Artificial intelligence-based method for identifying risk points of construction quality of multi-ribbed slab floor

Abstract

The application provides a method for identifying risk points of construction quality of a multi-ribbed beam floor based on artificial intelligence; the method organizes three-stage inspection information of formwork installation, support erection and pre-pouring review into point records according to the axis network coordinates, extracts and splices inspection items, measured and photo video semantic features, inputs the point risk score and risk source process prompts in combination with the stage conditions; and performs risk joint checking expansion based on a support chain, a formwork row and a boundary reinforcing belt correlation diagram, generates a review list according to the joint checking score and the on-duty capacity, and forms a closed loop by writing back the review conclusion; the scheme realizes phased and traceable point risk identification and review priority sequencing, improves the risk positioning accuracy and the review efficiency, and reduces the missed detection rate.

Description

Technical Field

[0001] This invention relates to the field of construction quality monitoring and machine learning technology, and in particular to an artificial intelligence-based method for identifying construction quality risk points in closely spaced ribbed floor slabs. Background Technology

[0002] As a large-span, economical floor slab construction method, the construction of ribbed beam floor slabs typically involves key processes such as formwork installation, support erection, and pre-pouring verification. On-site quality management is based on a grid coordinate system, locating each floor and axis line to specific points. It collects multi-source information including inspection results, measured data, and photographic / video descriptions, and conducts verification and rectification based on the supporting structure, formwork arrangement, and boundary reinforcement strips. The management objective is to promptly locate potential risk points, rationally arrange the verification sequence, and form a closed loop within limited manpower and shifts, thereby reducing the transmission of quality hazards before pouring.

[0003] Existing technologies mostly employ mobile inspections and form-based data entry, combining inspection item templates and thresholds to judge measured data, and retaining photos and videos as supporting evidence. Some systems are based on BIM or construction drawings, establishing ledgers for points according to axis grids and recording inspection information by process stage; risk screening typically relies on rule bases, statistical scoring, or supervised models to score and summarize inspection items, classify images and text, or compare measured values ​​against thresholds. Review arrangements are mostly based on fixed priorities, experience, or simple weighted sorting, with some solutions providing association retrieval based on adjacent axes or components in the same row, and updating records through review terminals to achieve basic data closure.

[0004] However, existing methods generally lack a unified expression and comparable mechanism for multi-source information on the same network points under three-stage conditions, making it difficult to form stable features at the point scale. Risk propagation often remains at the level of nearest neighbors or empirical correlations, lacking structured propagation and intensity quantification based on support chains, shell arrays, and boundary reinforcement zones as defined paths. Review tasks often fail to combine the available review capacity for the shift with the constructed path to generate a list of points with joint verification relationships, and closed-loop verification is insufficient in corresponding to the risk source processes.

[0005] Therefore, a method for identifying construction quality risk points in closely spaced ribbed floor slabs that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an artificial intelligence-based method for identifying construction quality risk points in ribbed beam floor slabs. The core technical problem to be solved by this application is: in the construction scenario of ribbed beam floor slabs, how to uniformly model multi-source inspection information under the axis grid points and stage dimensions, and determine the specific points that need to be reviewed first and their joint inspection sources under the constraints of construction relationships, generate an executable list for the shift, and realize closed-loop management.

[0007] The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to embodiments of the present invention includes:

[0008] S1. Obtain on-site inspection information during the installation, support erection, and pre-pouring verification stages of the ribbed beam floor slab formwork. Divide the area into points according to the axis grid coordinates, and write the inspection item records, actual measurement records, and photo and video descriptions into the corresponding points to form a point process inspection record.

[0009] S2. Extract construction stage identifiers and inspection item statuses from the point process inspection records to form phased point characteristics, and establish point connections based on the structural relationships of support chains, formwork rows, and boundary reinforcement strips to form a structural association diagram;

[0010] S3. Input the phased point features into the diffusion-type anomaly discrimination network, and perform diffusion back-calculation under the constraints of construction phase identification. Calculate the point risk score of each point's deviation from the qualified point process. Decompose the point risk score according to the dimensions of formwork installation, support erection, and pre-pouring review to form a risk source process prompt. When the point risk score meets the trigger threshold, determine the risk trigger point.

[0011] S4. Based on the constructed association diagram, perform risk linkage expansion on the risk trigger points, and limit the linkage expansion path to the connection of the same support chain, the connection of the same mold shell, and the connection of the boundary reinforcement strip. Use the risk score of the point as the linkage expansion strength, calculate the linkage expansion score of the connected points along the linkage expansion path, and record the linkage relationship corresponding to the connected points.

[0012] S5. Based on the threshold of the number of verifiable points for the shift, sort and extract the joint investigation extended scores to generate a list of verifiable points for the shift. The list of verifiable points for the shift includes point identifiers, verification order identifiers, and joint investigation relationships.

[0013] S6. Receive the review conclusions corresponding to the review point list for the shift, form a review conclusion mark for the point, and write the review conclusion mark for the point into the point process inspection record to obtain the closed-loop point process inspection record.

[0014] Optionally, S1 is as follows:

[0015] Collect on-site inspection information during the stages of formwork installation, support erection, and pre-pouring verification; read the grid coordinates and generate point markers that correspond one-to-one with the grid coordinates;

[0016] The inspection item records corresponding to the location identifiers are coded in a state according to the preset inspection item number order to form a 48-dimensional inspection item state vector.

[0017] The measured records corresponding to the point markers are numerically aggregated according to the preset measured field order to form a sixteen-dimensional measured record vector.

[0018] The photo and video descriptions corresponding to the location markers are vectorized and mapped according to a preset semantic description lexicon to form a 32-dimensional photo and video description vector. The inspection item records, actual measurement records, photo and video descriptions and stage markers are written into the location process inspection record.

[0019] Optionally, S2 is as follows:

[0020] Read the construction stage identifier corresponding to the point identifier from the point process inspection record, and extract the inspection item status vector, actual measurement record vector, and photo and video description vector;

[0021] The inspection item status vector, measured record vector, and photo and video description vector are concatenated to form a 96-dimensional phased point feature, and the construction stage identifier is 3D one-hot encoded as the stage condition input for the diffusion anomaly discrimination network.

[0022] Based on the grid coordinates of the location markers and the structural layout information of the ribbed beam floor slab, determine the corresponding support chain markers, formwork row markers, and boundary reinforcement strip markers.

[0023] Using point markers as nodes, nodes with the same support chain marker, the same mold shell row marker, and the same boundary reinforcement strip marker are connected to form a point connection diagram.

[0024] Optionally, S3 specifically refers to:

[0025] Before inputting the phased point features into the diffusion anomaly discrimination network, the construction phase identifier is 3D one-hot encoded and then fully connected to generate a 12-dimensional phase condition vector. The diffusion step size identifier is fully connected to generate a 16-dimensional step size condition vector.

[0026] The 96-dimensional phased point features are concatenated with the 12-dimensional phased condition vector and the 16-dimensional step size condition vector to form a 124-dimensional network input vector, and the network input vector is written into the initial step size of the diffusion backpropagation calculation.

[0027] A fully connected residual stacking structure without self-attention is used to perform layer-by-layer denoising calculation on the network input vector. The fully connected residual stacking structure consists of four residual blocks in sequence, and each residual block contains two fully connected layers with 256 fully connected neurons.

[0028] At the block input of each residual block, the stage condition vector and step condition vector are concatenated again, and the residual block output and residual block input are added together and then passed to the next residual block to maintain the diffusion back-reasoning path under the construction stage identification constraint.

[0029] The outputs of the four residual blocks are input into a dimension-reduced fully connected layer containing 128 fully connected neurons, and a 96-dimensional noise prediction vector is output through a denoising output layer to obtain a reconstructed vector of the phased location features. The inverse residual sequence is formed based on the difference between the phased location features and the reconstructed vector at each diffusion step.

[0030] The reverse residual sequence is divided into three residual segments according to the inspection item status vector, the measured record vector, and the photo and video description vector. The three residual segments are then mapped through the risk output layer to obtain a one-dimensional point risk score. At the same time, the three residual segments are mapped to obtain a three-dimensional risk source process prompt. The three-dimensional risk source process prompt corresponds one-to-one with the formwork installation dimension, support erection dimension, and pre-pouring review dimension.

[0031] The risk score of a location is determined by comparing it with the trigger threshold. When the risk score of a location meets the trigger threshold, the corresponding location is identified as a risk trigger location and output to the risk linkage extended calculation of the constructed association graph.

[0032] Optionally, when mapping the three residuals to obtain a one-dimensional location risk score through the risk output layer, the location risk score is calculated according to a metric function, which is specifically:

[0033] ;

[0034] in, To define the number of steps in the preset diffusion step size set, The step size number, The residual dimension number. To reverse the residual sequence In the The 96-dimensional residual vector at step size is the first... Dimensional residuals, For the first Each step size corresponds to a step size weight, which is fixed by the back-calculated update coefficient table. For construction phase identification The phase index obtained from the mapping, , , For risk output layer in stage index The corresponding segment weights, as learnable parameters of the risk output layer, are fixed during the offline training phase and applied to the state vector interval of the inspection item, the measured record vector interval, and the photo / video description vector interval, respectively. To provide a one-dimensional risk score, the risk output layer further aggregates the three residual intensities of each step and outputs a three-dimensional risk source process indicator through a fully connected mapping. Three-dimensional risk source process reminder It corresponds one-to-one with the dimensions of formwork installation, support erection, and pre-pouring verification.

[0035] Optionally, S4 specifically refers to:

[0036] The risk trigger points are input to construct a correlation graph, the nodes corresponding to the risk trigger points are located, and the risk scores of the risk trigger points are read as the initial linkage expansion strength.

[0037] Adjacency search is performed starting from the node corresponding to the risk trigger point. The adjacency search is limited to propagation only along the point connections. The point connections are composed of connections with the same support chain, connections with the same mold shell, and connections with the boundary reinforcement strip. Node relationships other than the point connections do not participate in the joint search expansion calculation.

[0038] For each point connection obtained through adjacency retrieval, extract the connection type identifier, select a preset attenuation coefficient based on the connection type identifier, and use the preset attenuation coefficient as the joint query extension path parameter for that point connection;

[0039] Using the initial joint investigation expansion strength as input, the first-level joint investigation expansion score is calculated for the connected points directly connected to the risk trigger point according to the joint investigation expansion path parameters. The risk trigger point identifier, connected point identifier, and connection type identifier are combined in the propagation order to generate the first-level joint investigation relationship.

[0040] The first layer of connected points are sequentially set as extension nodes, and the corresponding first layer of joint query extension score is used as the joint query extension strength of the extension node. Adjacency retrieval, connection type identifier extraction, joint query extension path parameter selection and joint query extension score calculation are repeated to generate multi-layer joint query relationships layer by layer, so that the joint query extension score propagates along the construction path of connection with the same support chain, connection with the same mold shell row, and connection with the boundary reinforcement strip.

[0041] The termination condition for the joint query expansion process is set as the number of expansion layers reaching the preset maximum number of expansion layers or the joint query expansion strength of the expansion node being lower than the joint query cutoff threshold. When the same connected point is accessed through multiple joint query expansion paths, the maximum value of the joint query expansion score of the connected point is taken, and the multi-level joint query relationship that generates the maximum value is determined as the joint query relationship of the connected point.

[0042] The output includes the identifiers of connected points, the score of the joint query extension, and the joint query relationship. The joint query extension result is used as the sorting input for generating the list of points to be reviewed during the shift.

[0043] Optionally, when the same connected point is accessed through multiple linked expansion paths, an optimization function is used to update the linked expansion score, wherein the optimization function is specifically:

[0044] ;

[0045] in, To expand the number of layers and satisfy , For the first The point identifier corresponding to the layer extension node. To and The connection points are identified by the point-to-point connection. Connecting point identifiers In the The extended score obtained from the join operation at each layer is initialized to zero before the first calculation and can be updated repeatedly within the same layer. Point identification In the The layer's joint expansion strength, and the zeroth layer satisfies And satisfy for any layer , Connect the points to the point identifier Point of view marker The connection type identifier, Identified by connection type The preset attenuation coefficient is found in the preset attenuation coefficient table. This is a maximum value operator used to retain a larger joint query extension score when multiple joint query extension paths exist for accessing the same connected point identifier.

[0046] Optional, S5 specifically includes:

[0047] Receive the output of the joint investigation and expansion results, and extract the point identifier, joint investigation and expansion score, and joint investigation relationship corresponding to each connected point. At the same time, extract the point identifier of the risk trigger point.

[0048] Write the location identifier of the risk trigger point into the set to be sorted, and merge records with duplicate location identifiers in the set to be sorted. When merging, retain the maximum join expansion score and retain the join relationship that generated the maximum join expansion score.

[0049] The set to be sorted is sorted from high to low according to the joint query expansion score. When the joint query expansion scores are the same, they are sorted in a secondary order according to the location risk score from high to low, and a review order queue is generated.

[0050] Based on the threshold number of reviewable points for the current shift, the review order queue is truncated and assigned review order identifiers in sequence to generate a review point list for the current shift. The review point list for the current shift includes point identifiers, review order identifiers, and linkage relationships.

[0051] Optionally, step S6 specifically includes:

[0052] Receive the review conclusions corresponding to each location identifier in the on-duty review point list. The review conclusions include the review check item results and a conclusion identifier indicating whether an abnormality is confirmed.

[0053] The review conclusions are structured and mapped to generate point review conclusion markers that correspond one-to-one with the point identifiers. The point review conclusion markers include a conclusion identifier and a corresponding risk source process prompt confirmation identifier.

[0054] The point verification conclusion is marked and written into the point process inspection record, which is consistent with the point identifier and the construction stage identifier, to form a closed-loop point process inspection record. The closed-loop point process inspection record is used as the input source for subsequent extraction of point features in stages.

[0055] The beneficial effects of this invention are:

[0056] (1) This proposal puts forward an improved diffusion-based anomaly discrimination method and conditional modeling technique. It constructs 96-dimensional phased location features based on 48-dimensional inspection items with fixed semantics, 16-dimensional measured data, and 32-dimensional photo / video descriptions. It introduces phase-specific hot conditions up to 12 dimensions and 16-dimensional step size conditions. It combines four-layer fully connected residual stacking with a preset step size coefficient table to perform diffusion backpropagation, and uses segmented residuals and phase-related segment weights to measure location risk and source process indications. Compared to discrimination methods based on rule thresholds or supervised classification, this improvement achieves comparable expression and reproducible backpropagation paths for the same location across processes under three-stage condition constraints. This enables risk identification to have stable dimensional semantics and interpretable process source indications, supporting the determination of risk trigger points based on thresholds.

[0057] (2) This proposal puts forward a novel structural correlation diagram joint investigation expansion method and path propagation technology. Taking points as nodes, it propagates only along three types of connections: the same support chain, the same formwork arrangement, and the boundary reinforcement zone. A preset attenuation coefficient is assigned according to the connection type, and the joint investigation expansion score and joint investigation relationship are generated by merging the maximum value of multiple paths and controlling the number of layers / intensity. Unlike the practice of simply treating adjacent axes as homogeneous adjacency or empirical nearest neighbor diffusion, this method transforms the point risk output by the model into the propagation intensity on the structural path under the constraint of the solid structural relationship. This ensures the consistency of risk expansion with the support system, formwork arrangement, and boundary reinforcement zone, and facilitates targeted joint investigation in the upstream and same row range of the structure.

[0058] (3) This proposal proposes a method and sorting and writing-back technique for on-duty execution and closed-loop management. Based on the joint sorting of joint inspection extended scores and point risk scores, a review list containing joint inspection relationships is generated under the constraints of on-duty capacity. When the review terminal returns the data, the source process prompts are confirmed by mapping the inspection item to the process dimension. The conclusion mark and confirmation mark are written back to the process record of the corresponding point and stage. Unlike the practice of allocating review resources only according to fixed weights or experience priorities, the overall approach integrates point-level risk identification, construction path joint inspection, executable list and closed-loop record into the same data structure and axis network index. This enables multi-source inspection information to be uniformly modeled under the point and stage dimensions, and determines the specific points that need to be reviewed first and their joint inspection sources under the constraints of construction relationships. This supports the on-site progress according to the list and the consistent accumulation of subsequent data. Attached Figure Description

[0059] 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:

[0060] Figure 1 This is a flowchart of a method for identifying construction quality risk points in a closely spaced ribbed beam floor slab based on artificial intelligence, as proposed in this invention.

[0061] Figure 2 This invention presents a flowchart of the process inspection record construction for a ribbed beam floor slab construction quality risk point identification method based on artificial intelligence.

[0062] Figure 3 This is a flowchart illustrating the generation of a phased point feature and structural correlation diagram for a method for identifying construction quality risk points in a closely spaced ribbed slab floor system based on artificial intelligence, as proposed in this invention.

[0063] Figure 4 The flowchart of the diffusion-based anomaly discrimination network calculation for the artificial intelligence-based method for identifying construction quality risk points in closely spaced ribbed floor slabs proposed in this invention is shown below.

[0064] Figure 5 This invention presents a flowchart of the extended risk linkage diagram for identifying construction quality risk points in a closely spaced ribbed beam floor slab, based on artificial intelligence.

[0065] Figure 6 This is a flowchart illustrating the process of generating a checkpoint list for the construction quality risk point identification method for closely spaced ribbed beam floor slabs proposed in this invention.

[0066] Figure 7This is a flowchart of the review conclusion writing and closed-loop point inspection record of the method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence proposed in this invention.

[0067] Figure 8 This invention presents a floor slab grid-based risk linkage diagram for identifying construction quality risk points in ribbed slab floor systems, based on artificial intelligence. The diagram uses grid intersections as points, with solid blue lines, dashed green lines, and orange dots representing structural associations formed by connections to the same support chain, formwork shell, and boundary reinforcement strip, respectively. Red lightning bolts represent risk trigger points identified by a diffusion-based anomaly discrimination network. Surrounding points are arranged in shades from dark to light to represent the linkage expansion intensity obtained by attenuation propagation along the structural path. Black circle numbers represent the review list and sequence extracted during the shift under constraints on the number of verifiable items. The method integrates "anomaly identification—structural linkage—review assignment," avoiding random sampling and single-point alarms, improving positioning efficiency and review hit rate, and forming a closed-loop data support for continuous optimization. Detailed Implementation

[0068] In Example 1, reference Figures 1 to 8 A method for identifying construction quality risk points in ribbed beam floor slabs based on artificial intelligence, comprising:

[0069] S1. Obtain on-site inspection information during the installation, support erection, and pre-pouring verification stages of the ribbed beam floor slab formwork. Divide the area into points according to the axis grid coordinates, and write the inspection item records, actual measurement records, and photo and video descriptions into the corresponding points to form a point process inspection record.

[0070] S2. Extract construction stage identifiers and inspection item statuses from the point process inspection records to form phased point characteristics, and establish point connections based on the structural relationships of support chains, formwork rows, and boundary reinforcement strips to form a structural association diagram;

[0071] S3. Input the phased point features into the diffusion-type anomaly discrimination network, and perform diffusion back-calculation under the constraints of construction phase identification. Calculate the point risk score of each point's deviation from the qualified point process. Decompose the point risk score according to the dimensions of formwork installation, support erection, and pre-pouring review to form a risk source process prompt. When the point risk score meets the trigger threshold, determine the risk trigger point.

[0072] S4. Based on the constructed association diagram, perform risk linkage expansion on the risk trigger points, and limit the linkage expansion path to the connection of the same support chain, the connection of the same mold shell, and the connection of the boundary reinforcement strip. Use the risk score of the point as the linkage expansion strength, calculate the linkage expansion score of the connected points along the linkage expansion path, and record the linkage relationship corresponding to the connected points.

[0073] S5. Based on the threshold of the number of verifiable points for the shift, sort and extract the joint investigation extended scores to generate a list of verifiable points for the shift. The list of verifiable points for the shift includes point identifiers, verification order identifiers, and joint investigation relationships.

[0074] S6. Receive the review conclusions corresponding to the review point list for the shift, form a review conclusion mark for the point, and write the review conclusion mark for the point into the point process inspection record to obtain the closed-loop point process inspection record.

[0075] In this embodiment, step S1 specifically includes:

[0076] Step S1 is used to organize the inspection information of the ribbed beam floor construction site into a point process inspection record that can be directly consumed by the diffusion anomaly discrimination network. In order to ensure that the subsequent diffusion back calculation can be executed in stages under the construction stage identifier constraint, this embodiment forms the inspection information of different construction stages under the same axis coordinate into record entries, and solidifies the construction stage identifier and the dimensional structure of the three types of input vectors in each record.

[0077] The data collected includes on-site inspection information from the formwork installation, support erection, and pre-pouring verification stages. This information includes inspection item records, measured data, photo and video descriptions, and corresponding grid coordinates. The grid coordinates are marked as follows: It consists of horizontal axis numbers and vertical axis numbers, which will Input the point identifier generation rules, using floor number and The axis numbering is spliced ​​to generate point markers. and will and Binding ensures that each set of grid coordinates corresponds to a unique point identifier within the same floor; the construction stage identifier is denoted as... The values ​​are limited to three categories: formwork installation, support erection, and pre-pouring verification. They are used to generate stage condition vectors in subsequent steps and run through the residual block splicing, thereby avoiding the mixing and comparison of the phased point features of different process stages.

[0078] Mark the location The corresponding inspection item records are input into a preset inspection item number table. This preset inspection item number table contains a fixed set of forty-eight inspection items and provides an order in which the inspection items are numbered. Each inspection item is coded according to its review result, using a three-value coding rule: qualified is recorded as one, unqualified as negative one, and not inspected as zero. The coded values ​​of the forty-eight inspection items are arranged sequentially according to the preset inspection item number order to obtain a forty-eight-dimensional inspection item state vector. By fixing the order of inspection item numbers, different locations and shifts can form... Maintain consistency in dimensional semantics to ensure the stability of the splicing dimensions of subsequent phased point features;

[0079] Mark the location The corresponding measured records are input into a preset measured field table, which contains sixteen measured fields with a given field order. Measured values ​​are read for each measured field and their values ​​are aggregated. The sixteen measured values ​​are then arranged sequentially according to the preset measured field order to obtain a sixteen-dimensional measured record vector. When a measured field is missing, the measured value of that field is set to zero, while maintaining... The order of the fields remains unchanged to ensure that the input dimension of the subsequent diffusion-based anomaly detection network is constant;

[0080] Mark the location The corresponding photo and video descriptions are input from a preset semantic description lexicon. This lexicon is constructed during system initialization based on photo and video description text from historical location process inspection records. A preset word frequency threshold is used to filter the semantic description word set, and an offline word vector training model is used to generate 32-dimensional word vectors from the set, forming a lookup table mapping from semantic description words to 32-dimensional word vectors. Simultaneously, a fixed set of 32-dimensional unknown word vectors is fixed in the preset semantic description lexicon to represent unincluded words. Maximum matching word segmentation is performed on the photo and video descriptions using a preset word segmentation dictionary. The segmentation results are mapped to the corresponding 32-dimensional word vectors one by one. Each word is weighted according to its frequency of occurrence in the photo and video description. The weights are multiplied dimension-wise by the corresponding 32-dimensional word vectors and then accumulated word by word to obtain the 32-dimensional photo and video description vector. When word segmentation fails to match the preset semantic description vocabulary, a 32-dimensional unknown word vector is used for accumulation, combining inspection records, actual measurement records, photo and video descriptions, and construction stage identifiers. Jointly write the point process inspection record and in China-Israel point markers Construction phase markings As an index field, make the same Record entries are created for each of the three construction phases, allowing for independent retrieval of data in subsequent steps. Direct extraction , , It also splices together 96-dimensional phased point features and can simultaneously extract construction phase identifiers. Used as input for stage-specific conditional constraints, satisfying the consistency requirements of the diffusion-type anomaly detection network for input dimensions and stage constraints.

[0081] In this embodiment, step S2 specifically includes:

[0082] Step S2 is used to convert the point process inspection records into phased point features that can be directly input into the diffusion-based anomaly discrimination network, and simultaneously establish a construction correlation diagram for risk linkage expansion, so that the point risk score output by subsequent diffusion back-calculation can be propagated along the construction path of the support chain, shell row, and boundary reinforcement zone. The point process inspection record is denoted as... The location is marked as The axis coordinates are marked as The construction phase is marked as The state vector of the inspection item is denoted as The measured recorded vector is denoted as The vector describing the photo and video is denoted as The phased location characteristics are denoted as The unique heat vector of the stage is denoted as The structural layout information for closely spaced ribbed floor slabs includes the support chain arrangement, denoted as follows: The arrangement of the mold shells is denoted as follows: The layout of the boundary reinforcement zone is denoted as follows: The support chain identifier is determined by the aforementioned arrangement table and denoted as follows. The mold shell is marked as The boundary reinforcement strip is marked as The construction of the association graph is denoted as The set of nodes in the constructed association graph is denoted as . The set of points connected is denoted as ;

[0083] Point marking As an index, records are checked from the point-of-care process. Read the corresponding entry and read the construction stage identifier from the entry. Simultaneously, extract the check item state vector from the entry. Measured recording vector Photo and video description vectors ,in It is fixed at forty-eight dimensions, and the order of the dimensions is consistent with the preset check item number order. It is fixed at sixteen dimensions, and the order of the dimensions is consistent with the preset order of the measured fields. The format is fixed at 32 dimensional and obtained by lookup mapping and weighted accumulation of a preset semantic descriptor lexicon, using point-based identifiers. Construction phase markings The process of jointly locating points is checked and recorded, so that the same grid coordinates can be separated into separate input data entries at different construction stages;

[0084] Check item status vector Measured recording vector Photo and video description vectors The features are assembled in a fixed order to form 96-dimensional, phased point features. The fixed order is to first assemble. Then splice it together Then splice it together Thus The first 48 dimensions correspond to the inspection item status vector, the middle 16 dimensions correspond to the measured record vector, and the last 32 dimensions correspond to the photo and video description vector, thus identifying the construction stage. Three-dimensional one-heat encoding is performed to obtain three-dimensional stage one-heat vectors. The formwork installation corresponds to the first dimension taking one and the remaining dimensions taking zero; the support erection corresponds to the second dimension taking one and the remaining dimensions taking zero; the pre-pouring verification corresponds to the third dimension taking one and the remaining dimensions taking zero. The stage-specific heat vectors are then used. As the source of stage conditions for the diffusion anomaly discrimination network, the subsequent diffusion back-calculation is mapped to the qualified point process representation space of each stage under the construction stage identification constraint.

[0085] According to the location markings Associated grid coordinates Determine the structural grouping identifier and grid coordinates based on the structural layout information of the closely spaced ribbed floor slab. The support chain arrangement table consists of horizontal axis numbers and vertical axis numbers. Mold Layout Table Layout table of boundary reinforcement strip During the system initialization phase, the data is generated from the construction drawings or building information model. During parsing, the range of axis numbers covered by the support chains, formwork rows, and boundary reinforcement strips in the plane is fixed as a grid interval, and a mapping is established between the grid intervals and their corresponding identifiers. The grid intervals do not overlap and cover the grid coordinates of all points. The grid intervals do not overlap and cover the grid coordinates of all points. The grid interval allows for the exclusion of some point grid coordinates, and the grid coordinates are... Input support chain layout table And locate the grid interval into which it falls to determine the support chain identifier. , axis coordinates Input mold shell layout table And locate the grid interval into which it falls to determine the mold shell row identifier. , axis coordinates Input boundary reinforcement strip layout table And locate the grid intervals that fall within the boundary to determine the boundary reinforcement zone markers. When the axis coordinates When the boundary reinforcement zone does not fall within any boundary reinforcement zone grid interval, the boundary reinforcement zone will be marked. The value is assigned to zero to represent a non-boundary reinforcement zone point;

[0086] Point marking As a node, construct the association graph. All location identifiers are aggregated into a node set. Point connection set Composed of connections with the same support chain, connections with the same mold shell, and connections with boundary reinforcement strips, all bearing the same support chain identifier. The nodes are extracted into a group, according to the axis coordinates. The horizontal axis numbers are sorted in ascending order, and a support chain is established between any two adjacent nodes after sorting. The starting point identifier, ending point identifier, and connection type of the connection are written into the point connection set. Those with the same mold shell markings The nodes are extracted into a group, according to the axis coordinates. The longitudinal axis numbers are sorted in ascending order, and after sorting, a connection is established between two adjacent nodes using the same mold shell, which is then written into the point connection set. Those with the same and non-zero boundary reinforcement band markings The nodes are extracted into a group, according to the axis coordinates. The horizontal axis numbers are sorted in ascending order, and boundary reinforcement strips are established between adjacent nodes after sorting and written into the point connection set. By connecting the points The connection type is fixed to ensure that subsequent risk investigation and expansion are limited to the construction of the correlation diagram only under the constraints of the same support chain connection, same mold shell row connection, and boundary reinforcement strip connection. spread.

[0087] In this embodiment, step S3 specifically includes:

[0088] Step S3 is used to perform diffusion back-calculation on the phased point features and output the point risk score, thereby realizing anomaly identification based on the process performance of qualified points under the construction stage identification constraint, and using the risk trigger point as the starting point for risk linkage expansion in the construction association diagram. The phased point features are denoted as follows. The location is marked as The construction phase is marked as The diffusion step size is marked as The stage condition vector is denoted as The step size conditional vector is denoted as The network input vector is denoted as The diffuse anomaly detection network is denoted as The block state of the residual stack is denoted as The noise prediction vector is denoted as The reconstructed vector is denoted as The reverse residual sequence is denoted as The risk score for each location is recorded as follows: The risk source process prompt is recorded as follows: The trigger threshold is denoted as The number of steps in the preset diffusion step size set is denoted as And for each step length, a fixed back-dated update coefficient table is generated. The back-dated update coefficient table includes noise intensity parameters, fidelity coefficients and noise coefficients used to initialize disturbances, combination coefficients used for denoising updates between steps, and step length weights used for pooling point risk scores.

[0089] Mark the construction stage Performing three-dimensional one-heat encoding yields three-dimensional stage one-heat vectors. ,Will The input-stage conditional fully connected mapping layer generates a twelve-dimensional stage conditional vector. Diffusion step size identifier The input step-size conditional fully connected mapping layer generates a 16-dimensional step-size conditional vector. diffusion step size identifier The steps are taken from a preset diffusion step size set, which is written and fixed in order during system deployment. The stage condition fully connected mapping layer and the step size condition fully connected mapping layer serve as a diffusion-based anomaly detection network. The constituent layers are trained together and their parameters are fixed during the offline training phase;

[0090] 96-dimensional phased point features With the twelve-dimensional stage condition vector Sixteen-dimensional step-size conditional vector The 124-dimensional network input vector is generated by splicing the vectors in a fixed order. The fixed order is to first assemble. Then splice it together Then splice it together The initial step size for the diffusion backpropagation calculation is set to the maximum step size in the preset diffusion step size set, and... Write the back-pull state cache corresponding to the initial step size. To obtain the back-pull vector of the initial step size, generate a 96-dimensional Gaussian noise vector from the pseudo-random number generator, and then divide the phased point features. The initial step size input vector is obtained by multiplying the initial step size by the fidelity coefficient in the updated coefficient table dimension by dimension, multiplying the 96-dimensional Gaussian noise vector by the noise coefficient of the initial step size dimension by dimension, and then adding the two dimension by dimension. The random seed of the pseudo-random number generator is identified by the point. Generate by combining the construction date;

[0091] A fully connected residual stack structure is used to perform layer-by-layer denoising computation on the initial step-size input vector. The fully connected residual stack structure consists of four residual blocks arranged sequentially. Each residual block contains a first fully connected layer and a second fully connected layer. The first fully connected layer consists of 256 fully connected neurons, and the second fully connected layer also consists of 256 fully connected neurons. The block input of the first residual block is initialized to be the sum of the initial step-size input vector and... , The concatenated vector;

[0092] The stage condition vector is concatenated again at the block input of each residual block. With step size condition vector The splicing results are sequentially input into the first fully connected layer and the second fully connected layer to obtain the block output. The block output and the block input are then added together with the dimension-wise residuals to obtain the block fused output. and output the block fusion. The construction phase identifier constraint is passed to the next residual block as its block input, so that the construction phase identifier constraint continues to participate in the diffusion back-propagation path update within each residual block;

[0093] The output of the fourth residual block is fused into the output. The dimensionality-reduced state is obtained by inputting a fully connected layer containing 128 fully connected neurons into a dimensionality-reduced layer, and then inputting the dimensionality-reduced state into a denoising output layer to obtain a 96-dimensional noise prediction vector. Based on the combined coefficients of the current step size in the back-inference update coefficient table, a denoising update is performed on the vector to be back-inferred at the current step size. The denoising update includes a noise prediction vector. Generate the reconstruction vector at the current step size And based on the reconstructed vector The combination coefficients generate the vector to be reversed for the next step. For a preset set of diffusion step sizes, noise prediction and denoising updates are repeatedly performed in step-size order to form a reconstructed vector sequence. For each step size, the... With corresponding step size Performing dimension-by-dimensional interpolation yields a 96-dimensional residual vector for that step size, which is then aggregated into a reverse residual sequence in step-size order. ;

[0094] Back-engineering the residual sequence The 96-dimensional residual vector for each step is divided into three residual segments according to the dimensional interval. The first residual segment corresponds to the 48-dimensional interval of the inspection item status vector, the second residual segment corresponds to the 16-dimensional interval of the measured record vector, and the third residual segment corresponds to the 32-dimensional interval of the photo and video description vector. For each step, the absolute value of the three residual segments is taken dimension-by-dimensionally and then summed to obtain the residual strength for that step. The residual strength of each step is then correlated with the construction stage identifier. The common input risk output layer, the risk output layer for back-deriving residual sequences Calculate the risk score of the location Location risk score Calculated by metric function:

[0095] ;

[0096] in, To define the number of steps in the preset diffusion step size set, The step size number, The residual dimension number. To reverse the residual sequence In the The 96-dimensional residual vector at step size is the first... Dimensional residuals, For the first Each step size corresponds to a step size weight, which is fixed by the back-calculated update coefficient table. For construction phase identification The phase index obtained from the mapping, , , For risk output layer in stage index The corresponding segment weights, as learnable parameters of the risk output layer, are fixed during the offline training phase and applied to the state vector interval of the inspection item, the measured record vector interval, and the photo / video description vector interval, respectively. To provide a one-dimensional risk score, the risk output layer further aggregates the three residual intensities of each step and outputs a three-dimensional risk source process indicator through a fully connected mapping. Three-dimensional risk source process reminder It corresponds one-to-one with the dimensions of formwork installation, support erection, and pre-pouring verification.

[0097] Risk score of the location With trigger threshold To determine the risk score of the location. Meets the trigger threshold At that time, mark the corresponding location. The identified risk trigger points are output to the extended calculation of the risk linkage graph, and the trigger threshold is determined. During the system deployment phase, based on the historical closed-loop point process inspection records, point samples with qualified review conclusions are selected. The point risk scores of the point samples are calculated and sorted, and the risk scores corresponding to the preset percentiles are taken as the trigger thresholds.

[0098] In this embodiment, step S4 specifically includes:

[0099] Step S4 is used to perform risk linkage expansion on the risk trigger points in the constructed association graph. It transforms the point risk scores output by the diffusion-based anomaly discrimination network into linkage expansion scores that can be propagated along the construction path, and generates linkage expansion results that can be directly sorted in step five. The constructed association graph is denoted as... The set of nodes is denoted as The set of points connected is denoted as The location identifier of the risk trigger point is denoted as The risk score for each location is recorded as follows: The joint investigation extension strength is recorded as The extended score for joint lookup is recorded as The connection type identifier is denoted as The preset attenuation coefficient is denoted as The joint investigation cutoff threshold is denoted as The preset maximum number of expansion layers is denoted as The joint query relationship is recorded as The joint investigation relationship consists of risk trigger point identifiers, connected point identifiers, and connection type identifiers arranged in the order of propagation.

[0100] Input the risk trigger points to construct a relationship graph. In the node set The central positioning point is marked as The node reads the risk score corresponding to the location identifier. ,Will The initial join expansion strength is assigned to the starting node. The starting node is recorded as the zero-level extension node. To support the merging of maximum values ​​for subsequent multi-path access, the maximum known joint query extension score and corresponding joint query relationship are maintained for each visited point identifier. When the point identifier is visited for the first time, the maximum joint query extension score is initialized to zero and the joint query relationship is initialized to empty.

[0101] Adjacency search is performed starting from the zeroth-level extended node, and the adjacency search starts from the set of nodes. Read the point connections adjacent to the current extended node, and extract the point identifiers and connection type identifiers at both ends of the point connection. Limit the adjacency search to propagate only along the point connection. The point connection is limited to the same support chain connection, the same mold shell row connection, and the boundary reinforcement strip connection. When the connection type identifier of the point connection does not belong to the above three categories, stop accessing the point connection.

[0102] Extract the connection type identifier for each point connection obtained through adjacency retrieval. And identified according to the connection type Select the preset attenuation coefficient from the preset attenuation coefficient table. ,Will As a parameter for the extended path of the connection at this point, the preset attenuation coefficient table contains a mapping record from the connection type identifier to the attenuation coefficient, and each attenuation coefficient is a value greater than zero and less than or equal to one.

[0103] With initial joint lookup expansion strength For input, identify the risk trigger point. For each directly connected point, calculate the first-level join extension score and construct the first-level join relationship. For each point marker Connect the points to the linked point markers and read their preset attenuation coefficients. The first-level join expansion score is calculated as follows: and The product, and the risk trigger point is marked. Connecting point identifiers and connection type identifiers are combined in the order of propagation to form the first-level linkage relationship. The first layer of joint investigation relationship Bind to the connected point identifier;

[0104] The connected points in the first layer are sequentially designated as expansion nodes, and expansion is performed layer by layer. The number of expansion layers is denoted as . The zeroth layer represents the nodes corresponding to the risk trigger points. For each layer expansion, the linkage expansion strength of the previous layer's expansion nodes is read as the current propagation input. Adjacency retrieval, connection type identifier extraction, and preset attenuation coefficient selection are performed. The linkage expansion scores of connected points are then updated according to the optimization function.

[0105] ;

[0106] in, To expand the number of layers and satisfy , For the first The point identifier corresponding to the layer extension node. To and The connection points are identified by the point-to-point connection. Connecting point identifiers In the The extended score obtained from the join operation at each layer is initialized to zero before the first calculation and can be updated repeatedly within the same layer. Point identification In the The layer's joint expansion strength, and the zeroth layer satisfies And satisfy for any layer , Connect the points to the point identifier Point of view marker The connection type identifier, Identified by connection type The preset attenuation coefficient is found in the preset attenuation coefficient table. This is a maximum value operator used to retain a larger joint query extension score when multiple joint query extension paths exist for accessing the same connected point identifier.

[0107] After completion After the update, Assigned value And using the previous level join relationship as the path prefix, the current connection type identifier is appended to generate a new multi-level join relationship, if and only if If the score exceeds the maximum linked query extension score currently recorded for the connected point identifier, a new multi-level linked query relationship will be used to replace the linked query relationship for the connected point identifier.

[0108] Set the termination condition for the joint query expansion process to reach the preset maximum expansion layer. Or the joint query expansion strength of the node to be expanded is lower than the joint query cutoff threshold. When the first Layer extension node point identifier satisfy Stop from the point marker. Continuing to expand outwards, for all connected point identifiers generated before the termination condition is met, their maximum joint query expansion score and their corresponding joint query relationship are used as the final result entries, and only one joint query relationship is retained for each connected point identifier.

[0109] The results obtained from the joint investigation expansion process are compiled into the joint investigation expansion results. The joint investigation expansion results include the connected point identifiers, joint investigation expansion scores, and joint investigation relationships. The joint investigation expansion results are used as the sorting input for generating the on-duty review point list in step five.

[0110] In this embodiment, step S5 specifically includes:

[0111] Step S5 converts the joint investigation expansion result output from Step 4 into an executable list of on-duty review points, ensuring that the sorting input simultaneously includes the joint investigation expansion score obtained from the constructed path propagation and the point risk score output by the diffusion-based anomaly discrimination network. The joint investigation expansion result is denoted as... The location identifier of the risk trigger point is denoted as The set to be sorted is denoted as The extended score for joint lookup is recorded as The joint query relationship is recorded as The risk score for each location is recorded as follows: The review sequence queue is denoted as The threshold for the number of locations that can be verified during the shift is recorded as follows: The review order identifier is recorded as The list of checkpoints for the shift is recorded as follows: ;

[0112] Receive the extended results of the join query output from step four. ,Will Read and extract the location identifier and extended score corresponding to each record one by one. Joint investigation relationship Simultaneously, the location identifiers of the risk trigger points are read from the location risk score output records in step three. and their location risk scores And read according to the point identifier index The risk score corresponding to each location marker within the area. This allows the secondary sorting to directly use the location risk score;

[0113] Mark the location of the risk trigger point. Write to the unsorted set and for Write the join extended score and join relation, where the join extended score is assigned a value. Assigning an empty join relation to indicate that it is the starting point for join expansion, and then assigning the join expansion results to the join relation. Each record in the dataset is written into the unsorted set according to its location identifier. When the set to be sorted When records with duplicate location identifiers are found, a merge is performed. During the merge, the maximum join extension score for that location identifier is retained, along with the join relationships that generated the maximum join extension score. To determine which join relationships to retain when the maximum join extension scores are the same, the join relationship level is defined as the join relationship. The number of connection type identifiers is determined, and when the maximum join extension score is reached by multiple records simultaneously, the join relationship with the smaller number of join relationship layers is retained;

[0114] Sets to be sorted A review order queue is generated by sorting the joint investigation scores from highest to lowest. When the scores for joint investigations are the same, the risk score at each location will be used to determine the outcome. Secondary sorting is performed from highest to lowest. When the extended score and the location risk score are the same, the location is sorted in ascending order by the character code of the location identifier to ensure the queue order is reproducible. The review order queue is then checked. Each queue element contains a location identifier, a joint investigation extension score, a location risk score, and a joint investigation relationship, which are used to directly generate a list of locations to be reviewed on duty after interception;

[0115] Read the threshold number of verifiable locations during the shift The threshold number of verifiable locations per shift The parameters are provided by the system configuration or entered and written into the shift parameter table by the quality inspector at the start of the shift. For the review sequence queue Cut off from the front of the team to the back of the team From the queue elements, we obtain the list of checkpoints for the current shift. ,when The number of elements is less than At that time, Write all data into the on-duty review point list Check the list of locations for the shift. Assign review sequence identifiers sequentially according to queue order. The review sequence identifiers are incremented from the beginning, resulting in a final list of review points for the shift. It should include at least location identifiers, review sequence identifiers, and joint investigation relationships, so that quality inspectors can review the locations sequentially according to the review sequence identifiers and simultaneously check the corresponding construction path source locations according to the joint investigation relationships.

[0116] In this embodiment, step S6 specifically includes:

[0117] Step S6 establishes a closed-loop feedback mechanism for the on-duty review actions, enabling the location risk scores and risk source process prompts output by the diffusion-based anomaly detection network to be verified by the review conclusions. The verification results are then written back to the location process inspection record as input for subsequent extraction of phased location features. The on-duty review location list is recorded as follows: The location is marked as The construction phase is marked as The point process inspection record is recorded as follows The risk source process prompts output by the diffusion-type anomaly detection network are recorded as follows: The review conclusion is recorded as The results of the review and inspection items are recorded as follows: The conclusion indicating whether an anomaly has been confirmed is marked as follows: The risk source process prompt confirmation mark is recorded as The verification conclusion of the location is marked as The closed-loop point process inspection record is recorded as follows: ;

[0118] Receive the list of checkpoints for the current shift Verification conclusions corresponding to the markings at each location. Review conclusion The data is transmitted back from the verification terminal, which then displays the location markers. In conjunction with the joint inspection, quality inspectors are required to enter the results of each pre-set verification inspection item. And enter a conclusion indicating whether an abnormality has been confirmed. Review and check the results of the inspection items. The inspection result is represented by a key-value pair sequence of inspection item number and inspection result code, where the inspection result code takes the value of... Indicates failure, value is This indicates a successful conclusion, confirming the absence of any anomalies. Values This indicates that an anomaly has been confirmed, and the value is [value]. This indicates that no abnormalities have been detected.

[0119] Review conclusions Perform structured mapping to generate point verification conclusion markers that correspond one-to-one with the point identifiers. Structured mapping includes a unified coding system for conclusion identifiers and risk source process prompt confirmation identifiers: Directly write the point verification conclusion marker, and simultaneously read the point identifier from the diffusion anomaly detection network. Output of risk source process prompts The risk source process prompt As a three-dimensional vector, the three-dimensional components correspond one-to-one with the dimensions of formwork installation, support erection, and pre-pouring verification. The dimension corresponding to the largest value among the three-dimensional components is used as the prompt dimension index, which is used to check the results of the review items. Obtain underlying evidence comparable to the suggested dimension index, and pre-establish the inspection item-process dimension mapping table in the system. Mapping table Each inspection item number is mapped to one of the following dimensions: formwork installation, support erection, or pre-pouring review. (Mapping table) During the system deployment phase, the process configuration file is imported and generated. This configuration file is stored in key-value pairs of "inspection item number—process dimension identifier" and written into the system parameter library. The result code for the inspection is The inspection item numbers are extracted into a failed set and mapped using a mapping table. Count the number of times the set that failed to pass appears in the dimension corresponding to the suggested dimension index. Furthermore, if the number of occurrences is greater than zero, a confirmation flag will be displayed for the process originating the risk. Set as ,when Or when the number of occurrences is zero, Set as The verification results of the points are marked as binary pairs. ;

[0120] Mark the site verification conclusion. Write point process check record Records that are consistent with both the site markers and the construction phase markers form a closed-loop site process inspection record. During writing, the point identifier is first read from the output record of the diffusion anomaly detection network. Corresponding construction stage identifiers and with As a key in the point process inspection record In the location record item, when there are multiple records that meet the requirements When writing to a record, select the record with the latest timestamp as the writing target, and update the conclusion identifier field of the target record to... Update the risk source process prompt confirmation identifier field to The review timestamp and reviewer identifier are written into the closed-loop field of the record item to mark the completion of the closed loop for the record item. The point process inspection record after completion of the writing is recorded as the closed-loop point process inspection record. and will This serves as the input source for extracting phased point features based on point location identifiers and construction phase identifiers.

[0121] The above description is only a preferred embodiment 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 method for identifying construction quality risk points in closely spaced ribbed floor slabs based on artificial intelligence, characterized in that, include: S1. Obtain on-site inspection information during the installation, support erection, and pre-pouring verification stages of the ribbed beam floor slab formwork. Divide the area into points according to the axis grid coordinates, and write the inspection item records, actual measurement records, and photo and video descriptions into the corresponding points to form a point process inspection record. S2. Extract construction stage identifiers and inspection item statuses from the point process inspection records to form phased point characteristics, and establish point connections based on the structural relationships of support chains, formwork rows, and boundary reinforcement strips to form a structural association diagram; S3. Input the phased point features into the diffusion-type anomaly discrimination network, and perform diffusion back-calculation under the constraints of construction phase identification. Calculate the point risk score of each point's deviation from the qualified point process. Decompose the point risk score according to the dimensions of formwork installation, support erection, and pre-pouring review to form a risk source process prompt. When the point risk score meets the trigger threshold, determine the risk trigger point. S4. Based on the constructed association diagram, perform risk linkage expansion on the risk trigger points, and limit the linkage expansion path to the connection of the same support chain, the connection of the same mold shell, and the connection of the boundary reinforcement strip. Use the risk score of the point as the linkage expansion strength, calculate the linkage expansion score of the connected points along the linkage expansion path, and record the linkage relationship corresponding to the connected points. S5. Based on the threshold of the number of verifiable points for the shift, sort and extract the joint investigation extended scores to generate a list of verifiable points for the shift. The list of verifiable points for the shift includes point identifiers, verification order identifiers, and joint investigation relationships. S6. Receive the review conclusions corresponding to the review point list for the shift, form a review conclusion mark for the point, and write the review conclusion mark for the point into the point process inspection record to obtain the closed-loop point process inspection record.

2. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 1, characterized in that, S1 specifically refers to: Collect on-site inspection information during the stages of formwork installation, support erection, and pre-pouring verification; read the grid coordinates and generate point markers that correspond one-to-one with the grid coordinates; The inspection item records corresponding to the location identifiers are coded in a state according to the preset inspection item number order to form a 48-dimensional inspection item state vector. The measured records corresponding to the point markers are numerically aggregated according to the preset measured field order to form a sixteen-dimensional measured record vector. The photo and video descriptions corresponding to the location markers are vectorized and mapped according to a preset semantic description lexicon to form a 32-dimensional photo and video description vector. The inspection item records, actual measurement records, photo and video descriptions and stage markers are written into the location process inspection record.

3. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 1, characterized in that, S2 specifically refers to: Read the construction stage identifier corresponding to the point identifier from the point process inspection record, and extract the inspection item status vector, actual measurement record vector, and photo and video description vector; The inspection item status vector, measured record vector, and photo and video description vector are concatenated to form a 96-dimensional phased point feature, and the construction stage identifier is 3D one-hot encoded as the stage condition input for the diffusion anomaly discrimination network. Based on the grid coordinates of the location markers and the structural layout information of the ribbed beam floor slab, determine the corresponding support chain markers, formwork row markers, and boundary reinforcement strip markers. Using point markers as nodes, nodes with the same support chain marker, the same mold shell row marker, and the same boundary reinforcement strip marker are connected to form a point connection diagram.

4. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 1, characterized in that, S3 specifically refers to: Before inputting the phased point features into the diffusion anomaly discrimination network, the construction phase identifier is 3D one-hot encoded and then fully connected to generate a 12-dimensional phase condition vector. The diffusion step size identifier is fully connected to generate a 16-dimensional step size condition vector. The 96-dimensional phased point features are concatenated with the 12-dimensional phased condition vector and the 16-dimensional step size condition vector to form a 124-dimensional network input vector, and the network input vector is written into the initial step size of the diffusion backpropagation calculation. A fully connected residual stacking structure without self-attention is used to perform layer-by-layer denoising calculation on the network input vector. The fully connected residual stacking structure consists of four residual blocks in sequence, and each residual block contains two fully connected layers with 256 fully connected neurons. At the block input of each residual block, the stage condition vector and step condition vector are concatenated again, and the residual block output and residual block input are added together and then passed to the next residual block to maintain the diffusion back-reasoning path under the construction stage identification constraint. The outputs of the four residual blocks are input into a dimension-reduced fully connected layer containing 128 fully connected neurons, and a 96-dimensional noise prediction vector is output through a denoising output layer to obtain a reconstructed vector of the phased location features. The inverse residual sequence is formed based on the difference between the phased location features and the reconstructed vector at each diffusion step. The reverse residual sequence is divided into three residual segments according to the inspection item status vector, the measured record vector, and the photo and video description vector. The three residual segments are then mapped through the risk output layer to obtain a one-dimensional point risk score. At the same time, the three residual segments are mapped to obtain a three-dimensional risk source process prompt. The three-dimensional risk source process prompt corresponds one-to-one with the formwork installation dimension, support erection dimension, and pre-pouring review dimension. The risk score of a location is determined by comparing it with the trigger threshold. When the risk score of a location meets the trigger threshold, the corresponding location is identified as a risk trigger location and output to the risk linkage extended calculation of the constructed association graph.

5. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 4, characterized in that, When mapping the three residuals through the risk output layer to obtain the one-dimensional location risk score, the location risk score is calculated according to the metric function, which is specifically: ； in, To define the number of steps in the preset diffusion step size set, The step size number, The residual dimension number. To reverse the residual sequence In the The 96-dimensional residual vector at step size is the first... Dimensional residuals, For the first Each step size corresponds to a step size weight, which is fixed by the back-calculated update coefficient table. For construction phase identification The phase index obtained from the mapping, , , For risk output layer in stage index The corresponding segment weights, as learnable parameters of the risk output layer, are fixed during the offline training phase and applied to the state vector interval of the inspection item, the measured record vector interval, and the photo / video description vector interval, respectively. To provide a one-dimensional risk score, the risk output layer further aggregates the three residual intensities of each step and outputs a three-dimensional risk source process indicator through a fully connected mapping. Three-dimensional risk source process reminder It corresponds one-to-one with the dimensions of formwork installation, support erection, and pre-pouring verification.

6. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 1, characterized in that, S4 specifically refers to: The risk trigger points are input to construct a correlation graph, the nodes corresponding to the risk trigger points are located, and the risk scores of the risk trigger points are read as the initial linkage expansion strength. Adjacency search is performed starting from the node corresponding to the risk trigger point. The adjacency search is limited to propagation only along the point connections. The point connections are composed of connections with the same support chain, connections with the same mold shell, and connections with the boundary reinforcement strip. Node relationships other than the point connections do not participate in the joint search expansion calculation. For each point connection obtained through adjacency retrieval, extract the connection type identifier, select a preset attenuation coefficient based on the connection type identifier, and use the preset attenuation coefficient as the joint query extension path parameter for that point connection; Using the initial joint investigation expansion strength as input, the first-level joint investigation expansion score is calculated for the connected points directly connected to the risk trigger point according to the joint investigation expansion path parameters. The risk trigger point identifier, connected point identifier, and connection type identifier are combined in the propagation order to generate the first-level joint investigation relationship. The first layer of connected points are sequentially set as extension nodes, and the corresponding first layer of joint query extension score is used as the joint query extension strength of the extension node. Adjacency retrieval, connection type identifier extraction, joint query extension path parameter selection and joint query extension score calculation are repeated to generate multi-layer joint query relationships layer by layer, so that the joint query extension score propagates along the construction path of connection with the same support chain, connection with the same mold shell row, and connection with the boundary reinforcement strip. The termination condition for the joint query expansion process is set as the number of expansion layers reaching the preset maximum number of expansion layers or the joint query expansion strength of the expansion node being lower than the joint query cutoff threshold. When the same connected point is accessed through multiple joint query expansion paths, the maximum value of the joint query expansion score of the connected point is taken, and the multi-level joint query relationship that generates the maximum value is determined as the joint query relationship of the connected point. The output includes the identifiers of connected points, the score of the joint query extension, and the joint query relationship. The joint query extension result is used as the sorting input for generating the list of points to be reviewed during the shift.

7. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 6, characterized in that, When the same connected point is accessed through multiple linked search extension paths, an optimization function is used to update the linked search extension score. Specifically, the optimization function is: ； in, To expand the number of layers and satisfy , For the first The point identifier corresponding to the layer extension node. To and The connection points are identified by the point-to-point connection. Connecting point identifiers In the The extended score obtained from the join operation at each layer is initialized to zero before the first calculation and can be updated repeatedly within the same layer. Point identification In the The layer's joint expansion strength, and the zeroth layer satisfies And satisfy for any layer , Connect the points to the point identifier Point of view marker The connection type identifier, Identified by connection type The preset attenuation coefficient is found in the preset attenuation coefficient table. This is a maximum value operator used to retain a larger joint query extension score when multiple joint query extension paths exist for accessing the same connected point identifier.

8. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 1, characterized in that, S5 specifically refers to: Receive the output of the joint investigation and expansion results, and extract the point identifier, joint investigation and expansion score, and joint investigation relationship corresponding to each connected point. At the same time, extract the point identifier of the risk trigger point. Write the location identifier of the risk trigger point into the set to be sorted, and merge records with duplicate location identifiers in the set to be sorted. When merging, retain the maximum join expansion score and retain the join relationship that generated the maximum join expansion score. The set to be sorted is sorted from high to low according to the joint query expansion score. When the joint query expansion scores are the same, they are sorted in a secondary order according to the location risk score from high to low, and a review order queue is generated. Based on the threshold number of reviewable points for the current shift, the review order queue is truncated and assigned review order identifiers in sequence to generate a review point list for the current shift. The review point list for the current shift includes point identifiers, review order identifiers, and linkage relationships.

9. The method for identifying construction quality risk points of closely spaced ribbed beam floor slabs based on artificial intelligence according to claim 1, characterized in that, Step S6 is as follows: Receive the review conclusions corresponding to each location identifier in the on-duty review point list. The review conclusions include the review check item results and a conclusion identifier indicating whether an abnormality is confirmed. The review conclusions are structured and mapped to generate point review conclusion markers that correspond one-to-one with the point identifiers. The point review conclusion markers include a conclusion identifier and a corresponding risk source process prompt confirmation identifier. The point verification conclusion is marked and written into the point process inspection record, which is consistent with the point identifier and the construction stage identifier, to form a closed-loop point process inspection record. The closed-loop point process inspection record is used as the input source for subsequent extraction of point features in stages.

Images