displacement sensor detects that the horizontal displacement of a support component exceeds the limit, it can generate entity attribute information of "support component - status - displacement exceeding the limit". If the on-site image further identifies that the support component is tilted and the safety specification text records that the same area has abnormal deformation, then the confidence level of the "support component displacement exceeding the limit" entity and its representation relationship with "foundation pit instability risk" can be enhanced.

[0114] Secondly, different modal data in the risk evidence chain that point to the same entity attribute or the same relationship but have different representation results are identified as conflicting evidence.

[0115] Specifically, when different modal data correspond to the same construction area, the same work object, and the same or similar time period, but give different status judgments for the same entity attribute, or give different relationship judgments for the same association, the corresponding data can be identified as conflict evidence.

[0116] For example, if structured sensor monitoring data indicates that the tilt angle of a tower crane exceeds the safety threshold, while on-site image recognition results show that the tower crane is in a normal posture, then the structured sensor monitoring data and on-site images constitute conflicting evidence regarding the entity attribute of "tower crane posture state".

[0117] Regarding correlations, when safety regulations indicate that a certain environmental factor can induce a specific risk, while on-site monitoring results and construction status indicate that the environmental factor has not yet affected the target work object, this can also be marked as conflicting evidence at the relational level. For example, the regulations may indicate that strong winds may induce risks in lifting operations, but wind speed sensors show that the current wind speed is within a safe range. In this case, the regulations represent a general risk correlation, while the sensor data represents the actual risk status at the current moment. The two can be marked as conflicting evidence with differences in applicable conditions, for further judgment based on spatiotemporal conditions.

[0118] In some embodiments, conflict evidence can be classified according to the type of conflict. The conflict types may include numerical conflict, state conflict, attribute conflict, relational conflict, and temporal conflict. Specifically, numerical conflict refers to differences in values ​​given by different sources for the same monitoring indicator exceeding a preset range; state conflict refers to different sources classifying the same work object as being in a normal or abnormal state; attribute conflict refers to different sources providing different descriptions of the location, type, or responsible party of an entity; relational conflict refers to different sources providing different results regarding whether there is a connection between two entities; and temporal conflict refers to inconsistencies in the order in which risks occur, as recorded by different sources.

[0119] Next, the modal source credibility of each conflicting piece of evidence and the temporal proximity between the collection time of each conflicting piece of evidence and the occurrence time of the risk event were determined.

[0120] The credibility of a modal source can be determined based on both the type and quality of the evidence source. For structured sensor monitoring data, the credibility can be determined based on sensor calibration status, equipment failure status, sampling continuity, outlier ratio, and the positional relationship between the monitoring point and the target object. For on-site images, the credibility can be determined based on image clarity, target occlusion degree, acquisition angle, shooting range, confidence level of the recognition model output, and the stability of the target in continuous images. For safety specification texts, the credibility can be determined based on the specification's level of validity, applicable project type, applicable construction stage, and the degree of match between the clauses and the current risk event.

[0121] Temporal proximity is used to characterize the temporal closeness between the state reflected by evidence and the actual state of a risk event. The occurrence time of a risk event can be the actual occurrence time of a risk event that has already occurred, or the first occurrence time of the risk state determined based on monitoring data and on-site images. When the exact occurrence time cannot be determined, the occurrence time of the risk event can be represented as a risk occurrence time interval.

[0122] For conflicting evidence, the time difference between its collection time or the time of occurrence of the state it describes and the time of occurrence of the risk event can be compared. The smaller the time difference, the more accurately the evidence reflects the actual state at the time of the risk event, and the higher its temporal proximity. The larger the time difference, the more likely the evidence reflects a historical state before or after the risk event, and its temporal proximity is relatively low.

[0123] Furthermore, when determining temporal proximity, the effective duration of different data modalities can be considered. For example, structured sensor monitoring data is typically used to reflect the immediate state near the acquisition time, while safety specification texts are used to reflect constraints that do not change with the specific acquisition time. Therefore, based on the data modality and state duration corresponding to the evidence, the effective time range of various types of evidence can be determined, and it can be judged whether the occurrence of the risk event falls within this effective time range.

[0124] Finally, based on the modality source credibility and temporal proximity, retained evidence is identified from conflicting evidence, and the unretained conflicting evidence and their conflict states are associated with the retained evidence.

[0125] Specifically, conflicting evidence can be comprehensively compared based on the modal source credibility and temporal proximity of each piece of conflicting evidence. When the modal source credibility and temporal proximity of a particular piece of conflicting evidence are both higher than those of other conflicting evidence, this piece of conflicting evidence can be identified as reserved evidence, and the result represented by the reserved evidence can be used as the current valid state of the corresponding entity attribute or relationship.

[0126] When one piece of conflicting evidence has a high modal source credibility but low temporal proximity, while another piece of conflicting evidence has a relatively low source credibility but high temporal proximity, the comparison priority can be determined based on the type of information to be judged. For example, for information with strong time sensitivity, such as the immediate operating status of equipment, personnel location, and the current status of protective facilities, temporal proximity can be given priority; for relatively stable information, such as component type, construction specification requirements, and equipment rated parameters, modal source credibility can be given priority.

[0127] When comparisons of multiple conflicting pieces of evidence yield similar results, making it impossible to determine a single piece of evidence to retain, definitive updates to entity attributes or relationships can be temporarily withheld. Instead, the corresponding entity or relationship can be marked as pending verification, and a field verification task can be generated. Once new structured sensor monitoring data, field images, or manual verification results are received, the retained evidence can be re-determined.

[0128] Conflicting evidence that is not retained is not directly deleted, but is associated with retained evidence as historical evidence or objectionable evidence. Specifically, the source of the conflicting evidence that is not retained can be recorded, including the time of collection, the original representation result, the credibility of the source, the temporal proximity, the conflict type, and the reason for not retaining it. Relationships such as "conflict with", "replaced", or "pending review" between it and retained evidence can be established in the knowledge graph.

[0129] For example, a sensor detects that the tower crane's tilt angle exceeds the limit at 10:00 AM, while the on-site video shows the tower crane's attitude is normal at 10:02 AM. If it is further determined that the sensor has exceeded its calibration period, and other attitude sensors did not detect any abnormalities during the same period, the reliability of the sensor data source can be reduced. The "normal attitude" state corresponding to the on-site video can be identified as retained evidence, and the "tilt angle exceeding the limit" data can be associated with this retained evidence as conflicting evidence. The conflicting state can be marked as "suspected sensor abnormality".

[0130] For example, on-site video footage at 9:00 AM showed no cracks in the foundation pit support components, while displacement sensors detected a sudden increase in displacement at 11:00 AM, and on-site video footage acquired at 11:05 AM identified cracks in the support components. If the risk event occurred at 11:00 AM, the video evidence and displacement monitoring evidence from 11:05 AM have a high temporal proximity, and the evidence of "cracks in the support components" and "sudden increase in displacement" can be identified as retained evidence. Although the evidence of "no cracks found" at 9:00 AM is not retained as the current state, it can be retained as historical evidence before the risk state evolved, thus forming a state evolution process of "normal state - increased displacement - crack appearance - support instability risk".

[0131] Through the above processing, the credibility of entities and relationships can be enhanced when multimodal evidence is consistent with each other. When multimodal evidence is conflicting, evidence adjudication can be made by combining the quality of data sources and the time of risk occurrence. At the same time, the source and conflict information of unadopted evidence can be retained, making the risk information in the knowledge graph traceable and providing an evidentiary basis for subsequent local subgraph updates, risk link tracing and dynamic early warning.

[0132] In summary, step S103 first constrains risk information generated at different times based on the sequential relationship between construction procedures. Then, combining the spatial topological relationships between the construction area, work objects, environmental factors, monitoring indicators, and risk sources, it filters and associates entities and relationships that may point to the same risk event. Subsequently, the entities and relationships representing risk causes, risk states, and risk outcomes can be organized into a risk evidence chain according to the risk formation sequence. Based on this, the confidence level of consistent evidence across different modalities can be enhanced. For evidence with inconsistent representations, the credibility of its modal source and its temporal proximity to the time of the risk event are comprehensively compared to determine the evidence to be retained. The remaining conflicting evidence and conflict states are associated and saved, thus forming a risk knowledge expression that combines procedural logic, spatial correlation, evidentiary support, and conflict traceability.

[0133] S104. Instantiate the multi-level ontology architecture based on the entity and relation set, construct the initial knowledge graph of construction project safety risks, and then bind the risk knowledge nodes in the initial knowledge graph with their corresponding spatial locations and time series according to the spatiotemporal tags to establish spatiotemporal co-occurrence associations.

[0134] In some embodiments, constructing an initial knowledge graph of construction project safety risks may include:

[0135] Entities in the entity and relation set are used as graph nodes, and relations in the entity and relation set are used as graph edges. Entity category, modality source, association strength, confidence level, risk status, and spatiotemporal label are used as attribute information of the corresponding graph node or graph edge, and are stored in the graph database to construct the initial knowledge graph.

[0136] In some embodiments, constructing an initial knowledge graph of construction project safety risks refers to organizing and storing the set of entities and relationships obtained after entity and relation extraction, unified semantic mapping, cross-modal association, and multimodal evidence cross-validation in the form of a graph structure, thereby forming a graph data structure that can express the objects of construction project safety risks and their relationships. The initial knowledge graph is used to represent the state of risk knowledge formed based on the currently collected construction site data, and serves as the basis for subsequent new data access, local subgraph location, incremental updates, and dynamic evolution of the target knowledge graph.

[0137] In this embodiment, a graph node refers to a data unit in a knowledge graph used to represent entity objects, corresponding to entities in the entity and relation set. For example, risk sources, construction procedures, environmental factors, monitoring indicators, protective measures, and risk events can all be stored as graph nodes. Among them, risk source nodes can represent risk objects such as foundation pit deformation, formwork support instability, and lack of edge protection; construction procedure nodes can represent operation stages such as earthwork excavation, formwork erection, concrete pouring, and hoisting; monitoring indicator nodes can represent monitoring objects such as displacement values, settlement values, verticality deviation, wind speed, or electrical load; protective measure nodes can represent control measures such as support reinforcement, setting up guardrails, suspending work, or recalculating; and risk event nodes can represent event outcomes such as collapse, electric shock, fall from height, or mechanical injury.

[0138] In this embodiment, a graph edge refers to a data unit in a knowledge graph used to represent the relationships between different graph nodes, corresponding to the relationships in the entity and relation set. Graph edges can represent the accompanying relationship between construction procedures and risk sources, the inducing relationship between environmental factors and risk sources, the representational relationship between monitoring indicators and risk sources, the control relationship between protective measures and risk sources, and the evolutionary relationship between risk sources and risk events. Thus, the knowledge graph can not only store single risk entities, but also represent the influence paths and association links between multiple risk entities.

[0139] In this embodiment, attribute information refers to the descriptive information attached to graph nodes or edges, used to further define the source, type, status, and credibility of the nodes or edges. Specifically, it includes entity category, modal source, association strength, confidence level, risk status, and spatiotemporal label. Entity category indicates the ontology category to which the graph node belongs; modal source indicates that the information corresponding to the node or edge comes from one or more of the following: safety specification text, structured sensor monitoring data, and on-site images; association strength indicates the strength of the association between different entities; confidence level indicates the credibility of the entity or relationship; risk status indicates the risk level or abnormal state corresponding to the entity or relationship; and spatiotemporal label indicates the acquisition time, work area code, measurement point number, component number, or other spatial location information corresponding to the node or edge.

[0140] In this embodiment, the graph database refers to a data storage system used to store graph nodes, graph edges, and their attribute information. By using entities as graph nodes and relationships as graph edges, and storing the corresponding attribute information in the graph database, construction project safety risk information can be expressed in a structured manner in the form of nodes and edges, thereby facilitating subsequent risk queries, link tracing, spatiotemporal co-occurrence analysis, and incremental updates of local subgraphs.

[0141] Specifically, when constructing the initial knowledge graph, the entity and relation set is first read, and each entity is converted into a corresponding graph node according to its category. Secondly, according to the relationship type between entities, each relation is converted into a graph edge connecting the corresponding graph node. Then, the entity category, modality source, association strength, confidence level, risk status, and spatiotemporal label are written into the attribute fields of the corresponding graph node or graph edge. Finally, the graph data after encapsulating nodes, edges, and attributes is stored in the graph database to form the initial knowledge graph of construction project safety risks.

[0142] For example, when the entity and relation set includes the "excavation of foundation pit" construction procedure entity, the "deformation of foundation pit" risk source entity, the "abnormal displacement monitoring value" monitoring indicator entity, and the "support and reinforcement" protection measure entity, the above entities can be constructed as graph nodes respectively. An accompanying relationship edge can be established between "excavation of foundation pit" and "deformation of foundation pit", a representation relationship edge can be established between "abnormal displacement monitoring value" and "deformation of foundation pit", and a control relationship edge can be established between "support and reinforcement" and "deformation of foundation pit". At the same time, the modal source, risk status, confidence level, correlation strength and spatiotemporal label corresponding to each node and edge can be stored as attribute information.

[0143] Therefore, by converting the set of entities and relationships into graph nodes and graph edges, and storing entity categories, modal sources, association strength, confidence levels, risk status, and spatiotemporal labels as attribute information in the graph database, an initial knowledge graph containing both risk objects and risk relationships can be constructed. This provides a data foundation for subsequent risk node binding based on spatiotemporal labels, establishment of spatiotemporal co-occurrence associations, local subgraph localization, and dynamic updating of the knowledge graph.

[0144] In some embodiments, establishing a spatiotemporal co-occurrence association may include:

[0145] Based on the spatiotemporal labels, determine whether different risk knowledge nodes are in the same or adjacent spatial ranges, and determine whether there are overlapping time windows between the time series corresponding to different risk knowledge nodes.

[0146] When different risk knowledge nodes simultaneously satisfy spatial range association and time window overlap, corresponding spatiotemporal co-occurrence associations are established in the initial knowledge graph.

[0147] In some embodiments, establishing spatiotemporal co-occurrence associations refers to, based on the initial knowledge graph having already formed graph nodes and edges, further utilizing the spatiotemporal labels carried by each risk knowledge node to determine whether different risk knowledge nodes are simultaneously associated in spatial location and temporal range, and establishing corresponding spatiotemporal association relationships when conditions are met. Spatiotemporal co-occurrence associations are used to indicate that different risk knowledge nodes are not only semantically related, but also co-occur within the same or adjacent spatial range of the construction site, and within the same or similar time periods, thereby reflecting the possible on-site accompanying relationships, induced relationships, or evolutionary relationships between risks.

[0148] In this embodiment, risk knowledge nodes refer to graph nodes in the initial knowledge graph that are related to the safety risks of construction projects. These may include risk source nodes, construction process nodes, environmental element nodes, monitoring indicator nodes, protective measure nodes, and risk event nodes. Spatiotemporal tags refer to the time and space attributes corresponding to risk knowledge nodes, which may include information such as collection time, work area code, measurement point number, component number, spatial coordinates, or construction stage. Through spatiotemporal tags, the specific time and location of the occurrence or identification of a certain risk knowledge node can be determined.

[0149] In this embodiment, the same spatial range refers to different risk knowledge nodes corresponding to the same work area, the same measuring point, the same component, or the same spatial coordinate range; adjacent spatial ranges refer to different risk knowledge nodes that are not in the exact same position, but their corresponding work areas, measuring points, or components are adjacent or have an influence relationship in the construction space. For example, adjacent foundation pit zones, adjacent work surfaces on the same floor, and work areas within the same hoisting influence range can all be identified as adjacent spatial ranges.

[0150] In this embodiment, an overlapping time window refers to a time interval that is the same or overlaps between the time series corresponding to different risk knowledge nodes. For example, if a certain risk source node is in an abnormal state during a certain time period, and another monitoring indicator node or construction process node is also collected, recorded, or identified during the same time period, then it can be considered that there is an overlap in time windows between the two. The time window can be set to minute level, hour level, day level, or construction stage level according to the needs of construction safety management; this embodiment does not make specific limitations on this.

[0151] Specifically, when establishing spatiotemporal co-occurrence associations, the spatiotemporal labels carried by each risk knowledge node in the initial knowledge graph are first read to determine the spatial location and time series corresponding to each node. Secondly, it is determined whether different risk knowledge nodes are in the same spatial range or adjacent spatial ranges. Thirdly, it is determined whether there are overlapping time windows between the time series corresponding to different risk knowledge nodes. When different risk knowledge nodes simultaneously meet the two conditions of spatial range association and time window overlap, the corresponding spatiotemporal co-occurrence associations are established in the initial knowledge graph.

[0152] For example, when a work area simultaneously contains a "foundation pit excavation" construction process node, a "foundation pit displacement anomaly" monitoring indicator node, and a "slope crack" risk source node within the same time window, a spatiotemporal co-occurrence correlation can be established between these nodes because they spatially correspond to the same or adjacent work areas and temporally belong to the same or similar construction stages. Similarly, when an abnormal verticality deviation corresponding to a component and a structural tilt identified in on-site images occur in the same or adjacent component, and their time windows overlap, a spatiotemporal co-occurrence correlation can be established between the two to indicate a field co-occurrence relationship between the monitored anomaly and the on-site visual risk.

[0153] In some embodiments, spatiotemporal co-occurrence associations can be stored either as graph edges in the initial knowledge graph or as attribute information of existing graph edges. When stored as graph edges, they can be used to directly connect risk knowledge nodes that meet the spatiotemporal co-occurrence conditions; when stored as attribute information, they can be used to enhance the spatiotemporal interpretability of existing associations. Spatiotemporal co-occurrence associations can also carry attribute information such as co-occurrence spatial range, co-occurrence time window, co-occurrence source, and co-occurrence status, so as to facilitate subsequent risk link tracing and incremental updates of local subgraphs.

[0154] Therefore, by judging whether different risk knowledge nodes satisfy spatial range association and time window overlap based on spatiotemporal labels, and establishing spatiotemporal co-occurrence association when the conditions are met, the initial knowledge graph can not only express the semantic relationship between risk objects, but also express the spatiotemporal companion relationship of risk objects in the construction site. This provides a more accurate spatiotemporal association basis for subsequent risk cause analysis, risk link tracing, local subgraph positioning, and dynamic early warning.

[0155] S105. Continuously access newly added construction site data and process the construction site data as multi-source heterogeneous data to obtain new entities and relations; compare the semantic state corresponding to the new entities and relations with the existing state in the initial knowledge graph, locate the affected local subgraphs according to the spatiotemporal labels, and incrementally update the local subgraphs to obtain the target knowledge graph of construction safety risks after dynamic evolution.

[0156] In some embodiments, incrementally updating a local subgraph may include:

[0157] The update scope is determined based on the spatiotemporal tags carried by the newly added entities and relationships;

[0158] Within the scope of the update, the risk status, confidence level, and association strength corresponding to the newly added entities and relationships will be matched with the existing nodes and relationships in the local subgraph.

[0159] When the matching result indicates a new risk, add the corresponding node or relationship to the local subgraph;

[0160] When the matching result indicates a change in risk status, update the attribute information of the corresponding node or relationship;

[0161] When the matching result indicates that the original risk has failed, add historical status tags to the corresponding node or relationship and retain its spatiotemporal traceability information.

[0162] In some embodiments, incremental updates of local subgraphs refer to updating only the affected local areas of the graph after new construction site data is added, based on the existing and continuously operating knowledge graph of construction safety risks. This does not require a full reconstruction of the entire knowledge graph, thereby enabling the graph to evolve dynamically and update in real time.

[0163] Specifically, determining the update scope based on the spatiotemporal tags carried by newly added entities and relationships involves first reading the data collection time and spatial location information attached to the new entities and relationships, and then determining their corresponding work area, measurement point range, component range, or construction stage range. This allows for the location of local subgraph regions in the initial knowledge graph that match or are adjacent to this spatiotemporal scope. This update scope is used to limit the area of ​​effect of subsequent update operations, avoiding redundant calculations of irrelevant graph regions.

[0164] After determining the update scope, the risk status, confidence level, and association strength corresponding to the newly added entities and relationships within that scope are matched with existing nodes and relationships in the local subgraph. This involves semantically comparing and analyzing the new data with the node and edge attributes in the existing knowledge graph. Specifically, risk status indicates the level of risk or degree of anomalousness, confidence level indicates the reliability of the information, and association strength indicates the strength of the relationship between entities. By comparing these three types of attributes with the corresponding attributes of existing nodes and relationships in the local subgraph, it can be determined whether the new information is consistent with existing knowledge, conflicts with it, or belongs to a new category.

[0165] In this embodiment, when the matching result is a new risk, it means that no existing node or relationship with semantic consistency or match with the new entity or relationship was found in the local subgraph. In this case, the corresponding node or relationship can be added to the local subgraph to expand the risk representation capability of the knowledge graph. For example, when new monitoring data identifies a new type of structural anomaly, and there is no corresponding risk source node in the original graph, the risk source node and its associated relationship are added.

[0166] When the matching result indicates a change in risk status, it means that the newly added entity or relation is semantically consistent with an existing node or relation in the local subgraph, but its risk status, confidence level, or association strength has changed. For example, if the displacement value of the same monitoring point changes from a normal state to an abnormal state, then it is only necessary to update the risk status attribute of the corresponding node or update the association strength of the corresponding relation to reflect the evolution of the risk.

[0167] When the matching result indicates that the original risk has become invalid, it means that the newly added data shows that the original risk node or relationship is no longer valid or has been resolved, such as when monitoring indicators return to normal or on-site images confirm that the risk has been rectified. In this case, a historical status tag can be added to the corresponding node or relationship to mark it as historical risk information, while retaining its spatiotemporal traceability information to support subsequent risk retrospective analysis and construction safety audits.

[0168] In some embodiments, historical status tags are used to identify the failure status of the risk node or relationship in the time dimension, but do not delete its record in the knowledge graph; spatiotemporal tracing information is used to record the work area location, collection time and related event information corresponding to the risk node or relationship in the historical stage, thereby ensuring that the knowledge graph has a complete risk evolution recording capability.

[0169] Therefore, by determining the update scope based on spatiotemporal labels and classifying and updating the risk within a local subgraph for new risks, changes in risk status, and risk failures, a local incremental update mechanism for the knowledge graph of construction safety risks can be realized. This reduces computational overhead while ensuring the real-time performance of the graph and enhances the dynamic expression and historical traceability of risk status.

[0170] like Figure 2 As shown, the partial node architecture of this target knowledge graph uses the "core risk source" as the central node, and standardizes the directed semantic relationships between multi-level entities. For example, it clarifies the risk-inducing mechanism of environmental factors, the dynamic accompanying state of construction procedures and risks, the risk control role of protective measures, and the evolution path of risk events caused by risk out of control. This unified semantic framework not only provides a standardized template for cross-modal fusion and spatiotemporal binding of multi-source heterogeneous data, completely solving the data fragmentation problem, but also lays the underlying graph structure foundation for subsequent intelligent risk tracing and dynamic early warning.

[0171] In some embodiments, after step S105, the following may also be included:

[0172] Based on the target knowledge graph, the safety risk assessment of construction projects is carried out, and the risk chain is traced along the relationship between environmental elements, construction procedures, risk sources and risk events.

[0173] Based on the construction process, query the associated risk sources, monitoring indicators, and protective measures;

[0174] When a risk source or risk event reaches a preset risk state after cross-verification with multimodal evidence, dynamic early warning information is output.

[0175] In some embodiments, construction safety risk assessment based on target knowledge graphs refers to using a constructed and continuously updated construction safety risk knowledge graph to comprehensively analyze the relationships between risk sources, construction procedures, environmental factors, monitoring indicators, protective measures, and risk events at the construction site, thereby identifying and judging the formation path, propagation path, and evolution path of risks, and providing decision support for construction site safety management.

[0176] In this embodiment, risk link tracing refers to performing reverse or forward tracing analysis of risk events or risk sources along predefined association paths in a knowledge graph. Specifically, by traversing multiple paths along the relationships between environmental elements, construction procedures, risk sources, and risk events, the possible triggering factors, development process, and scope of impact of a risk event or risk source can be determined. For example, starting from a risk event node, its corresponding risk source node, construction procedure node, and environmental element node can be traced step by step to form a complete risk evolution link, achieving a structured expression of the causes of risk.

[0177] In some embodiments, querying the associated risk sources, monitoring indicators, and protective measures based on a construction process refers to using a construction process node as the query entry point to retrieve risk source nodes, monitoring indicator nodes, and protective measure nodes directly or indirectly associated with that construction process node from the knowledge graph. Here, risk sources represent the types of safety risks that the construction process may cause, monitoring indicators represent monitoring parameters that characterize the risk status, and protective measures represent the safety control measures that should be taken for the construction process and its associated risk sources. This query method enables the linked analysis of risk identification and safety measures starting from the construction process.

[0178] In some embodiments, when a risk source or risk event reaches a preset risk state after cross-validation of multimodal evidence, dynamic early warning information is output. This means that during the operation of the knowledge graph, multi-source data updates are continuously received, and the risk state of the risk source or risk event is evaluated through a multimodal evidence cross-validation mechanism. When its risk state reaches a preset risk threshold or meets the high-risk judgment conditions, the early warning mechanism is triggered and dynamic early warning information is output.

[0179] In this embodiment, the preset risk status can be medium risk, high risk, or abnormal status, etc.; the dynamic early warning information can include the location of the risk, the corresponding construction procedure, the associated risk source, the abnormal situation of the monitoring indicators, and the recommended protective measures, thereby providing timely risk warnings and handling basis for construction site managers.

[0180] Therefore, by using target knowledge graphs for risk assessment, risk link tracing, construction process association query, and dynamic early warning output, it is possible to achieve full-link analysis and real-time response to safety risks at construction sites, thereby improving the accuracy of risk identification, the interpretability of risk propagation paths, and the timeliness and practicality of safety early warnings.

[0181] Therefore, this invention constructs a multi-level ontology architecture for construction project safety risks, uniformly defining entity categories such as risk sources, construction procedures, environmental elements, monitoring indicators, protective measures, and risk events, and setting the relationships between entities, modal evidence binding rules, and spatiotemporal attribute slots. Based on this, it collects multi-source heterogeneous data such as safety specification texts, structured sensor monitoring data, and on-site images, preprocesses them, and adds spatiotemporal tags composed of the collection time and spatial location. Under the constraints of the multi-level ontology architecture, it extracts entities and relationships from the multi-source heterogeneous data after adding spatiotemporal tags, maps the risk information to a unified semantic space, and performs further processing. Cross-modal association and multimodal evidence cross-validation are performed to obtain a set of entities and relations carrying modality source, association strength, confidence level, and spatiotemporal labels. Subsequently, a multi-level ontology architecture is instantiated based on the entity and relation set to construct an initial knowledge graph of construction safety risks. Risk knowledge nodes in the graph are bound to spatial location and time series to establish spatiotemporal co-occurrence associations. At the same time, newly added construction site data is continuously added, and the semantic state corresponding to the newly added entities and relations is compared with the existing state in the initial knowledge graph. The affected local subgraphs are located according to the spatiotemporal labels and incrementally updated to obtain a dynamically evolved target knowledge graph of construction safety risks.

[0182] Through the above methods, this invention can unify scattered data such as safety specification texts, structured sensor monitoring data, and on-site images into a single semantic framework, avoiding the problems of semantic inconsistency and difficulty in associating risk information from different sources; it can perform cross-modal fusion and evidence verification of sensor anomalies, image recognition results, log records, and specification constraints around the same risk source, improving the accuracy of risk identification and risk relationship construction, and reducing false alarms and missed alarms caused by a single data source; it can establish the correspondence between risk knowledge nodes and specific work areas, components, measuring points, and construction stages based on spatiotemporal tags, forming a queryable and traceable risk link; at the same time, after the access of new on-site data, only the affected local subgraphs are incrementally updated, avoiding the computational overhead of full reconstruction, so that the construction project safety risk knowledge graph can dynamically evolve with the changes in the construction site status, thereby improving the timeliness and adaptability of construction safety risk assessment, source tracing analysis, and dynamic early warning.

[0183] like Figure 3 As shown, the construction engineering safety risk knowledge graph construction method based on multi-source data fusion of the present invention operates through conventional components of a computing device, including a processor, memory, communication interface and bus.

[0184] Of course, it is understood that each embodiment of the present invention can achieve one of the effects individually, and the combination of multiple embodiments of the present invention can achieve all the above effects. However, it is not required that each embodiment of the present invention achieve all the above advantages and effects, because each embodiment of the present invention can constitute a separate technical solution and make one or more contributions to the prior art. Each embodiment of the present invention is not affected, and the solution can still be implemented even if one embodiment is deleted.

[0185] For any module structures not specifically defined in this invention, the existing technical descriptions shall prevail. The prior art mentioned in the foregoing background and specific embodiments sections can be considered part of this invention and used to understand the meaning of certain technical features or parameters.

Claims

1. A method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion, characterized in that, The method includes: A multi-level ontology architecture for construction project safety risks is constructed. The multi-level ontology architecture includes risk sources, construction procedures, environmental elements, monitoring indicators, protective measures, and entity categories of risk events. The association relationships between each entity category, modal evidence binding rules, and spatiotemporal attribute slots are defined. Collect multi-source heterogeneous data from the construction site and preprocess it. Add spatiotemporal tags containing the collection time and spatial location to the preprocessed data. The multi-source heterogeneous data includes safety specification text, structured sensor monitoring data, and on-site images. Under the constraints of the multi-level ontology architecture, entities and relationships are extracted from multi-source heterogeneous data after spatiotemporal tags are added. Based on the sequential relationship between construction procedures and the spatial topological relationship between construction areas, work objects and risk sources, entities and relationships pointing to the same risk event are organized into risk evidence chains. The confidence of consistent evidence in the risk evidence chains is enhanced, and conflicting evidence is compared in terms of source credibility and temporal proximity to obtain a set of entities and relationships carrying modal source, association strength, confidence, and spatiotemporal tags. The multi-level ontology architecture is instantiated based on the entity and relation set to construct an initial knowledge graph of construction project safety risks. Then, risk knowledge nodes in the initial knowledge graph are bound to their corresponding spatial locations and time sequences according to the spatiotemporal tags to establish spatiotemporal co-occurrence associations. New construction site data is continuously added and processed as multi-source heterogeneous data to obtain new entities and relationships. The semantic state corresponding to the new entities and relationships is compared with the existing state in the initial knowledge graph. The affected local subgraphs are located according to the spatiotemporal tags, and the local subgraphs are incrementally updated to obtain the dynamically evolved target knowledge graph of construction safety risks.

2. The method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 1, characterized in that: The modal evidence binding rules are used to determine the corresponding data corroboration sources for each entity category; wherein, the risk source and the risk event are corroborated by at least one of the safety specification text, the structured sensor monitoring data, and the on-site images, the monitoring indicators are corroborated by the structured sensor monitoring data, and the protective measures are corroborated by the safety specification text; the spatiotemporal attribute slots are used to determine whether each entity category needs to be bound to the acquisition time, work area code, measurement point number, or component number.

3. The method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 2, characterized in that, The process of collecting and preprocessing multi-source heterogeneous data from construction sites includes: The safety specification text is cleaned, segmented, and denoised; the structured sensor monitoring data is clock-aligned, missing values ​​are processed, and normalized; and the on-site images are subjected to frame extraction, deduplication, and distortion correction.

4. The method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 3, characterized in that, The extraction of entities and relations from multi-source heterogeneous data after the addition of spatiotemporal labels includes: Extract risk sources, construction procedures, protective measures, and regulatory constraints from the aforementioned safety specification text; Based on the correspondence between monitoring indicators and risk sources in the multi-level ontology architecture, the structured sensing monitoring data is transformed into risk status semantic tags. Identify defect-related risk sources, protection-deficiency-related risk sources, or non-compliance-related risk events from the on-site images.

5. A method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in any one of claims 1 to 4, characterized in that, Based on the sequential relationship between construction procedures and the spatial topological relationship between construction areas, work objects, and risk sources, entities and relationships pointing to the same risk event are organized into a risk evidence chain, including: Based on the spatiotemporal tags carried by each entity and relationship, candidate entities and relationships that occur in the same construction area and whose time intervals meet preset conditions are filtered. Based on the multi-level ontology architecture, the construction procedures, work objects, environmental elements, monitoring indicators, risk sources, and protective measures corresponding to the candidate entities are determined respectively. Based on the sequential relationship between construction procedures and the spatial topological relationship between work objects, environmental elements, monitoring indicators and risk sources, the pointing relationship of the candidate entities and candidate relationships to the same risk event is determined. Based on the aforementioned pointing relationship, candidate entities and candidate relationships that respectively represent risk causes, risk states, and risk outcomes are connected in the order of risk formation to obtain the risk evidence chain.

6. The method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 5, characterized in that, The process of enhancing the confidence of consistent evidence in the risk evidence chain and comparing the source credibility and temporal proximity of conflicting evidence includes: Different modal data in the risk evidence chain that point to the same entity attribute or the same relationship and have the same representation result are identified as consistent evidence. Based on the number of modalities of the consistent evidence and the credibility of the modal source, the confidence of the corresponding entity or relationship is increased. Different modal data in the risk evidence chain that point to the same entity attribute or the same relationship but have different representation results are identified as conflicting evidence. Determine the modal source credibility of each conflicting piece of evidence and the temporal proximity between the time of collection of each conflicting piece of evidence and the time of occurrence of the risk event; Based on the modality source credibility and the temporal proximity, retained evidence is determined from the conflicting evidence, and the unretained conflicting evidence and their conflict states are associated with the retained evidence.

7. The method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 1, characterized in that, The initial knowledge graph for constructing safety risks in construction projects includes: The entities in the entity and relation set are used as graph nodes, the relations in the entity and relation set are used as graph edges, and the entity category, modality source, association strength, confidence level, risk status, and spatiotemporal label are used as attribute information of the corresponding graph node or graph edge. All of these are stored in the graph database to construct the initial knowledge graph.

8. The method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 7, characterized in that, The establishment of spatiotemporal co-occurrence correlation includes: Based on the spatiotemporal tags, determine whether different risk knowledge nodes are in the same or adjacent spatial ranges, and determine whether there are overlapping time windows between the time series corresponding to different risk knowledge nodes. When different risk knowledge nodes simultaneously satisfy spatial range association and time window overlap, a corresponding spatiotemporal co-occurrence association is established in the initial knowledge graph.

9. A method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 1 or 8, characterized in that, After obtaining the dynamically evolved target knowledge graph of construction project safety risks, it also includes: Based on the target knowledge graph, the safety risks of construction projects are assessed, and the risk chain is traced along the relationship between environmental elements, construction procedures, risk sources and risk events. Based on the construction procedures, query the associated risk sources, monitoring indicators, and protective measures; When the risk source or risk event reaches a preset risk state after cross-verification with multimodal evidence, dynamic early warning information is output.

10. A method for constructing a knowledge graph of construction project safety risks based on multi-source data fusion as described in claim 1 or 7, characterized in that, The incremental update of the local subgraph includes: The update range is determined based on the spatiotemporal tags carried by the newly added entities and relationships; Within the scope of the update, the risk status, confidence level, and association strength corresponding to the newly added entities and relationships are matched with the existing nodes and existing relationships in the local subgraph; When the matching result indicates a new risk, add the corresponding node or relationship to the local subgraph. When the matching result indicates a change in risk status, update the attribute information of the corresponding node or relationship; When the matching result indicates that the original risk has failed, add historical status tags to the corresponding node or relationship and retain its spatiotemporal traceability information.

Images