A deep learning-based construction supervision intelligent management system and method

Abstract

The application discloses a kind of construction supervision intelligent management system and method based on deep learning, it is related to construction supervision field, including following steps: acquisition and pre-processing construction image data and supervision rule constraint;Construction supervision basic data set is input Grounding DINO network and carries out graph-text semantic collaborative identification and process stage collection;Correlation arrangement is before and after construction state and carries out irreversible feature locking and link risk marking;According to GraphRAG framework, irreversible node identification data and supervision rule constraint are carried out supervision atlas search, conclusion element arrangement and fact node backstepping;Fact supporting matching and fact verification are carried out to necessary construction fact data and construction site state identification data;Determine supervision state and configuration management action, generate construction supervision management result.The application utilizes deep visual identification and atlas enhanced search, realizes construction fact verification, has the advantages of accurate identification, evidence collection in time.

Description

Technical Field

[0001] This invention relates to the field of construction supervision, and in particular to a construction supervision intelligent management system and method based on deep learning. Background Technology

[0002] Current construction process management typically relies on manual inspections, on-site photography, construction logs, and acceptance records for status confirmation, and on verifying the construction objects, areas, times, and stages according to supervision rules. In actual construction, processes such as rebar tying, formwork installation, concrete pouring, and concealed works acceptance are characterized by continuous progression. Once some construction conditions are covered by subsequent processes, it becomes difficult to reconfirm the original connection points, material surfaces, reserved locations, and pre-concealment conditions through on-site inspection. Although existing methods have begun to utilize image recognition technology to assist in identifying construction objects, most remain at the level of single-frame target recognition, manual record comparison, or rule / clause retrieval.

[0003] Existing technologies lack the ability to jointly process construction image data, supervision rule constraints, and construction procedure progress data. This makes it difficult to accurately arrange the connection between the current construction state and subsequent construction states, and also makes it difficult to promptly identify irreversible nodes that affect the verification of the current construction state after the formation of subsequent construction states. At the same time, existing supervision conclusions are usually derived directly from manual judgment or rule matching, lacking a processing link that deduces necessary construction facts from supervision conclusions and then verifies the temporal consistency, spatial consistency, and state integrity of construction facts. This can easily lead to insufficient factual support before concealment, unclear direction for supplementary evidence collection, and insufficient accuracy of construction supervision management results.

[0004] Therefore, how to provide a construction supervision intelligent management system and method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a construction supervision intelligent management system and method based on deep learning. This invention utilizes deep visual recognition and graph enhancement retrieval to realize the verification of construction facts, and has the advantages of accurate identification and timely evidence collection.

[0006] A construction supervision intelligent management method based on deep learning according to an embodiment of the present invention includes the following steps: Acquire construction image data and supervision rule constraints, perform preprocessing, and generate a basic dataset for construction supervision. The basic dataset of construction supervision is input into the Grounding DINO network for collaborative recognition of text and semantics and collection of work process stages to generate construction site status recognition data; Based on the construction process advancement data, the construction status is linked and arranged, and irreversible features are locked and connection risks are marked by combining the construction site status identification data to generate irreversible node identification data. Based on the GraphRAG framework, the data on irreversible node identification and the constraints of supervision rules are used to perform supervision graph retrieval, conclusion element arrangement and fact node back-inference to generate necessary construction fact data. The necessary construction fact data is matched with the construction site status identification data for fact support, and the matching results are verified to generate the necessary construction fact verification data. Based on the necessary construction fact verification data, the supervision status is determined and management actions are configured to generate construction supervision management results.

[0007] Optionally, the preprocessing includes image frame extraction, time stamping, construction area marking, rule-based text structuring, and invalid data removal.

[0008] Optionally, the generation of the construction site status identification data specifically includes: Image frames, time stamps, construction area markers, and structured supervision rule constraints are extracted from the construction supervision basic dataset. Recognition text prompts are generated based on the process name, construction object, and confirmation conditions in the supervision rule constraints. The image frame is input into the image encoder of the Grounding DINO network to extract multi-scale visual features of the construction scene texture, construction object outline, area boundary and visible feature location in the image frame, and generate construction visual feature data. The text prompts are input into the text encoder to encode the semantic content in the process name, construction object, construction area and confirmation conditions, and generate semantic feature data of the supervision prompts. The construction visual feature data and the supervision prompt semantic feature data are input into the feature enhancer to perform cross-modal interaction between image area features and text prompt features, generating image and text enhanced feature data; Input the enhanced image and text feature data into the language-guided query selector to filter construction object candidates, construction area candidates, and key visible feature candidates that match the semantic feature data of the supervision prompts, and generate construction status candidate identification data. The construction status candidate identification data is input into the cross-modal decoder to perform location decoding, category matching, and confidence configuration on the construction object candidate, construction area candidate, and key visible feature candidate to generate the construction status identification result. Based on the time marker, construction area marker, and process name in the supervision rules, the construction status identification results are aggregated into process stages to generate construction site status identification data.

[0009] Optionally, the generation of the irreversible node identification data specifically includes: Acquire construction process progress data and arrange fields according to process name, construction object, construction area, operation time, and record source to generate basic process progress data. The construction process progress data includes construction plan, process sequence, construction ledger, and acceptance record. Based on the process sequence, operation time, construction area and construction object in the basic data of process advancement, the current construction status and subsequent construction status are linked and arranged to generate construction status connection data. By combining the construction site status identification data, the irreversible feature lock is performed on the construction status connection relationship data. The connection relationship that affects the current construction status review after the formation of subsequent construction status is configured as an irreversible connection relationship, and irreversible connection candidate data is generated. The irreversible connection candidate data is matched with the construction site status identification data, and connection risk is marked according to the factual support and subsequent construction status to generate connection risk mark data. The connection risk mark data includes progress mark, correction mark, unconfirmable mark and verifiable completion mark. Based on the connection risk marking data, the node status of the irreversible connection candidate data is arranged, and the arranged node status is associated with the irreversible connection relationship to generate irreversible node identification data. The irreversible node identification data includes irreversible connection relationship, current construction status, subsequent construction status and node status.

[0010] Optionally, the generation of the necessary construction fact data specifically includes: Irreversible node identification data and supervision rule constraints are transformed into nodes and configured with relational edges according to the GraphRAG framework to generate supervision retrieval graph data; Based on the GraphRAG framework, the supervision retrieval graph data is divided into communities, community content is collected, retrieval questions are configured, and local neighborhoods are expanded to generate a supervision association retrieval subgraph. Based on the supervision association retrieval subgraph, the conclusion elements are arranged to determine the supervision conclusion data to be formed for the current construction status, node status and supervision rule constraints. The supervision conclusion data includes the conclusion-related process, conclusion-related object, conclusion basis conditions, conclusion handling content, conclusion-related area and conclusion-related time range. The supervision conclusion data is back-traced by fact nodes. The facts of the construction object, construction area, key visible features, time, and process stage required to support the establishment of the supervision conclusion are back-collected to generate fact node back-tracing data. Based on the fact nodes, duplicate, irrelevant, and expired fact nodes are filtered out, and fact fields are arranged according to construction object, construction area, key visible features, time stamp, and process stage to generate necessary construction fact data.

[0011] Optionally, the generation of the necessary construction fact verification data specifically includes: Perform fact-support matching on necessary construction fact data and construction site status identification data to generate fact-support matching data; Based on factual support matching data, the time stamps and process stage sequence between necessary construction fact data and construction site status identification data are verified for temporal consistency, and temporal consistency verification data is generated. Based on factual support data, spatial consistency verification is performed on the construction area, construction object, and key visible features between the necessary construction fact data and the construction site status identification data, generating spatial consistency verification data. By combining temporal consistency verification data and spatial consistency verification data, the integrity of the factual support for the construction object, key visible features, and process stages is verified. The verification results of factual support matching data, temporal consistency verification data, spatial consistency verification data, and integrity status judgment results are then compiled to generate the necessary construction factual verification data.

[0012] Optionally, the generation of the construction supervision management results specifically includes: Based on the necessary construction fact verification data, the verification status of construction objects, construction areas, key visible features and process stages is classified into status categories to generate supervision status classification data. The supervision status classification data includes verification passed status, verification pending correction status, and verification unconfirmed status. Based on the supervision status classification data, it is determined whether the current construction status meets the conclusion basis conditions in the supervision conclusion data, and management actions are configured in combination with the conclusion handling content to generate supervision management action data. The supervision management action data includes management actions that can be promoted, management actions for supplementary evidence collection, management actions for review and confirmation, and management actions that cannot be confirmed. The system categorizes data such as current construction status, subsequent construction status, node status, and supervision status, and compiles the result fields for supervision management data to generate construction supervision management results.

[0013] According to an embodiment of the present invention, a construction supervision intelligent management system based on deep learning includes: The preprocessing module is used to acquire and preprocess construction image data and supervision rule constraints to generate a basic dataset for construction supervision. The status recognition module is used to input the basic dataset of construction supervision into the Grounding DINO network for collaborative recognition of text, image and semantics and collection of process stages, and to generate construction site status recognition data; The node identification module is used to arrange the construction status in relation to the construction process based on the construction sequence data, and to lock irreversible features and mark connection risks in combination with the construction site status identification data to generate irreversible node identification data. The fact generation module is used to perform supervision map retrieval, conclusion element arrangement, and fact node back-inference according to the GraphRAG framework for irreversible node identification data and supervision rule constraints, and generate necessary construction fact data. The fact verification module is used to match necessary construction fact data with construction site status identification data for fact support, and to perform fact verification on the matching results to generate necessary construction fact verification data. The results generation module is used to determine the supervision status and configure management actions based on necessary construction fact verification data, and generate construction supervision management results.

[0014] The beneficial effects of this invention are: This invention combines construction image data, supervision rule constraints, and construction sequence data for joint processing. This transforms construction status identification from relying solely on manual inspection or single-frame image judgment. Instead, it first generates a basic construction supervision dataset, then utilizes the Grounding DINO network to perform image-text-semantic collaborative identification of construction objects, construction areas, and key visible features, grouping them according to sequence stages. This enables more accurate acquisition of the current construction status during continuous construction, reducing identification errors caused by scattered image information, unstructured rule text, and unclear sequence stages, thus providing a stable data foundation for subsequent supervision judgments.

[0015] This invention uses construction sequence data to correlate and arrange the current and subsequent construction states, and combines this with on-site status identification data to lock irreversible features and mark connection risks. This allows for the timely identification of irreversible nodes that, once a subsequent construction state is formed, may affect the verification of the current construction state. Through this process, in continuous processes such as rebar tying, formwork installation, concrete pouring, and concealed acceptance, it can proactively locate connection points that may prevent the verification of pre-concealed facts, reducing the risks of omissions in construction facts, difficulties in subsequent corrections, and delays in supervisor confirmation.

[0016] This invention also utilizes the GraphRAG framework to perform supervision graph retrieval, conclusion element arrangement, and fact node back-inference on irreversible node identification data and supervision rule constraints, generating necessary construction fact data. Furthermore, it performs fact support matching and fact verification between this necessary construction fact data and construction site status identification data. This creates a continuous support chain from supervision rules, current construction status, necessary construction facts to management actions, providing clear factual basis for supervision status determination and management action configuration. This improves the directionality of corrective evidence collection, reduces invalid review work, and enhances the accuracy and traceability of construction supervision management results. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based intelligent management method for construction supervision proposed in this invention. Figure 2 This is a flowchart illustrating the necessary construction fact data generation process for the intelligent construction supervision management method based on deep learning proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-2 A deep learning-based intelligent management method for construction supervision includes the following steps: Acquire construction image data and supervision rule constraints, perform preprocessing, and generate a basic dataset for construction supervision. The basic dataset of construction supervision is input into the Grounding DINO network for collaborative recognition of text and semantics and collection of work process stages to generate construction site status recognition data; Based on the construction process advancement data, the construction status is linked and arranged, and irreversible features are locked and connection risks are marked by combining the construction site status identification data to generate irreversible node identification data. Based on the GraphRAG framework, the data on irreversible node identification and the constraints of supervision rules are used to perform supervision graph retrieval, conclusion element arrangement and fact node back-inference to generate necessary construction fact data. The necessary construction fact data is matched with the construction site status identification data for fact support, and the matching results are verified to generate the necessary construction fact verification data. Based on the necessary construction fact verification data, the supervision status is determined and management actions are configured to generate construction supervision management results.

[0020] In this embodiment, preprocessing includes image frame extraction, time stamping, construction area marking, rule text structuring, and invalid data removal; Image frame extraction involves extracting construction image data at a rate of 1 frame per second. When the original video frame rate is lower than 1 frame / second, the original frame is retained. The extracted image frames are uniformly adjusted to 1280×720 pixels, and a consecutive frame number is assigned to each image frame. The time stamping includes converting the image frame acquisition time into millisecond-level time stamps. Image frames with missing acquisition times are filled in according to the adjacent valid frame times and frame sequence numbers. Image segments with more than 5 consecutive missing frames are configured as invalid image segments. Construction area marking involves dividing each image frame into 64 construction areas using an 8×8 grid, and assigning area numbers in order from left to right and from top to bottom. The rule text structuring includes splitting the supervision rule constraints into fields according to clause number, process name, construction object, confirmation conditions and disposal requirements. When the length of a single rule text exceeds 120 characters, it is divided into segments according to semicolons or periods. When the length of a single rule text is less than 8 characters, it is merged into adjacent rule texts under the same clause number or the same process name. Invalid data removal includes removing image frames with a resolution below 100, an average brightness of less than 20 or greater than 235, or an effective screen area of ​​less than 70%, as well as removing rule texts with more than two missing rule fields.

[0021] In this embodiment, the generation of construction site status identification data specifically includes: Image frames, time stamps, construction area markers, and structured supervision rule constraints are extracted from the construction supervision basic dataset. Recognition text prompts are generated based on the process name, construction object, and confirmation conditions in the supervision rule constraints. The generation of recognition text prompts includes processing of process object fields, confirmation condition fields, visible feature fields, area limitation fields, and concatenation of prompt statements; Process object field processing includes splitting the process name and construction object in the structured supervision rule constraints, configuring the construction object as the main identification word, and configuring the process name as the process limiting word. The number of main identification words in each identification text prompt is configured to be 1, and the number of process limiting words is not more than 2. The confirmation condition field processing includes filtering the confirmation conditions in the structured supervision rule constraints using status words, configuring the condition fields representing completion status, installation status, connection status, pre-coverage status, and appearance status as status recognition words, with no more than 5 status recognition words in each recognition text prompt; Visible feature field processing includes filtering the visibility of component boundaries, material surfaces, connection parts, reserved locations, and pre-concealment states that can be presented in construction images from the state recognition words, generating key visible feature words. The number of key visible feature words in each recognition text prompt shall not exceed 5. The processing of the area restriction field includes configuring the area number and area location in the construction area mark, and binding the area field with the process restriction word and the main identification word to generate area identification words. The number of area identification words in each identification text prompt shall not exceed 3. The prompt text concatenation process involves combining fields in the order of main identification word, process limiting word, region identification word, key visible feature word, and status identification word, and removing duplicate and empty fields to generate identification text prompts. The length of the concatenated identification text prompts shall not exceed 80 characters. If it exceeds 80 characters, the fields shall be retained in the order of main identification word, key visible feature word, status identification word, process limiting word, and region identification word. The image frame is input into the image encoder of the Grounding DINO network to extract multi-scale visual features of the construction scene texture, construction object outline, area boundary and visible feature location in the image frame, and generate construction visual feature data. Specifically, the process involves: inputting the image frame into the image encoder of the Grounding DINO network; dividing the image frame into image blocks with a side length of 16 pixels; associating each image block with the region number in the construction area marker to generate image block embedding data; encoding the pixel grayscale changes, texture continuity, edge gradients, and structural boundary positions in the image block embedding data to generate basic visual feature data; aggregating adjacent image block features from the basic visual feature data, merging features of image blocks belonging to the same construction area into regional visual features, and aggregating continuous edge features across regions into regional boundary features to generate hierarchical visual feature data; spatiotemporally fusing the hierarchical visual feature data with time markers, region numbers, and image block positions to generate visual feature data with spatiotemporal constraints; hierarchically organizing the visual feature data with spatiotemporal constraints according to three scales: image block, construction area, and whole frame image to generate multi-scale visual feature data; and aggregating the multi-scale visual feature data according to construction scene texture, construction object outline, region boundary, and visible feature position to generate construction visual feature data. The text prompts are input into the text encoder to encode the semantic content in the process name, construction object, construction area and confirmation conditions, and generate semantic feature data of the supervision prompts. Specifically, the process involves: inputting the identified text prompts into a text encoder; marking the semantic fields of the process name, construction object, construction area, and confirmation conditions in the identified text prompts to generate a prompt field sequence; performing semantic embedding and context-related encoding on the prompt field sequence; writing the subordinate relationship between the process name and the construction object, the positional relationship between the construction object and the construction area, and the state relationship between the construction object and the confirmation conditions into the semantic encoding result to generate prompt context semantic data; configuring semantic weights on the prompt context semantic data, configuring the construction object and confirmation conditions as the primary identification semantics, configuring the process name and the construction area as the limiting identification semantics, and removing invalid semantic fields to generate supervisor prompt semantic feature data. Invalid semantic fields include empty fields, duplicate fields, fields not associated with process names, and fields not associated with construction objects; The construction visual feature data and the supervision prompt semantic feature data are input into the feature enhancer to perform cross-modal interaction between image area features and text prompt features, generating image and text enhanced feature data; Specifically, the process involves: inputting construction visual feature data and supervision prompt semantic feature data into a feature enhancer; visually organizing the construction area, image block location, area boundary features, and visible feature location in the construction visual feature data to generate an image area feature sequence; organizing the main recognition semantics, limited recognition semantics, and state recognition semantics in the supervision prompt semantic feature data to generate a text prompt feature sequence; unifying the feature dimensions of the image area feature sequence and the text prompt feature sequence, and performing cross-modal association according to the matching relationships between construction area and area semantics, construction object outline and object semantics, and visible feature location and state recognition semantics to generate image-text associated feature data; configuring text semantic guidance weights for the image area features in the image-text associated feature data, and semantically suppressing image area features with cross-modal matching values ​​less than 0.30, while semantically enhancing image area features with cross-modal matching values ​​not less than 0.30 to generate semantically enhanced area feature data; and interactively updating the semantically enhanced area feature data and the text prompt feature sequence, writing the semantic information of the construction object, construction area, and key visible features into the corresponding image area features to generate image-text enhanced feature data. Input the enhanced image and text feature data into the language-guided query selector to filter construction object candidates, construction area candidates, and key visible feature candidates that match the semantic feature data of the supervision prompts, and generate construction status candidate identification data. Specifically, the process involves: inputting the enhanced image and text feature data into the language-guided query selector; organizing the semantic enhancement region features, text prompt features, and image-text associations within the enhanced image and text feature data into candidate queries to generate language-guided query data; filtering the region features in the language-guided query data that match the semantics of the construction object according to the main identification semantics in the supervision prompt semantic feature data to generate candidate construction objects; filtering the region features in the language-guided query data that match the semantics of the construction area and the area number according to the limited identification semantics in the supervision prompt semantic feature data to generate candidate construction areas; and filtering the region features in the language-guided query data that match the confirmation conditions and visibility features according to the status identification semantics in the supervision prompt semantic feature data. Feature candidate selection is performed on regional features that match the location and boundary features of the area to generate key visible feature candidates. For construction object candidates, construction area candidates, and key visible feature candidates, candidate association is performed within the same image frame, the same construction area, and the same work process stage. Candidates that do not match the semantic feature data provided by the supervisor are eliminated. Candidates with a semantic match value less than 0.25 with the semantic feature data provided by the supervisor are eliminated. If the number of candidates exceeds the retention limit, they are retained in descending order of semantic match value. Within each construction area of ​​each image frame, the number of retained construction object candidates does not exceed 10, the number of retained construction area candidates does not exceed 5, and the number of retained key visible feature candidates does not exceed 10, generating candidate identification data for construction status. The construction status candidate identification data is input into the cross-modal decoder to perform location decoding, category matching, and confidence configuration on the construction object candidate, construction area candidate, and key visible feature candidate to generate the construction status identification result. Specifically, the process involves: inputting candidate construction status recognition data into a cross-modal decoder; expanding candidate features of construction objects, construction areas, and key visible features; and configuring the region location, image patch location, and region boundary features within the candidate features as the basis for location decoding; decoding the target bounding box location of the candidate features in the image frame based on the location decoding basis to generate candidate location data; performing category matching between the candidate location data and the construction object semantics, region semantics, and status recognition semantics in the supervision prompt semantic feature data to generate category matching results; calculating the recognition confidence based on the object semantic matching value, region semantic matching value, status semantic matching value, target bounding box integrity value, and candidate association value; and arranging the candidate location data, category matching results, recognition confidence, frame number, time stamp, and region number, while removing candidate recognition results with a recognition confidence less than 0.40 to generate the construction status recognition result. The calculation of the identification confidence includes: performing consistency judgments between the construction object candidates, construction area candidates, and key visible feature candidates and the construction area markers and state recognition semantics defined by the construction object semantics, area semantics, and respectively; configuring consistent judgment results as the corresponding semantic matching value 1, and inconsistent judgment results as the corresponding semantic matching value 0; configuring the proportion of the area of ​​the target box within the image frame in the candidate location data to the total area of ​​the target box as the target box integrity value; configuring the case where construction object candidates, construction area candidates, and key visible feature candidates exist simultaneously within the same construction area as candidates. The association value is set to 1. If only two candidate classes exist, the candidate association value is set to 0.67; if only one candidate class exists, the candidate association value is set to 0.33. The recognition confidence score is obtained by weighting and summing the object semantic matching value, region semantic matching value, state semantic matching value, target box integrity value, and candidate association value. The weights for object semantic matching value, region semantic matching value, state semantic matching value, target box integrity value, and candidate association value are all set to 0.25, 0.20, and 0.10 respectively. Based on the time marker, construction area marker, and process name in the supervision rules, the construction status identification results are aggregated into process stages to generate construction site status identification data. Specifically, the construction status identification results are collected according to time markers and construction area markers, and associated with the process names in the supervision rules to determine the process stage; the construction status identification results under the same time marker, the same construction area, and the same process stage are merged, and the result with the highest identification confidence is selected as the stage representative identification result; the stage representative identification results are arranged into fields to generate construction site status identification data.

[0022] In this embodiment, the generation of irreversible node identification data specifically includes: Acquire construction process progress data and arrange fields according to process name, construction object, construction area, operation time, and record source to generate basic process progress data. Construction process progress data includes construction plan, process sequence, construction ledger, and acceptance record. The construction plan is used to determine the connection relationship between planned processes, the process sequence is used to determine the relationship between the current construction status and subsequent construction status, and the construction ledger and acceptance record are used to determine the formation of the construction status. Based on the process sequence, operation time, construction area, and construction object in the basic data of process advancement, the current construction status and subsequent construction status are linked and arranged to generate construction status connection relationship data. The construction status connection relationship data includes the current process name, the subsequent process name, the current construction object, the connection occurrence area, and the connection time range. The connection time range consists of the end time of the current construction status to the start time of the subsequent construction status. By combining the construction site status identification data, the irreversible feature lock is performed on the construction status connection relationship data. The connection relationship that affects the current construction status review after the formation of subsequent construction status is configured as an irreversible connection relationship, and irreversible connection candidate data is generated. Specifically, the following steps are taken: The current and subsequent process names in the construction status connection data are verified sequentially; current and subsequent construction statuses with continuous progression within the same construction area are configured as connection states to be locked; based on the current construction object, connection location, and connection time range, the visible state of the current construction object in the connection state to be locked is reviewed; if the current construction object or key visible features have no recognition results for two consecutive image frames after the formation of the subsequent construction state, it is configured as a review-restricted state; if the material exposure state or the visible state of the connection part changes before and after the formation of the subsequent construction state, it is configured as a visible state change; and connection states to be locked that are in a review-restricted state or a visible state change are configured as irreversible connection relationships; irreversible connection relationships are arranged as candidates according to the current process name, subsequent process name, current construction object, connection location, and connection time range to generate irreversible connection candidate data. The irreversible connection candidate data is matched with the construction site status identification data, and connection risk is marked according to the factual support and subsequent construction status to generate connection risk mark data. The connection risk mark data includes the advance mark, the correction mark, the unconfirmed mark, and the verifiable completion mark. Specifically, the following steps are performed: The current process name, current construction object, connection location, and connection time range in the irreversible connection candidate data are matched with the process stage, construction object, construction area, time marker, key visible features, and identification confidence level in the construction site status identification data to generate a current construction status matching result. Identification results in the current construction status matching result that match the construction object, construction area, process stage, have key visible features, have a time marker within the connection time range, and have an identification confidence level of not less than 0.40 are configured as valid factual support; identification results with missing matches in the construction object, construction area, key visible features, and process stage are configured as missing factual support. Based on the subsequent process name and connection time range in the irreversible connection candidate data, the following steps are performed: The status identification results of subsequent process stages in the same construction area of ​​the construction site status identification data are verified to form a status. Cases where there are subsequent process stage identification results and the time stamp is later than the time stamp in the current construction status matching result are configured as subsequent construction has been formed, and cases where there are no subsequent process stage identification results are configured as subsequent construction has not been formed. Irreversible connection candidate data with valid factual support and no subsequent construction is configured as progressable markers, irreversible connection candidate data with missing factual support and no subsequent construction is configured as needing correction markers, irreversible connection candidate data with missing factual support and subsequent construction is configured as unconfirmable markers, and irreversible connection candidate data with valid factual support and subsequent construction is configured as verifiable completion markers, generating connection risk marker data. Based on the connection risk marking data, the node status of the irreversible connection candidate data is arranged, and the arranged node status is associated with the irreversible connection relationship to generate irreversible node identification data. The irreversible node identification data includes irreversible connection relationship, current construction status, subsequent construction status and node status.

[0023] In this embodiment, the generation of necessary construction fact data specifically includes: Irreversible node identification data and supervision rule constraints are transformed into nodes and configured with relational edges according to the GraphRAG framework to generate supervision retrieval graph data; Specifically, the irreversible connection relationships, current construction status, subsequent construction status, node status, and connection occurrence areas associated with irreversible connection relationships in the irreversible node identification data are configured as connection relationship nodes, current status nodes, subsequent status nodes, node status nodes, and area nodes, respectively. The clause numbers, process names, construction objects, confirmation conditions, and handling requirements in the supervision rule constraints are configured as rule clause nodes, process nodes, object nodes, condition nodes, and handling nodes, respectively. Rule attribution edges are configured based on the attribution relationship between clause numbers and process names, construction objects, and confirmation conditions; state connection edges are configured based on the sequential relationship between the current construction status and subsequent construction status; and connection attribution edges are configured based on the connection attribution relationship between irreversible connection relationships and the current and subsequent construction statuses. Based on the marking relationship between irreversible connection relationships and node status, risk marking edges are configured; based on the subordinate relationship between process name and construction object, object constraint edges are configured; based on the status affiliation relationship between the current construction status and subsequent construction status and process name and construction object, respectively, status process edges and status object edges are configured; based on the rule constraint relationship between confirmation conditions and disposal requirements, rule disposal edges are configured; based on the positional relationship between the current construction status, subsequent construction status and connection occurrence area, status area edges are configured; the number of condition nodes under the same clause number is configured to not exceed 12; the number of relationship edges directly associated with the same current status node is configured to not exceed 20; the total number of current status nodes and subsequent status nodes associated with the same connection relationship node is configured to not exceed 4; and supervision retrieval map data is generated. Based on the GraphRAG framework, the supervision retrieval graph data is divided into communities, community content is collected, retrieval questions are configured, and local neighborhoods are expanded to generate a supervision association retrieval subgraph. Specifically, the process involves: using rule clause nodes, process nodes, object nodes, condition nodes, disposal nodes, current status nodes, subsequent status nodes, connection relationship nodes, node status nodes, and region nodes as community division objects; configuring graph nodes with the same process node, object node, region node, or connection relationship node into the same search community, with no more than 80 graph nodes in each search community; aggregating community content for rule clause nodes, process nodes, object nodes, condition nodes, disposal nodes, current status nodes, subsequent status nodes, connection relationship nodes, node status nodes, and region nodes within the same search community to generate community search content; determining the construction object and process stage to be searched based on the current construction status in the irreversible node identification data, determining the risk relationship type to be searched based on the irreversible connection relationship, determining the risk status range to be searched based on the node status, and limiting the connection impact range in conjunction with the subsequent construction status to generate search question data; and finally, generating search question data based on the construction object, process stage, risk relationship type, and risk status range in the search question data. To determine the target search communities, the content is matched with the scope of influence and the scope of connection. The number of target search communities is configured to not exceed 5. If more than 5 are selected, they are retained in descending order of the number of matches based on the construction object, process stage, risk relationship type, and node status. The current status nodes, rule clause nodes, and condition nodes that match the scope of construction object, process stage, and risk status in the target search communities are configured as central nodes. Object nodes are searched along the state object edges and object constraint edges associated with the central nodes. Subsequent status nodes are searched along the state connection edges connected to the central nodes. Connection relationship nodes and node status nodes are searched along the risk marker edges connected to the central nodes. Regional nodes are searched along the state region edges connected to the central nodes. Disposal nodes are searched along the rule disposal edges associated with the central nodes. The number of hops for the relationship edges of local neighborhood expansion is configured to not exceed 3. The number of adjacent nodes retained for each central node is configured to not exceed 20. The central nodes, as well as the found object nodes, subsequent status nodes, connection relationship nodes, node status nodes, regional nodes, and disposal nodes, are aggregated into a subgraph to generate a supervision association search subgraph. Based on the supervision association retrieval subgraph, the conclusion elements are arranged according to the association content between the current construction status, node status, and supervision rule constraints. The supervision conclusion data to be formed in the current construction status is determined. The supervision conclusion data includes the conclusion-related process, the conclusion-related object, the conclusion basis conditions, the conclusion handling content, the conclusion-related area, and the conclusion-related time range. Among them, the process node connected to the current status node in the supervision association retrieval subgraph is determined as the conclusion-related process; the object node connected to the current status node is determined as the conclusion-related object; the condition node that is associated with the current status node through the process node and object node and has a subordinate relationship with the rule clause node is determined as the conclusion basis conditions; the handling node that has a rule handling relationship with the rule clause node and condition node is determined as the conclusion handling content; the area node that is associated with the current status node, object node, and connection relationship node is determined as the conclusion-related area; and the connection time range in the irreversible node identification data is determined as the conclusion-related time range. The supervision conclusion data is back-traced by fact nodes. The facts of the construction object, construction area, key visible features, time, and process stage required to support the establishment of the supervision conclusion are back-collected to generate fact node back-tracing data. Specifically, based on the conclusion-related processes, conclusion-related objects, conclusion-based conditions, conclusion-handling content, conclusion-related areas, and conclusion-related time ranges in the supervision conclusion data, determine the factual support scope required for the establishment of the supervision conclusion; configure the conclusion-related objects as construction object facts, the conclusion-related areas as construction area facts, and configure the component boundaries, material surfaces, connection parts, reserved positions, and pre-concealment states involved in the conclusion-based conditions as key visible feature facts; configure the time markers and conclusion-related time ranges required to support the establishment of the supervision conclusion as time facts; configure the conclusion-related processes, the process stage to which the current construction state belongs, and the process stage to which the subsequent construction state belongs as process stage facts; and perform reverse aggregation according to the support relationships between the supervision conclusion data and the construction object facts, construction area facts, key visible feature facts, time facts, and process stage facts to generate fact node back-inference data; Based on the fact nodes, duplicate, irrelevant, and expired fact nodes are filtered out. Fact fields are then arranged according to construction object, construction area, key visible features, time markers, and process stages to generate necessary construction fact data. The number of necessary construction fact data corresponding to the same supervision conclusion data is configured to not exceed 30, of which the number of facts about construction object does not exceed 8, the number of facts about construction area does not exceed 5, the number of facts about key visible features does not exceed 10, the number of time facts does not exceed 4, and the number of facts about process stages does not exceed 3. The filtering of duplicate fact nodes includes: configuring fact nodes that are identical in terms of construction object facts, construction area facts, key visible feature facts, time facts, and process stage facts in the back-inference data as duplicate fact nodes, and retaining the fact nodes that form the most supporting relationships with the supervision conclusion data; The screening of irrelevant fact nodes includes: configuring fact nodes in the back-inference data of fact nodes that do not form a supporting relationship with the conclusion-related process, conclusion-related object, conclusion-based conditions, conclusion-handling content, conclusion-related area, and conclusion-related time range in the supervision conclusion data as irrelevant fact nodes, and removing them from the back-inference data of fact nodes; The filtering of invalid fact nodes includes: configuring fact nodes whose time facts are more than 300 seconds earlier than the start time of the conclusion-related time range, whose construction area facts are inconsistent with the conclusion-related area, and whose process stage facts are inconsistent with the current construction status or subsequent construction status as invalid fact nodes, and removing them from the fact node back-inference data.

[0024] In this embodiment, the generation of necessary construction fact verification data specifically includes: The necessary construction fact data and construction site status identification data are matched to generate fact support matching data. Fact support matching includes aligning the construction objects, construction areas, key visible features, time markers and process stages in the necessary construction fact data with the construction site status identification data, comparing them item by item and matching the status markers. Based on the factual support matching data, the time stamps and process stage sequence between the necessary construction fact data and the construction site status identification data are verified for temporal consistency, and temporal consistency verification data is generated. The temporal consistency verification includes comparing the time stamps, conclusion association time ranges, current construction status and subsequent construction status in the factual support matching data, verifying the scope attribution and marking the process stage sequence. Based on the fact-supported matching data, spatial consistency verification is performed on the construction area, construction object, and key visible features between the necessary construction fact data and the construction site status identification data, generating spatial consistency verification data. Spatial consistency verification includes verifying the region affiliation, comparing the object region, and marking the feature region between the construction area, construction object, and key visible features in the fact-supported matching data and the construction area, construction object, and key visible features in the construction site status identification data. Combining temporal consistency verification data and spatial consistency verification data, the completeness of factual support for construction objects, key visible features, and process stages is verified. The verification results of factual support matching data, temporal consistency verification data, spatial consistency verification data, and completeness determination results are then compiled to generate necessary construction factual verification data. The completeness verification includes collecting supporting items, comparing the completeness of verification results from temporal consistency verification data and spatial consistency verification data with the construction objects, key visible features, and process stages in the factual support matching data, and determining their completeness.

[0025] In this embodiment, the generation of construction supervision management results specifically includes: Based on the necessary construction fact verification data, the verification status of construction objects, construction areas, key visible features and process stages is classified into status categories to generate supervision status classification data. The supervision status classification data includes verification passed status, verification pending correction status and verification unconfirmed status. The status classification includes: configuring the case where the fact-supporting matching data, temporal consistency verification data, spatial consistency verification data, and complete status judgment result are all passed as the verification passed status; configuring the case where the fact-supporting matching data has missing items, but the temporal consistency verification data, spatial consistency verification data, and complete status judgment result do not show any inconsistencies as the verification pending correction status; configuring the case where the temporal consistency verification data, spatial consistency verification data, or complete status judgment result has inconsistencies as the verification unconfirmed status. Based on the supervision status classification data, it is determined whether the current construction status meets the conclusion basis conditions in the supervision conclusion data. In combination with the conclusion handling content, management actions are configured to generate supervision management action data, which includes management actions that can be promoted, management actions for supplementary evidence collection, management actions for review and confirmation, and management actions that cannot be confirmed. The management action configuration includes: configuring a "progressable management action" when the supervision status classification data is in the "verification passed" state and the current construction status meets the conclusion basis conditions; configuring a "correction and evidence collection management action" when the supervision status classification data is in the "verification passed" or "verification pending correction" state and the current construction status does not meet the conclusion basis conditions; configuring a "review and confirmation management action" when the supervision status classification data is in the "verification pending correction" state and the current construction status meets the conclusion basis conditions; and configuring an "unconfirmable" management action when the supervision status classification data is in the "verification unconfirmable" state. The system categorizes data such as current construction status, subsequent construction status, node status, and supervision status, and compiles the result fields for supervision management data to generate construction supervision management results.

[0026] A deep learning-based intelligent management system for construction supervision includes: The preprocessing module is used to acquire and preprocess construction image data and supervision rule constraints to generate a basic dataset for construction supervision. The status recognition module is used to input the basic dataset of construction supervision into the Grounding DINO network for collaborative recognition of text, image and semantics and collection of process stages, and to generate construction site status recognition data; The node identification module is used to arrange the construction status in relation to the construction process based on the construction sequence data, and to lock irreversible features and mark connection risks in combination with the construction site status identification data to generate irreversible node identification data. The fact generation module is used to perform supervision map retrieval, conclusion element arrangement, and fact node back-inference according to the GraphRAG framework for irreversible node identification data and supervision rule constraints, and generate necessary construction fact data. The fact verification module is used to match necessary construction fact data with construction site status identification data for fact support, and to perform fact verification on the matching results to generate necessary construction fact verification data. The results generation module is used to determine the supervision status and configure management actions based on necessary construction fact verification data, and generate construction supervision management results.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a construction supervision scenario of a residential building's main structure. This scenario included one basement level and 18 above-ground levels. Six fixed cameras and two mobile inspection cameras were connected to the site, collecting construction footage from three consecutive days. This footage was then linked to 12 construction plans, 18 process sequence records, 36 construction ledger records, 9 acceptance records, and 8 supervision rules. In this scenario, rebar tying, formwork installation, concrete pouring, and concealed acceptance proceeded continuously. As the current construction status was overwritten by subsequent construction statuses, it became difficult to re-verify connection points, reserved locations, and pre-concealment conditions. This easily led to problems such as missing construction facts, lack of supporting evidence for supervision conclusions, and delayed supplementary evidence collection.

[0028] In application, the system inputs construction images, construction plans, work sequence, construction ledgers, acceptance records, and supervision rules into the processing flow. These supervision rules cover 4 types of work process names, 6 types of construction objects, 9 types of confirmation conditions, and 5 types of handling requirements. The system generates recognition text prompts based on the work process name, construction object, and confirmation conditions. It then inputs the construction images and recognition text prompts into the Grounding DINO network to perform image-text-semantic collaborative recognition of reinforcing bars, formwork, embedded parts, connection points, reserved locations, material surfaces, and their pre-concealment state. This results in 126 sets of construction object recognition results, 48 ​​sets of construction area recognition results, 92 sets of key visible feature recognition results, and 18 sets of work process stage aggregation results.

[0029] For consecutive rebar tying and formwork installation states within the same construction area, the system configures the rebar tying state as the current construction state and the formwork installation state as the subsequent construction state, determining the connection time range based on the work time in the construction log. This scenario generates 15 sets of construction state connection relationship data, of which 9 sets are configured as irreversible connection candidate data, further forming 4 sets of "progressable" markers, 3 sets of "pending correction" markers, 1 set of "unconfirmable" markers, and 1 set of "verifiable completion" markers.

[0030] Subsequently, the system constructs a supervision retrieval graph data according to the GraphRAG framework, forming 8 rule clause nodes, 4 process nodes, 6 object nodes, 9 condition nodes, 5 disposal nodes, 9 current status nodes, and 9 subsequent status nodes. It then uses the supervision-related retrieval subgraph to infer necessary construction fact data. In this scenario, 27 necessary construction fact data items are generated, including 7 construction object facts, 4 construction area facts, 9 key visible feature facts, 4 time facts, and 3 process stage facts.

[0031] Finally, the system performs factual support matching, temporal consistency verification, spatial consistency verification, and status integrity verification on the necessary construction factual data and construction site status identification data. Based on the verification passed status, verification pending correction status, and verification unconfirmable status, it configures the management actions that can be promoted, the correction and evidence collection management actions, the review and confirmation management actions, and the unconfirmable management actions, respectively, and generates the construction supervision management results.

[0032] Based on the construction images collected over three consecutive construction days, eight supervision rules, 15 sets of construction status connection data, and 27 necessary construction fact data from the above embodiments, a comparative statistical analysis of the traditional manual verification method and the method of this invention yields the following performance data: Table 1. Performance Comparison of the Invention and Traditional Construction Supervision Management Methods

[0033] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0034] In terms of the accuracy of construction status identification, the traditional method achieves 86.4%, while the method of this invention achieves 90.8%, an improvement of 4.4 percentage points. This improvement mainly comes from the Grounding DINO network's collaborative recognition of image frames and recognition text prompts, and the inclusion of construction objects, construction areas, and key visible features in the recognition process. This reduces the instability in judging the process stage when relying solely on manual inspection or single visual recognition.

[0035] In terms of the matching rate of key visible features, the traditional method achieves 82.7%, while the method of this invention achieves 88.9%, an improvement of 6.2 percentage points. This improvement is related to the reverse deduction and factual support matching of necessary construction factual data. The system does not directly judge the supervision conclusion, but first deduces the facts of the construction object, the construction area, key visible features, time, and process stage, and then compares them item by item with the construction site status identification data, making it easier to match content such as connection parts, reserved locations, and pre-concealment status.

[0036] In terms of recall rate for identifying irreversible connections, the traditional method achieves 80.5%, while the method of this invention achieves 86.1%, an improvement of 5.6 percentage points. This is because the present invention correlates the current construction state with subsequent construction states and locks in irreversible features based on the formation of subsequent construction processes, enabling more accurate identification of the connection relationships that affect the verification of the current construction state after the formation of subsequent processes.

[0037] In terms of the pass rate of necessary construction fact verification, the traditional method is 78.9%, while the method of this invention is 85.2%, an improvement of 6.3 percentage points. This result is related to the verification of temporal consistency, spatial consistency, and state integrity. The system performs joint verification of time markers, construction areas, construction objects, key visible features, and process stages, reducing misjudgments caused by missing factual support and incomplete verification fields.

[0038] In terms of the accuracy of evidence correction, the traditional method achieves 76.8%, while the method of this invention achieves 82.4%, an improvement of 5.6 percentage points. This is because the present invention configures management actions based on the verification pass status, verification pending correction status, and verification unconfirmable status, enabling the evidence correction object to be generated around the missing construction facts, rather than requiring general manual review. The average time for a single fact verification is reduced from 42.6 seconds to 35.8 seconds, mainly because the GraphRAG framework first completes the supervision map retrieval and fact node back-inference, reducing the repeated retrieval of irrelevant rules and image content.

[0039] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A construction supervision intelligent management method based on deep learning, characterized in that, Includes the following steps: Acquire construction image data and supervision rule constraints, perform preprocessing, and generate a basic dataset for construction supervision. The basic dataset of construction supervision is input into the Grounding DINO network for collaborative recognition of text and semantics and collection of work process stages to generate construction site status recognition data; Based on the construction process advancement data, the construction status is linked and arranged, and irreversible features are locked and connection risks are marked by combining the construction site status identification data to generate irreversible node identification data. Based on the GraphRAG framework, the data on irreversible node identification and the constraints of supervision rules are used to perform supervision graph retrieval, conclusion element arrangement and fact node back-inference to generate necessary construction fact data. The necessary construction fact data is matched with the construction site status identification data for fact support, and the matching results are verified to generate the necessary construction fact verification data. Based on the necessary construction fact verification data, the supervision status is determined and management actions are configured to generate construction supervision management results.

2. The intelligent construction supervision management method based on deep learning according to claim 1, characterized in that, The preprocessing includes image frame extraction, time stamping, construction area marking, rule-based text structuring, and invalid data removal.

3. The intelligent construction supervision management method based on deep learning according to claim 1, characterized in that, The generation of the construction site status identification data specifically includes: Image frames, time stamps, construction area markers, and structured supervision rule constraints are extracted from the construction supervision basic dataset. Recognition text prompts are generated based on the process name, construction object, and confirmation conditions in the supervision rule constraints. The image frame is input into the image encoder of the Grounding DINO network to extract multi-scale visual features of the construction scene texture, construction object outline, area boundary and visible feature location in the image frame, and generate construction visual feature data. The text prompts are input into the text encoder to encode the semantic content in the process name, construction object, construction area and confirmation conditions, and generate semantic feature data of the supervision prompts. The construction visual feature data and the supervision prompt semantic feature data are input into the feature enhancer to perform cross-modal interaction between image area features and text prompt features, generating image and text enhanced feature data; Input the enhanced image and text feature data into the language-guided query selector to filter construction object candidates, construction area candidates, and key visible feature candidates that match the semantic feature data of the supervision prompts, and generate construction status candidate identification data. The construction status candidate identification data is input into the cross-modal decoder to perform location decoding, category matching, and confidence configuration on the construction object candidate, construction area candidate, and key visible feature candidate to generate the construction status identification result. Based on the time marker, construction area marker, and process name in the supervision rules, the construction status identification results are aggregated into process stages to generate construction site status identification data.

4. The intelligent construction supervision management method based on deep learning according to claim 1, characterized in that, The generation of the irreversible node identification data specifically includes: Acquire construction process progress data and arrange fields according to process name, construction object, construction area, operation time, and record source to generate basic process progress data. The construction process progress data includes construction plan, process sequence, construction ledger, and acceptance record. Based on the process sequence, operation time, construction area and construction object in the basic data of process advancement, the current construction status and subsequent construction status are linked and arranged to generate construction status connection data. By combining the construction site status identification data, the irreversible feature lock is performed on the construction status connection relationship data. The connection relationship that affects the current construction status review after the formation of subsequent construction status is configured as an irreversible connection relationship, and irreversible connection candidate data is generated. The irreversible connection candidate data is matched with the construction site status identification data, and connection risk is marked according to the factual support and subsequent construction status to generate connection risk mark data. The connection risk mark data includes progress mark, correction mark, unconfirmable mark and verifiable completion mark. Based on the connection risk marking data, the node status of the irreversible connection candidate data is arranged, and the arranged node status is associated with the irreversible connection relationship to generate irreversible node identification data. The irreversible node identification data includes irreversible connection relationship, current construction status, subsequent construction status and node status.

5. The intelligent construction supervision management method based on deep learning according to claim 1, characterized in that, The generation of the necessary construction fact data specifically includes: Irreversible node identification data and supervision rule constraints are transformed into nodes and configured with relational edges according to the GraphRAG framework to generate supervision retrieval graph data; Based on the GraphRAG framework, the supervision retrieval graph data is divided into communities, community content is collected, retrieval questions are configured, and local neighborhoods are expanded to generate a supervision association retrieval subgraph. Based on the supervision association retrieval subgraph, the conclusion elements are arranged to determine the supervision conclusion data to be formed for the current construction status, node status and supervision rule constraints. The supervision conclusion data includes the conclusion-related process, conclusion-related object, conclusion basis conditions, conclusion handling content, conclusion-related area and conclusion-related time range. The supervision conclusion data is back-traced by fact nodes. The facts of the construction object, construction area, key visible features, time, and process stage required to support the establishment of the supervision conclusion are back-collected to generate fact node back-tracing data. Based on the fact nodes, duplicate, irrelevant, and expired fact nodes are filtered out, and fact fields are arranged according to construction object, construction area, key visible features, time stamp, and process stage to generate necessary construction fact data.

6. The intelligent construction supervision management method based on deep learning according to claim 1, characterized in that, The generation of the necessary construction fact verification data specifically includes: Perform fact-support matching on necessary construction fact data and construction site status identification data to generate fact-support matching data; Based on factual support matching data, the time stamps and process stage sequence between necessary construction fact data and construction site status identification data are verified for temporal consistency, and temporal consistency verification data is generated. Based on factual support data, spatial consistency verification is performed on the construction area, construction object, and key visible features between the necessary construction fact data and the construction site status identification data, generating spatial consistency verification data. By combining temporal consistency verification data and spatial consistency verification data, the integrity of the factual support for the construction object, key visible features, and process stages is verified. The verification results of factual support matching data, temporal consistency verification data, spatial consistency verification data, and integrity status judgment results are then compiled to generate the necessary construction factual verification data.

7. The intelligent construction supervision management method based on deep learning according to claim 1, characterized in that, The generation of the construction supervision management results specifically includes: Based on the necessary construction fact verification data, the verification status of construction objects, construction areas, key visible features and process stages is classified into status categories to generate supervision status classification data. The supervision status classification data includes verification passed status, verification pending correction status, and verification unconfirmed status. Based on the supervision status classification data, it is determined whether the current construction status meets the conclusion basis conditions in the supervision conclusion data, and management actions are configured in combination with the conclusion handling content to generate supervision management action data. The supervision management action data includes management actions that can be promoted, management actions for supplementary evidence collection, management actions for review and confirmation, and management actions that cannot be confirmed. The system categorizes data such as current construction status, subsequent construction status, node status, and supervision status, and compiles the result fields for supervision management data to generate construction supervision management results.

8. A deep learning-based intelligent management system for construction supervision, executing the deep learning-based intelligent management method for construction supervision as described in any one of claims 1 to 7, characterized in that, include: The preprocessing module is used to acquire and preprocess construction image data and supervision rule constraints to generate a basic dataset for construction supervision. The status recognition module is used to input the basic dataset of construction supervision into the Grounding DINO network for collaborative recognition of text, image and semantics and collection of process stages, and to generate construction site status recognition data; The node identification module is used to arrange the construction status in relation to the construction process based on the construction sequence data, and to lock irreversible features and mark connection risks in combination with the construction site status identification data to generate irreversible node identification data. The fact generation module is used to perform supervision map retrieval, conclusion element arrangement, and fact node back-inference according to the GraphRAG framework for irreversible node identification data and supervision rule constraints, and generate necessary construction fact data. The fact verification module is used to match necessary construction fact data with construction site status identification data for fact support, and to perform fact verification on the matching results to generate necessary construction fact verification data. The results generation module is used to determine the supervision status and configure management actions based on necessary construction fact verification data, and generate construction supervision management results.

Images