CN-121981343-A - Topology implicit deduction and delay risk prediction method and device for construction site state

CN121981343ACN 121981343 ACN121981343 ACN 121981343ACN-121981343-A

Abstract

The invention provides a topology implicit deduction and delay risk prediction method and device for a construction site state, and relates to the technical field of construction intelligent prediction and risk assessment management. The visual language model is utilized to analyze a field image and a constraint prompt word set, a construction field semantic graph is constructed through logic verification and correction, a visual programming platform is utilized to construct a BIM model of the construction field and analyze the BIM semantic graph, a visual field consistent BIM semantic graph is constructed, virtual-real bidirectional graph matching is then carried out, a global optimal node matching set is solved, two types of semantic graphs are fused according to the global optimal node matching set, a dynamic time-varying semantic graph is constructed, state implicit reverse deduction and delay risk prediction are carried out, and finally a deduction prediction result is mapped to the BIM model, so that visual management of construction progress is realized. The invention can effectively solve the problems of incomplete component identification, low matching degree of virtual and real information and limited progress monitoring stability, and realize intelligent prediction and risk assessment management of the building construction site.

Inventors

WU ZHIMING
XIAO CHENG
LI XIUFANG
LIAN YUXIN
YANG RONGHUA
FANG TIANKE
KANG RUNCHI

Assignees

厦门理工学院

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. The method for predicting the topology implicit deduction and delay risk of the construction site state is characterized by comprising the following steps of: S1, acquiring a construction site image, constructing a constraint prompt word set, driving a visual language model to analyze, and constructing a construction site semantic graph after logic verification and correction; S2, constructing and analyzing a BIM model of a construction site by utilizing a visual programming platform, and generating a BIM semantic map aligned with the construction site semantic map; S3, performing bidirectional graph matching on the construction site semantic graph and the BIM semantic graph, and solving a global optimal node matching set of the weighted bipartite graph; S4, combining the global optimal node matching set, fusing a construction site semantic graph and a BIM semantic graph to construct a dynamic time-varying semantic graph with a time stamp, performing implicit reverse deduction on the blocked node based on building structure topology constraint, performing deferred risk prediction on a progress hysteresis component, and mapping reverse deduction and deferred risk prediction results to a BIM model to realize visual management of construction progress.
2. The method for predicting the topological implicit deduction and deferral risk of a construction site state according to claim 1, wherein the constraint prompt word set comprises five inference phase instructions: The method comprises the steps of a first stage of global priori instruction for driving the visual language model to output a construction stage label of a current scene, a second stage of white list constraint instruction for setting a legal component class set for limiting the visual language model to identify components only in the range of the legal component class set, a third stage of attribute extraction instruction for extracting identifiers, materials, apparent states and two-dimensional pixel coordinate bounding boxes of the components, a fourth stage of relationship inference instruction for limiting interaction relationship types among the components, wherein the interaction relationship types comprise supporting relationships, connecting relationships and host relationships, and a fifth stage of format output instruction for limiting the structural templates of node fields and edge fields.
3. The method for predicting the hidden deduction and delay risk of the topology of the construction site state according to claim 1, wherein the construction process of the construction site semantic map is as follows: inputting the construction site image and the constraint prompt word set into a visual language model to obtain an initial analysis result; Checking the grammar integrity of the initial analysis result by using a JSON analyzer, retrying if the check fails, extracting N component entities of a node list in the initial analysis result if the check succeeds, and calculating a verification score Wherein the assay score The formula of (2) is: ; Wherein, the 、、 Is a weight coefficient; is the total number of component entities; Is the first Confidence of individual component entities; The number of the effective components; Is the stage confidence; Then will And a preset threshold value Comparing if Judging that the output quality is unqualified, discarding the first Data of individual building block entities, if Marking as data to be corrected; Invoking a construction rule base composed of BIM component logic, structure stress rules and construction process knowledge to carry out logic verification and correction on node attributes and topological relations of the data to be corrected, and outputting high-quality structured perception data; construction of a construction site semantic graph comprising components as nodes and topological logic relations among the components as edges based on the high-quality structured perception data Expressed as: ; Wherein, the The node set is a construction site component node set and is used for representing a construction site entity component; the construction site edge set is used for representing the logical dependency relationship among the components; for a construction site node feature matrix, fusing semantic features and geometric coordinate features of the components; The construction site side relationship attribute matrix is used for representing type characteristics of the relationship among the components; The logic checking and correcting process specifically comprises the following steps: First, traversing the relation triples in the data to be corrected Wherein And (3) with Respectively represent components 、 Is provided with a plurality of component nodes, Representing the relation type, checking the relation type according to the construction rule base Whether or not to conform to a component node And (3) with Category labels of (c) And (3) with Engineering logic between the two; If the relation type is a relation triplet of the supporting relation, checking whether the identification direction accords with the gravity transmission level, and executing direction reversal correction according to the bearing priority defined by the construction rule base when the direction is wrong, so as to obtain a component node after the error relation correction; Then, based on the component nodes with the corrected error relation, carrying out pixel geometry verification on the relation triplets by utilizing the bounding box of each component node, namely verifying whether the vertical distance between the bottom ordinate of the supporting component and the top ordinate of the supporting component is within a preset pixel error range aiming at the supporting component belonging to the supporting relation pair in the component nodes, and checking the horizontal projection overlapping degree of the supporting component; if the geometric consistency score is lower than a preset consistency threshold, judging that the geometric consistency score is visual illusion, deleting the supporting member, and obtaining member nodes with the illusion relation removed; finally, introducing a minimal relation complement strategy to make up for potential omission of the visual language model on basic component relation recognition, wherein the method specifically comprises the following steps: Traversing the component nodes with the removed phantom relations, screening the component nodes with the access degree of 0 and the non-bottom base, and checking whether the component types and the space constraints of the component nodes meet the completion requirements so as to complete the foundation support relation missing from the target component nodes on the premise of meeting the completion conditions, thereby obtaining high-quality structured perception data; the complement requirement is that the target member is of the type required to be supported specified by the construction rule base, and that the target member has effective vertical proximity and horizontal projection overlap with the underlying potential support member.
4. The method for predicting the hidden deduction and delay risk of the topology of the construction site state according to claim 1, wherein the construction process of the BIM semantic graph is specifically as follows: based on the internal and external parameter rotation matrix of the virtual camera of the visual programming platform, the orthogonal unit direction vector of the local coordinate system is extracted, namely, the line of sight main axis direction vector Vector of upward direction Vector with right direction ; Calculating coordinates of a center point, a half width, a half height and four corner points of the far cutting surface, wherein the formula is as follows: ; ; ; ; ; ; ; Wherein, the The center point of the far clipping surface; position coordinates of the virtual camera in a BIM model three-dimensional space; The distance of the cutting surface is far; 、 The half width and the half height of the far cutting surface are respectively; 、 The horizontal view angle and the vertical view angle of the virtual camera are respectively; 、 The coordinates of the left and right upper corner points of the far cutting surface are respectively; 、 The coordinates of left and right lower corner points of the far cutting surface are respectively; similarly, the distance of the far clipping surface is replaced by the distance of the near clipping surface, and four corner coordinates of the near clipping surface are calculated; Sequentially connecting eight corner coordinates of the far cutting surface and the near cutting surface into a closed geometrical body according to a space sequence to obtain a virtual view cone entity; Is provided with For the complete set of components of the BIM model, For the internal closed space geometry of the room object where the virtual camera is located, the visibility of the component needs to meet the intersection constraint of the virtual view cone entity and the room space constraint simultaneously to obtain the BIM component node set The expression is: ; Wherein, the Representing a complete set of components of a BIM model The first of (3) A component entity; Is a virtual view cone entity; representing a boundary surface of a room; For BIM component node set Extracting geometric solid of component from arbitrary component in (a) And its centroid For geometric entity Performing uniform scaling to obtain expansion entity ; Defining a physical adjacency discriminant function to perform spatial interference detection on all pairs of expansion entities in a view, wherein the physical adjacency discriminant function is expressed as: ; Wherein, the Is a component node And (3) with Is a physical adjacency discriminant function; 、 Respectively as component nodes And (3) with When (1) expanding entity In this case, the two-member node is determined And (3) with There is a spatial contact; Based on the BIM component node set Combining the component category and engineering logic to construct a BIM edge set comprising a supporting relationship, a host relationship and an adjacent relationship ; For BIM component node set Extracting core information containing component category labels, family and type names and key type attribute parameters, and splicing the core information into a natural language text sequence according to category, attribute and value logic; Mapping the natural language text sequence of each component node and the projection bounding box of the component under the virtual view cone entity into continuous vector representation, namely node feature vector, thereby forming BIM node feature matrix The expression of the node characteristic vector is as follows: ; Wherein, the Is a component node Is defined by the node feature vector of (a); Embedding a model for the pre-trained text; Is a component node Is a natural language text sequence of (1); is a coordinate projection function; Is a component node Is a projection bounding box; represents d-dimensional real numbers; Representing vector stitching; numbering node indexes; adopts a single-heat coding mode to collect opposite edges The relation type of each side is characterized in that the code vector of each side meets the constraint that the dimension of the corresponding relation type is 1 and the other is 0, and the independent heat code vectors of all sides are arranged according to the corresponding relation of nodes to form a BIM side attribute matrix ; Thereby generating a complete BIM semantic graph Expressed as 。
5. The method for predicting topological implicit deduction and deferral risk of a construction site state according to claim 4, wherein the solving process of the global optimal node matching set is as follows: Firstly, calculating the semantic similarity of the node feature vectors in the construction site semantic graph and the BIM semantic graph, wherein the expression is as follows: ; Wherein, the Node feature vector for construction site semantic graph Node feature vector with BIM semantic graph Semantic similarity of (c); Is the modular length of the vector; , ; Based on the neighborhood node set of the construction site semantic graph and the BIM semantic graph, calculating a topological consistency index of the nodes, wherein the expression is as follows: ; Wherein, the Is a topological consistency index; for construction site semantic graph node Is a neighborhood node set; For BIM semantic graph nodes Is a neighborhood node set; Is that With its neighborhood nodes Semantic relationship types between them; Is that With its neighborhood nodes Semantic relationship types between them; represents a maximum value; Representing the number of statistical intersections of edges with the same relationship type in the two-node neighborhood relationship; 、 Respectively represent the collection 、 Is of a size of (2); numbering node indexes; And fusing semantic similarity and topological consistency to obtain the overall matching degree of the nodes, wherein the formula is as follows: ; Wherein, the Node feature vector for construction site semantic graph Node feature vector with BIM semantic graph The node overall matching degree of the (a); 、 Respectively the weight coefficients; To be used for For node matching weight, constructing a weighted bipartite graph, and the expression is: ; ; ; Wherein, the The method is a weighted bipartite graph; A node set for a construction site component; A set of nodes for a BIM component; Matching a weight matrix for the node; Is that And (3) with Is also the node matching weight of (2) First of matrix Line 1 Column elements; then under the constraint condition, adopting Hungary algorithm to solve and optimize weighted bipartite graph to obtain global optimal node matching set The expression is: ; ; Wherein, the In order to match the identification variable(s), Semantic graph node representing construction site Node with BIM semantic graph Establishing a matching relationship; Representing a mismatch; the representation is constrained; representing maximization.
6. The method for predicting the hidden deduction and delay risk of the topology of the construction site state according to claim 1, wherein the generation process of the dynamic time-varying semantic map is specifically as follows: Taking the BIM semantic graph as a reference, combining the global optimal node matching set, and carrying out semantic alignment fusion with the construction site semantic graph to obtain a construction fusion semantic graph with a timestamp mark Thereby obtaining a dynamic semantic graph sequence containing continuous monitoring time ; Wherein, the The method is a fusion node set, namely a set of all nodes in a construction site semantic graph and a BIM semantic graph; The method is a fusion edge set, namely a set of all edges in two graphs; the node feature matrix is a fusion node feature matrix, namely a set of all node features in the two graphs; coding for the attribute matrix of the fusion edge, namely all the edge relation types in the two graphs; The current monitoring time is the current monitoring time; For the moment of time Is used for fusing semantic graphs; Marking the construction state of each fusion node for each fusion node in the dynamic semantic graph sequence to record the construction state of each fusion node at different monitoring moments, so as to obtain state tracks at all moments and generate dynamic time-varying semantic graphs; The construction state comprises a completed construction state and successfully identified on site, a construction state, a progress lag state in which BIM exists but is not detected on site, and an uncertain state with insufficient confidence of shielding or identification.
7. The method for predicting the hidden deduction and deferral risk of a topology in a construction site state according to claim 6, wherein the process of performing the hidden reverse deduction of the state on the blocked node based on the architectural structure topology constraint is specifically as follows: For each fusion node in the dynamic time-varying semantic graph, maintaining the whole evolution track from the construction start to the completion by using a node state updating formula, wherein the node state updating formula is as follows: ; Wherein, the Is a fusion node At the moment of time The updated construction status of (2); Updating operators for the construction state; For the moment of time Node observation information obtained through field perception; is a fusion node At the fusion edge set A topology neighborhood state context in (a); If the blocked node marked as an uncertain state in the dynamic time-varying semantic graph is detected that the adjacent node with the degree of departure in the dynamic time-varying semantic graph is in a completed construction or in-construction state, updating the state of the blocked node into the completed or in-construction state according to the prior physical space of the lower bearing prior to the upper structure; if obtained The node state of the node is the completed construction or the in-construction state, the front support node is queried reversely, if the state of the queried front support node is the progress lag state and the engineering constraint priority is violated, the judgment is made For visual illusion, will Resetting the node state of the node to be in a progress lag state or an uncertain state, and recording a logic conflict log; the process of performing a deferred risk prediction on a progress-retarding component is specifically: Extracting nodes of progress lag states in the dynamic time-varying semantic graphs, and calculating actual lag time difference by combining planned progress time ; Performing sweep effect deduction of directed graph message transfer along the process dependency relationship edges in the graph based on the actual lag time difference; Based on the planned start time, calculating the delay offset of the predicted subsequent process by adopting a cascade delay propagation function and combining the attenuation coefficient of the free time difference to obtain the predicted start time, wherein the formula is as follows: ; ; Wherein, the Is a fusion node Is used for predicting the starting time of the machine; is a fusion node Is set up according to the planned start time; To point along the dependent edge A preposed process node set; 、 Respectively represent the first 、 A plurality of fusion nodes; To delay the transmission attenuation coefficient; Indicating that the maximum value is taken; Representing to take the minimum value; Representing fusion nodes Is a real lag time difference of (2); Representing fusion nodes To the point of Free time differences in the BIM construction plan; When (when) In the time-course of which the first and second contact surfaces, Delay is completely absorbed by buffer, and cascade delay propagation does not occur, when When the excess is transmitted to the rear node.
8. The method for predicting a topological implicit deduction and deferral risk of a construction site state according to claim 5, further comprising calculating a comprehensive consistency index based on the global optimal node matching set to comprehensively evaluate a consistency degree between the construction site state and a BIM model state; The expression of the comprehensive consistency index is as follows: ; ; ; Wherein, the Is a comprehensive consistency index; 、 Is a weight coefficient, satisfies ; The node matching rate is used for representing the finishing or correct identification proportion of the BIM design components on site; Is the structural retention; to meet the number of topologically consistent edges; the total number of the nodes theoretically visible in the current view in the BIM semantic graph is calculated; the total number of theoretical relation edges formed by the matching nodes in the BIM semantic graph is calculated; The number of pairs of nodes that are valid matches; And carrying out self-adaptive classification judgment on the construction site state based on the comprehensive consistency index to obtain the construction site structure state: When (when) When the construction site structure state is judged to be in a structure consistent state with the BIM model design; When (when) When the construction site structure state and the BIM model are judged to be in a local asymmetric state; When (when) When the construction site structure state and the BIM model are judged to be in an obvious asymmetric state; Wherein, the Judging a threshold value for state consistency; a threshold value is determined for the state abnormality.
9. The method for predicting the topological implicit deduction and delay risk of the construction site state according to claim 8, further comprising the steps of analyzing and compensating the construction site structure state to obtain an abnormal node set, wherein the abnormal node set is used for generating construction abnormal alarm information and auxiliary monitoring results, and the method is specifically as follows: If the construction site structure state is the structure consistent state, traversing the global optimal node matching set Respectively extracting the nodes of the semantic graph of the construction site from all the matched node pairs BIM semantic graph node Apparent state properties of (a) And planned construction state Comparison of 、 Outputting a component-level construction state list containing component marking states of finished, under construction or normal states on schedule; if the construction site structure state is a local asymmetric state, carrying out directional difference decomposition on unmatched nodes: Extracting node sets in BIM building blocks Is present but not in Finding BIM components corresponding to construction site nodes, and marking the BIM components as isolated BIM node sets Judging the device as a progress hysteresis component, and generating early warning information; extracting a set of construction site component nodes Is present but not in Finding out the construction site entity corresponding to the BIM semantic graph node, and recording the construction site entity as an isolated site node set Reading out If the label belongs to a preset temporary facility white list, marking the label as a reasonable temporary facility, and if the label does not belong to the temporary facility white list, marking the label as an abnormal newly added object or illegal material stacking; if the construction site structure state is an obvious asymmetric state, executing a three-stage self-adaptive step-by-step compensation mechanism, and if the three stages of self-adaptive step-by-step compensation mechanisms are invalid, triggering the judgment of the degradation of the hierarchy; the three-stage self-adaptive step-by-step compensation mechanism comprises the following steps: the first stage of compensation is the operation of identifying interference nodes and correcting weight, i.e. extracting the semantic map of the construction site The nodes of the medium class labels belonging to { construction machinery, operators, sundries }, are defined as interference node sets Introducing attenuation factors Weakening the matching weight of the interference node, and obtaining the comprehensive similarity after correction The expression is: ; Wherein, the Node feature vector for construction site semantic graph Node feature vector with BIM semantic graph The node overall matching degree of the (a); The matching nodes are the construction site semantic graphs; Based on the corrected integrated similarity Updating the matching weight matrix The method comprises the steps of carrying out matching set of global optimal nodes of a weighted bipartite graph, calculating comprehensive consistency indexes, and re-judging the structural state of a construction site, carrying out directional difference decomposition on unmatched nodes in a re-judging result for nodes in a local asymmetric state, carrying out second-stage compensation for nodes in a obviously asymmetric state in the re-judging result; the second phase compensation is BIM priori guided backtracking identification operation, namely extracting isolated BIM node set The theoretical projection bounding box coordinates corresponding to each node and semantic labels thereof automatically generate reverse guidance prompt words through a text splicing operator, then input the reverse guidance prompt words and the construction site images into a visual language model together, and update a node set of a construction site semantic graph if the model detects a target member with a confidence threshold higher than a preset confidence threshold Sum edge set If the target component is not detected or the confidence is insufficient, entering a third stage for compensation; the third stage of compensation is to judge the level degradation operation, i.e. extract the semantic graph of the construction site Global graph feature vector of (a) , Generated by a graph embedding method and used for representing macroscopic features of the whole field, and calculating Phase characteristics of each preset construction phase of BIM model The cosine similarity of (2) to find the best matched construction stage and output the macroscopic construction stage judgment result The formula is: ; Wherein, the Representing the index corresponding to the maximum value ; Is the first in BIM model Feature vectors of the preset construction stage; Representing the modular length of the vector; And defining the isolated BIM node set and the isolated site node set after the three-stage compensation processing as an abnormal node set.
10. A device for predicting a risk of topology implicit deduction and deferral of a construction site state, for implementing the method for predicting a risk of topology implicit deduction and deferral of a construction site state according to any one of claims 1 to 9, comprising: the construction site semantic graph construction unit is used for collecting construction site images, constructing a constraint prompt word set, driving a visual language model to analyze, and constructing a construction site semantic graph after logic verification and correction; The BIM semantic map construction unit is used for constructing and analyzing a BIM model of a construction site by utilizing the visual programming platform to generate a BIM semantic map aligned with the construction site semantic map; the bidirectional graph matching unit is used for carrying out bidirectional graph matching on the construction site semantic graph and the BIM semantic graph, and solving a global optimal node matching set of the weighted bipartite graph; The deduction prediction unit is used for combining the global optimal node matching set, fusing the construction site semantic graph and the BIM semantic graph to construct a dynamic time-varying semantic graph with a time stamp, performing implicit reverse deduction on the blocked node based on the topological constraint of the building structure, performing deferred risk prediction on the progress hysteresis component, and mapping the reverse deduction and deferred risk prediction result to the BIM model so as to realize visual management of the construction progress.

Description

Topology implicit deduction and delay risk prediction method and device for construction site state Technical Field The invention relates to the technical field of intelligent prediction management and risk assessment of building construction, in particular to a method and a device for predicting topology implicit deduction and deferred risk of a construction site state. Background In the field of building engineering, a building information model (Building Information Modeling, BIM) technology realizes collaborative management of a project full life cycle by integrating geometric information and construction data. The construction stage utilizes BIM as progress control carrier to combine camera equipment and degree of depth learning algorithm discernment scene key factor, judge the construction progress through comparing perception data and model component, be the research focus of current intelligent monitoring. However, existing construction site state monitoring techniques still have a number of drawbacks in practical engineering applications. Firstly, under a complex construction environment, the problems of incomplete identification of component information, missed detection, false detection or state misjudgment and the like are easily caused by shielding, illumination change, construction equipment movement and frequent personnel movement. Secondly, the traditional progress management focuses on design models and preset plans, and has limited direct perception capability on-site dynamic changes, so that inconsistent phenomenon is easy to occur in actual construction of on-site states and model preset information, and the actual construction current situation is difficult to reflect by judging only by means of the models. In addition, the existing virtual-real information matching method focuses on feature alignment, is limited by the locality of the field image and the integrity of model information, and has room for improvement in terms of crossing member granularity and semantic gap of pixel recognition results. The information asymmetry phenomenon can influence the accuracy of virtual-real comparison, and the stability and the interpretability of automatic monitoring are limited in the face of information loss or dynamic interference. In view of this, the present application has been made. Disclosure of Invention The invention aims to provide a topology implicit deduction and delay risk prediction method and device for a construction site state, which are used for solving at least one of the problems of incomplete component identification, low matching degree of virtual and real information, limited progress monitoring stability and the like caused by complex construction site environment in the prior art. In order to solve the technical problems, the invention is realized by the following technical scheme: A topology implicit deduction and deferred risk prediction method of a construction site state comprises the following steps: S1, acquiring a construction site image, constructing a constraint prompt word set, driving a visual language model to analyze, and constructing a construction site semantic graph after logic verification and correction; S2, constructing and analyzing a BIM model of a construction site by utilizing a visual programming platform, and generating a BIM semantic map aligned with the construction site semantic map; S3, performing bidirectional graph matching on the construction site semantic graph and the BIM semantic graph, and solving a global optimal node matching set of the weighted bipartite graph; S4, combining the global optimal node matching set, fusing a construction site semantic graph and a BIM semantic graph to construct a dynamic time-varying semantic graph with a time stamp, performing implicit reverse deduction on the blocked node based on building structure topology constraint, performing deferred risk prediction on a progress hysteresis component, and mapping reverse deduction and deferred risk prediction results to a BIM model to realize visual management of construction progress. Preferably, the constraint prompt word set includes five inference phase instructions: The method comprises the steps of a first stage of global priori instruction for driving the visual language model to output a construction stage label of a current scene, a second stage of white list constraint instruction for setting a legal component class set for limiting the visual language model to identify components only in the range of the legal component class set, a third stage of attribute extraction instruction for extracting identifiers, materials, apparent states and two-dimensional pixel coordinate bounding boxes of the components, a fourth stage of relationship inference instruction for limiting interaction relationship types among the components, wherein the interaction relationship types comprise supporting relationships, connecting relationships and host relationships, and a f