4r system method for ai application research and development of construction engineering site management

Abstract

The application relates to the technical field of intelligent construction, artificial intelligence and engineering safety management technology, and discloses a 4R system method for AI application research and development of building engineering site management, which comprises the following steps: S1, recognizing all things, acquiring site images, videos and multi-source sensing data, identifying all physical elements and constructing a standardized data sensing base; S2, recognizing states, quantifying entity space-time and semantic attributes, tracking state evolution, constructing a global state matrix and realizing digital mapping; S3, recognizing relationships, calculating entity space topology and time coordination relationship and generating a global interaction relationship matrix; and S4, recognizing business, aligning the matrix with an external reference dictionary, generating a multi-dimensional business matrix, outputting intervention instructions when an abnormality occurs and forming a management closed loop. The 4R system method for AI application research and development of building engineering site management has the advantages of systematic theoretical innovation and standardized architecture design.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technologies of intelligent construction, artificial intelligence and engineering safety management, specifically a 4R system method for the research and development of AI applications for on-site management of construction projects. Background Technology

[0002] Currently, the global construction industry is at a critical stage of transformation towards "Construction Industry 4.0". New-generation information technologies, represented by digital twins, cyber-physical systems, and artificial intelligence, are being gradually applied to core business aspects such as safety management, progress monitoring, and cost control at construction sites. They are becoming an important driving force for the digital and intelligent upgrading of traditional engineering construction models. Construction sites have complex characteristics of high dynamism, strong interaction, and multi-variable coupling. Building a scientific and systematic AI application research and development system is the core foundation for realizing intelligent management of all elements and the entire life cycle of construction sites.

[0003] Currently, the application of artificial intelligence technology in the field of construction site management has made some progress. However, in the process of technology research and development and engineering implementation, there are still several problems to be solved. At the data perception level, the targets at construction sites are complex and severely obscured. The construction of high-quality instance segmentation datasets relies on a large amount of manual annotation, which is costly and difficult to cover long-tail targets, thus limiting the generalization ability of the model. At the state tracking level, existing visual tracking algorithms are prone to frequent changes in target identity under conditions of dust interference, severe target obscuration, or violent mechanical rotation, making it difficult to continuously capture the dynamic state of the target. At the spatial analysis level, two-dimensional images have perspective distortion problems. Traditional alarm mechanisms based on pixel bounding box overlap ignore the true three-dimensional volume and relative motion vector of the target, which easily generates false alarms. At the business transformation level, the coupling degree between the underlying visual analysis logic of AI and national and industry statutory standards and Building Information Modeling (BIM) benchmarks needs to be improved, and information silos are easily formed between the technology and management links.

[0004] Furthermore, the industry lacks a systematic underlying theoretical support, and most AI algorithm models exhibit fragmented characteristics, making it difficult to abstract complex on-site risks into computable mathematical logic and effectively establish a complete closed loop from raw data to engineering decisions. Therefore, a 4R system approach for the research and development of AI applications in construction engineering site management is proposed to solve the above problems. Summary of the Invention

[0005] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a 4R system approach for the research and development of AI applications in construction engineering site management. It boasts advantages such as systematic theoretical innovation and standardized architecture design. Through a 4R decoupling mechanism, it abstracts highly dynamic and complex on-site physical elements into a generalized entity matrix and a business decision matrix, providing a universal standard mathematical foundation for intelligent perception of all elements and highly sophisticated digital twins. It also optimizes the allocation of perception and computing resources, solving the problems of existing technologies, such as a lack of systematic underlying theoretical support within the industry, fragmented AI algorithm models, difficulty in abstracting complex on-site risks into computable mathematical logic, and the inability to effectively establish a complete closed loop from raw data to engineering decisions.

[0006] Technical solution The technical solution of this invention to solve the above-mentioned technical problems is as follows: A 4R system method for the research and development of AI applications for on-site management of construction projects, comprising the following steps: S1, Recognize Everything: Acquire raw images, video data, and multi-source heterogeneous sensor data from the construction site, automatically segment and identify all physical elements on site, including personnel, machinery, materials, and environment, and abstract unstructured physical elements into a discrete generalized entity target matrix to form a standardized data perception base. S2. State Identification: Based on the generalized entity target matrix, high-dimensional feature quantization is performed on the spatiotemporal and semantic attributes of each entity. The time dimension is introduced to realize the evolution tracking of entity state. A diagonal global state matrix containing the heterogeneous state features of each entity at time t is constructed to complete the digital mapping of the physical world. S3. Relationship Identification: Based on the global state matrix, solve the spatial topological relationships and temporal coordination relationships between all entities in the scene, and construct a global interaction relationship matrix with row and column dimensions equivalent to the number of on-site targets, providing computable relationship features for the spatiotemporal interaction of targets. S4. Business Identification: The global state matrix and global interaction relationship matrix are logically aligned with a pre-built external benchmark dictionary that includes a management benchmark dictionary and a regulatory benchmark dictionary to generate a global business matrix covering the dimensions of safety, quality, schedule, and cost. When a business characteristic exceeds a preset threshold, an executable proactive intervention instruction is output, forming a complete closed loop from physical perception to management decision-making.

[0007] The beneficial effects of this invention are: 1) This 4R system approach for AI application development in construction engineering site management differs from the previous development dilemma of relying heavily on breakthroughs in single algorithms or fragmented model stacking. It breaks through by constructing a 4R theoretical framework specifically designed for construction sites. Through the decoupling mechanism of "recognizing everything", "recognizing status", "recognizing relationships" and "recognizing business", it abstracts the highly dynamic and complex physical elements of the site into a generalized entity matrix and a business decision matrix. This system greatly optimizes the allocation of perception computing resources and provides a standard mathematical foundation with strong universality for full-element intelligent perception and high-level digital twins.

[0008] 2) The 4R system method for the research and development of AI applications for construction site management has innovatively reconstructed the underlying logic. Under the visual spatiotemporal perception architecture, "state" and "relationship" are defined as a quantified state matrix and a global interaction relationship matrix. This mechanism has pushed the identification of risks on construction sites from vague "qualitative observation" to high-precision "quantitative calculation", and transformed post-event handling into precise intervention with scientific foresight.

[0009] 3) The 4R system method for AI application development in construction site management has extremely high software and hardware compatibility. It does not require major changes to the existing sensing hardware of the construction site. It can be connected to common monitoring terminals through the efficient empowerment of pure software algorithms.

[0010] Based on the above technical solution, the present invention can be further improved as follows.

[0011] Further, step S1 specifically includes: S1.1, constructing a generalized entity target matrix containing four categories of elements: personnel, machinery and vehicles, materials and protective equipment, and environment and static facilities; S1.2, performing brightness adjustment, contrast adjustment, and geometric transformation preprocessing on the original image or video data, and removing highly repetitive or similar frames by calculating image perceptual hash features to form a purified construction scene image dataset; S1.3, for targets with regular shapes, inputting the purified image into the Grounding DINO model for open-vocabulary target detection, using the DINOv2 model to extract foreground visual features and clustering to generate initial positioning anchors, and inputting the bounding box information and initial positioning anchors as joint conditional cues into the SAM visual basic model to output a high-precision pixel-level instance segmentation mask; S1.4, performing confidence evaluation and multi-scale enhanced consistency verification on the instance segmentation results, filtering high-confidence pseudo-labels with confidence scores higher than a preset threshold, and combining a very small number of manually labeled long-tail complex samples with a Mask... The R-CNN architecture is used for semi-supervised iterative training to generate an instance segmentation model with engineering generalization capabilities; S1.5, for targets with irregular shapes, a fully convolutional semantic segmentation network is used to extract global contextual features of the image and perform pixel-level multi-classification prediction to supplement background semantic information; S1.6, all entity recognition results and background semantic information are summarized and filled into the generalized entity target matrix.

[0012] Furthermore, step S2 specifically includes: S2.1, constructing a diagonal global state matrix at time t, where each diagonal element corresponds to the state features of an entity, and configuring differentiated state attribute domains according to the entity type; S2.2, using a three-dimensional manifold space mapping method to establish the perspective projection transformation relationship between image pixel coordinates and real-world coordinates, introducing manifold constraints on the physical surface of the construction site to achieve monocular visual absolute scale recovery, and quantifying the three-dimensional coordinate domain and geometric dimension parameters of the entity; S2.3, embedding multi-target cross-frame tracking logic in the large model inference network, and introducing multi-dimensional evidence joint modeling... The entity ID is continuously bound across frames, and the motion vector features of the entity are calculated. The motion vector features include instantaneous motion velocity and acceleration vectors. S2.4 For personnel and mechanical entities, microscopic action semantic features are extracted through spatiotemporal graph convolutional neural networks. For material and component entities, the life cycle stage is determined based on the target appearance evolution features by combining a multi-stage appearance evolution network with an ordinal regression judgment operator. S2.5 The visual tensor and infrared thermal imaging and ultrasonic multi-source heterogeneous sensing data are fused to extract the quantification parameters of material and component defects, and complete the closed-loop quantization of the full-element state matrix.

[0013] Further, step S3 specifically includes: S3.1, constructing a global interaction relationship matrix at time t, where each element represents the objective spatiotemporal interaction relationship between two entities; S3.2, defining relationship feature extraction operators and relationship feature calculation rules, and extracting the dimensionality-reduced feature submatrices from the global state matrix as calculation parameters; S3.3, based on the entity's three-dimensional coordinate domain, geometric dimension parameters, and feedforward motion vectors, constructing a three-dimensional dynamic safety envelope surface equation with time parameters as variables, and calculating the precise Euclidean distance and topological interference features between entities, where the topological interference features include topological features such as inclusion, intersection, or separation; S3.4, extracting the entity's action semantic classification variables, life cycle stage variables, and corresponding timestamp sequences from the global state matrix, and calculating the absolute coordination time of strongly coupled processes and the leading-lag topological logic between the working subjects; S3.5, filling the spatial relationship features and temporal relationship features into the spatial submatrix and temporal submatrix of the global interaction relationship matrix, respectively.

[0014] Furthermore, step S4 specifically includes: S4.1 Constructing a global business matrix composed of multiple business feature vectors, each business feature vector corresponding to a specific engineering management business dimension; S4.2 Establishing a multi-source driven external benchmark dictionary, including a management benchmark dictionary integrating BIM design and project management parameters, and a regulatory benchmark dictionary encapsulating mandatory provisions of industry regulations; S4.3 Using state selection operators and business relationship selection operators, filtering out discrete features required for specific businesses from the global state matrix and global interaction relationship matrix; S4.4 Constructing a business feature judgment mapping function, logically aligning the filtered discrete features with the external benchmark dictionary to generate feature vectors for each business dimension; S4.5 Summarizing all business feature vectors to form a global business matrix, triggering corresponding active intervention commands when the value of a business feature vector exceeds a preset threshold or logical red line.

[0015] Furthermore, the three-dimensional manifold space mapping method in step S2.2 is as follows: establish a perspective projection transformation equation including the camera intrinsic matrix, the camera extrinsic matrix, and the axial depth factor; perform homography transformation on the support surface manifold equation for the ground target and solve for the absolute horizontal coordinates in reverse; perform elevation benchmark manifold equation on the BIM model design axis for the high-altitude target and solve for the absolute scale factor in reverse; project the point set within the mask in reverse to the world coordinate system to construct a three-dimensional point cloud envelope, and quantize the absolute coordinate domain and geometric dimension parameters of the entity by solving for the minimum circumscribed cuboid.

[0016] Furthermore, the multidimensional evidence joint modeling method in step S2.3 specifically involves: fusing the cosine distance of the target appearance features with the motion smoothness based on optical flow field estimation to construct a multidimensional association cost matrix; using the Hungarian algorithm for bipartite graph optimal matching to achieve continuous cross-frame binding of entity IDs; and inversely calculating the instantaneous motion velocity and acceleration vector of the entity based on the coordinate changes of the entity in consecutive frames.

[0017] Furthermore, the method for calculating the topological interference features in step S3.3 is as follows: performing spatial volume intersection integration on the three-dimensional dynamic safety envelope surface of multiple entities to accurately determine the inclusion, intersection, or separation topological relationships between entities; and mapping the continuous spatial distance features and discrete topological interference features together to the spatial subarray of the global interaction relationship matrix.

[0018] Furthermore, the business feature determination mapping function can generate six types of business feature vectors: personnel compliance and regional safety, danger distance early warning, building entity quality defect monitoring, progress and production efficiency supervision, dynamic cost accounting and resource statistics, and environmental supervision and facility integrity monitoring; the proactive intervention instructions include on-site sound and light early warning, equipment linkage control, resource scheduling and emergency response activation instructions. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention; Figure 2 This is a diagram illustrating the 4R theory and technology framework of this invention; Figure 3 This is a flowchart of Embodiment 2 of the present invention; Figure 4 This is a flowchart of Embodiment 3 of the present invention; Figure 5 This is a flowchart of Embodiment 4 of the present invention; Figure 6 This is a flowchart of Embodiment 5 of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1, by Figure 1-2This paper presents a 4R system method for the development of AI applications in construction site management. The 4Rs of this invention include recognizing everything, recognizing status, recognizing relationships, and recognizing business. The method specifically includes the following steps: S1, Recognizing everything: Acquiring raw images, video data, and multi-source heterogeneous sensor data of the construction site, automatically segmenting and identifying all physical elements on site, including personnel, machinery, materials, and environment, and abstracting unstructured physical elements into a discrete generalized entity target matrix to form a standardized data perception foundation; S2, Recognizing status: Based on the generalized entity target matrix, performing high-dimensional feature quantization on the spatiotemporal and semantic attributes of each entity, introducing a time dimension to track the evolution of entity status, and constructing a heterogeneous status of each entity at time t. The features of the diagonal global state matrix complete the digital mapping of the physical world; S3, Relationship Identification: Based on the global state matrix, the spatial topological relationships and temporal coordination relationships between all entities in the scene are calculated, and a global interaction relationship matrix with row and column dimensions equivalent to the number of on-site targets is constructed, providing computable relationship features for the spatiotemporal interaction of targets; S4, Business Identification: The global state matrix and global interaction relationship matrix are logically aligned with a pre-built external benchmark dictionary containing management benchmark dictionaries and regulatory benchmark dictionaries to generate a global business matrix covering the dimensions of safety, quality, schedule, and cost. When the business features exceed the preset threshold, executable proactive intervention instructions are output, forming a complete closed loop from physical perception to management decision-making.

[0022] This application constructs a systematic 4R theoretical framework. Through a decoupling mechanism of recognizing everything, recognizing states, recognizing relationships, and recognizing business, it abstracts the highly dynamic and complex physical elements of construction sites into a standardized matrix structure. This provides a universal standard mathematical foundation for intelligent perception of all elements and highly digital twins, solving the problems of fragmented AI algorithm models and difficulty in abstracting complex on-site risks into computable mathematical logic in existing technologies. At the same time, it opens up a complete closed loop from raw data perception to engineering management decision-making, realizing intelligent management of all elements and the entire life cycle of construction sites. This significantly improves the efficiency and accuracy of on-site management, reduces the cost and risk of manual management, and provides solid theoretical and technical support for the digital transformation of the construction industry.

[0023] Example 2, by Figure 3This embodiment, based on Embodiment 1, provides a detailed explanation of step S1. Step S1 specifically includes: S1.1 Constructing a generalized entity target matrix containing four categories of elements: personnel, machinery and vehicles, materials and protective equipment, and environment and static facilities; S1.2 Preprocessing the original image or video data by adjusting brightness, contrast, and performing geometric transformations, and removing highly repetitive or similar frames by calculating image perceptual hash features to form a purified construction scene image dataset; S1.3 For targets with regular shapes, inputting the purified image into an open-vocabulary target detection model (Grounding DINO) based on a base model for open-vocabulary target detection, using a self-supervised visual Transformerv2 model (DINOv2) to extract foreground visual features and cluster them to generate initial positioning anchors, and using the bounding box information and initial positioning anchors as joint conditional prompts input to the Segment Anything model. Model (SAM), outputting a high-precision pixel-level instance segmentation mask; S1.4, performing confidence evaluation and multi-scale enhanced consistency verification on the instance segmentation results, filtering out high-confidence pseudo-labels with confidence scores higher than a preset threshold, and combining a very small number of manually labeled long-tailed complex samples, using a Mask Region-Convolutional Neural Network (Mask R-CNN) architecture for semi-supervised iterative training to generate an instance segmentation model with engineering generalization capabilities; S1.5, for irregularly shaped targets, using a fully convolutional semantic segmentation network to extract global contextual features of the image, performing pixel-level multi-classification prediction to supplement background semantic information; S1.6, summarizing all entity recognition results and background semantic information, and filling them into the generalized entity target matrix.

[0024] The expression for the generalized entity target matrix in step S1.1 is as follows: , in They represent: Personnel elements, such as workers and managers; Mechanical and vehicle elements, such as excavators, dump trucks, tower cranes, etc. Materials, physical components, and protective equipment, such as various building components and labor protection materials; Environmental and static facility elements, such as areas, on-site roads, fences, protective nets, openings, edges, and other static entities. In a specific scenario, it is assumed that the scenario includes... If there are multiple targets, then the scene target matrix can be represented as: In step S1.3, regular-shaped targets include workers, excavators, tower cranes, etc.; in step S1.4, the confidence assessment includes Grounding DINO and SAM prediction scores being greater than the threshold and feature cosine similarity meeting the standard; in step S1.5, irregular-shaped targets include wide-area sky, muddy environment, irregular construction waste, etc.

[0025] This embodiment refines the detailed implementation rules for the entire process of "identifying everything." It achieves standardized classification and management of all elements at the construction site by building a matrix based on four entity categories. Relying on image preprocessing and perceptual hashing deduplication technology, it effectively simplifies the data volume, removes invalid and redundant data, and improves the efficiency of subsequent model inference and training. It integrates Grounding DINO, DINOv2, and SAM multi-vision large-scale models to construct an automated pre-annotation pipeline, significantly reducing the workload and time cost of traditional pixel-level manual annotation. Through confidence verification and semi-supervised training, it efficiently utilizes pseudo-labels and a small number of manually annotated samples, significantly improving the model's generalization and recognition capabilities for long-tail conditions such as construction site occlusion, lighting changes, and complex terrain. It adopts a differentiated segmentation strategy for regular and irregular targets, balancing refined instance segmentation with global background semantic completion, ensuring that all entities of different sizes, regular and irregular elements at the construction site can be completely and accurately identified. This provides clean, complete, and highly accurate standardized input data for subsequent state quantification, relationship calculation, and business analysis.

[0026] Example 3, by Figure 4 This embodiment, based on Embodiment 2, provides a detailed explanation of step S2. Step S2 specifically includes: S2.1, constructing a diagonal global state matrix at time t, where each diagonal element corresponds to the state features of an entity, and configuring differentiated state attribute domains according to the entity type; S2.2, using a three-dimensional manifold space mapping method to establish the perspective projection transformation relationship between image pixel coordinates and real-world coordinates, introducing manifold constraints on the physical surface of the construction site to achieve monocular visual absolute scale recovery, and quantifying the three-dimensional coordinate domain and geometric dimension parameters of the entity; S2.3, embedding multi-target cross-frame tracking logic in the large model inference network. The system introduces multi-dimensional evidence joint modeling to achieve continuous cross-frame binding of entity IDs and calculates the motion vector features of entities, including instantaneous motion velocity and acceleration vectors; S2.4, for personnel and mechanical entities, microscopic action semantic features are extracted through spatiotemporal graph convolutional neural networks; for material and component entities, the life cycle stage is determined based on the target appearance evolution features by combining a multi-stage appearance evolution network with ordinal regression judgment operators; S2.5, by fusing visual tensor and infrared thermal imaging, ultrasonic multi-source heterogeneous sensing data, the system extracts the quantification parameters of material and component defects and completes the closed-loop quantization of the full-element state matrix.

[0027] In step S2.1 Diagonal global state matrix at time step The state matrix of each target This constitutes the quantitative expression for the full target state of the engineering site: ; in, It is a diagonal matrix, because Different target types, state diagonal matrix Presenting diverse expressions: when hour, It represents features such as the coordinate domain of the person, motion vectors, and semantics of skeletal movements (e.g., walking, smoking, working).

[0028] when hour, It characterizes features such as the mechanical coordinate domain, motion vectors, and semantics of mechanism actions (e.g., filling, unloading, and no-load).

[0029] when hour, Characterizes material coordinate domain, dimensional parameters, life cycle stage (e.g., idle, hoisted, installed), and defect quantification parameters (e.g., crack width, exposed reinforcement area, damage size, etc.).

[0030] when hour, It characterizes features such as the coordinate domain of the environment and static facilities.

[0031] In this embodiment, the three-dimensional manifold space mapping method in step S2.2 is as follows: establishing a perspective projection transformation equation that includes the camera intrinsic parameter matrix, the camera extrinsic parameter matrix, and the axial depth factor; specifically, for a certain entity... The discrete entity pixel mask and corresponding bounding box output from step S1 are received as input, and the image pixel coordinates are established. With real-world coordinates Perspective projection transformation equations between , in For the camera intrinsic parameter matrix, The calibrated camera extrinsic matrix, This is the axial depth factor. Furthermore, strong physical surface manifold constraints are introduced at the construction site to achieve absolute scale recovery for monocular vision: Establish simultaneous manifold equations for the support surface for the surface target Inverse homography transformation is used to solve for the absolute horizontal coordinates of the workers and mobile machinery. For high-altitude targets, the elevation benchmark manifold equations of the axes are designed using the BIM model. Inverse calculation of the unique absolute scale factor Finally, the entire set of points within the mask is inversely projected using the aforementioned manifold mapping operator to construct a 3D point cloud envelope of the entity surface in the world coordinate system. The global state matrix is ​​then precisely quantized by solving for the minimum bounding box (OBB) of this point cloud. middle absolute coordinate domain With geometric dimension parameter domain .

[0032] In this embodiment, step S2.3 addresses the issue of identity ID jumps caused by dust interference, severe target occlusion, or violent mechanical rotation, which are common problems at construction sites. Multi-dimensional evidence joint modeling is introduced. The multi-dimensional evidence joint modeling method involves fusing the cosine distance of target appearance features with motion smoothness based on optical flow field estimation to construct a multi-dimensional association cost matrix. Specifically, by fusing the similarity of target appearance features with optical flow motion smoothness, discrete entities are continuously bound to IDs, thereby extracting the "motion vector" feature from the target state matrix. For a given entity... Construct a multidimensional correlation cost matrix Calculate the cosine distance matrix of the appearance features. Compared with the motion smoothness cost matrix based on optical flow field estimation ,definition ,in and It uses adaptive weights; the Hungarian algorithm is employed for optimal bipartite graph matching to achieve cross-frame binding of entity IDs, thereby calculating the instantaneous velocity and acceleration vectors of the entities. , .

[0033] In step S2.4, for a certain entity The system receives the target region tensor (number of color channels, pixel height and width) of the entity in step S1 and the sequence of key topological points extracted as input. It consists of two parts: behavior-oriented behavior semantic spatiotemporal graph convolutional modeling and a multi-stage appearance evolution network oriented towards functional states. On the one hand, the key points of the worker's skeleton or the hinge points of the mechanical moving mechanism are constructed into a spatiotemporal graph in non-Euclidean space within the sliding time window, and the graph convolution feature extraction operator is used. (in The adjacency matrix is ​​used to characterize the microscopic connection topology of entities. The normalization degree matrix, For the target, the original set of coordinates of all skeletal keypoints (or mechanical articulation points) during this time period. Spatial partitioning and aggregation, along with one-dimensional temporal sliding window convolution, are performed on the feature transformation weights to map and output specific micro-action semantic variables. ; On the other hand, considering the unidirectional physical degradation or iteration characteristics of surface texture, color, and shape exhibited by materials and physical components (such as cast-in-place concrete and steel sections awaiting installation) as the process progresses, a multi-stage appearance evolution network is constructed. This network extracts high-dimensional visual feature vectors representing the target region. And introduces an ordinal regression decision operator. Apply physical constraints to the process stage to which the entity belongs.

[0034] in, For the network to determine the first The feature weight matrix learned at each lifecycle stage is used to process high-dimensional visual features. Adaptive weighting is applied to key texture and morphological information within the model; For the corresponding number The bias term for each stage is used to adjust the basic judgment threshold for that physical evolution stage. This enables a multi-stage image classification model to compare the appearance evolution of materials and entities, determine their lifecycle stages, and ultimately accurately decouple and populate the lifecycle stage variables. .

[0035] In step S2.5, the disease characteristic parameters include crack width, exposed reinforcement area ratio, and damage volume.

[0036] Example 4, by Figure 5 This embodiment, based on Embodiment 3, provides a detailed explanation of step S3. Step S3 specifically includes: S3.1, constructing a global interaction relationship matrix at time t, where each element represents the objective spatiotemporal interaction relationship between two entities; S3.2, defining relationship feature extraction operators and relationship feature calculation rules, and extracting the dimensionality-reduced feature submatrices from the global state matrix as calculation parameters; S3.3, based on the entity's three-dimensional coordinate domain, geometric dimension parameters, and feedforward motion vectors, constructing a three-dimensional dynamic safety envelope surface equation with time parameters as variables, and calculating the precise Euclidean distance and topological interference features between entities, including topological features such as inclusion, intersection, or separation; S3.4, extracting the entity's action semantic classification variables, lifecycle stage variables, and corresponding timestamp sequences from the global state matrix, and calculating the absolute coordination time of strongly coupled processes and the leading-lag topological logic between the main working entities; S3.5, filling the spatial relationship features and temporal relationship features into the spatial submatrix and temporal submatrix of the global interaction relationship matrix, respectively.

[0037] Specifically, the global interaction matrix at time t in step S3.1 is as follows: At any time, for all within the scene One objective, define a global relation matrix. The rows and columns of the matrix represent the number of targets on-site. The expression is as follows: in, Representative scenarios Goals and Objective spatiotemporal interaction matrix between targets (when When the matrix is ​​empty, it is calculated from the target state matrix.

[0038] Calculate the spatiotemporal interaction matrix between targets based on the state matrix of each target. ,make For the goal The total number of interactions between them is calculated as follows: in, Indicate target In the The relational eigenvalues ​​under a relation are mathematically expressed as follows: in, As a feature extraction operator, specific feature scalars or vectors (such as extracting only 3D coordinates and lifecycle stages) are extracted from each target state matrix to form a dimensionality-reduced feature submatrix, which serves as the feature matrix. Calculation parameters; In the first The rules for calculating relation features under type relations calculate different relation features based on the different relation types. When the type belongs to the relationship between spatial distance and interference, Output the relationship characteristics between targets, such as Euclidean distance and motion interference coefficient; when the... When the type belongs to spatial topological relations, Output topological features such as envelope, intersection, and separation between targets; when the... When the type belongs to a time relationship, Output the characteristics of the collaboration duration and state sequence between targets.

[0039] The method for calculating the topological interference features in step S3.3 is as follows: Perform spatial volume intersection integration on the three-dimensional dynamic safety envelope surface of multiple entities to accurately determine the inclusion, intersection, or separation topological relationships between entities; specifically: extract the global state matrix from step S2. 3D manifold coordinate domain of a specific entity With geometric dimension set As input, combined with the feedforward motion vector , To construct a three-dimensional dynamic safety envelope equation with time as the variable for operators and construction machinery. Abandoning traditional qualitative state determination, it directly substitutes it into the spatial relationship calculation operator: through point-to-polygon distance calculation and spatial vector operation, it outputs the precise Euclidean distance and motion interference coefficient between entities; by performing spatial volume intersection integration operation on the dynamic envelope surface of multiple entities (i.e., It accurately determines whether there are absolute topological interference features such as inclusion, intersection or separation between entities; and finally maps the continuous spatial distance features and discrete topological features calculated above to the spatial subarray of the global interaction relation matrix.

[0040] A framework for resolving process collaboration and temporal interaction relationships driven by specific state variables; the system addresses complex process intersections in highly dynamic construction environments by analyzing the global state matrix. Extracting semantic classification variables of actions from the target audience Lifecycle stage variables The corresponding timestamp sequence is used as input and imported into the time-dimensional collaboration and timing analysis operator. By calculating the time overlap integral function of different work subjects (such as a tower crane in the "lifting" state and a worker in the "hooking" state) within a specific collaborative action interval, the absolute collaboration duration of strongly coupled processes is output. At the same time, a process sequence logic judgment model is introduced to perform time-series cross-comparison of the life cycle evolution nodes and action sequences of each entity to clarify the leading and lagging topological logic between work subjects. Finally, the above-mentioned collaboration time span and process timing logic characteristics are accurately quantified and filled into the time subarray of the global interaction relationship matrix, providing objective data support for upper-level business scheduling and efficiency evaluation.

[0041] Example 5, by Figure 6 This embodiment, based on Embodiment 4, provides a detailed explanation of step S4. Step S4 specifically includes: S4.1 Constructing a global business matrix composed of multiple business feature vectors, each business feature vector corresponding to a specific engineering management business dimension; S4.2 Establishing a multi-source driven external benchmark dictionary, including a management benchmark dictionary integrating BIM design and project management parameters, and a regulatory benchmark dictionary encapsulating mandatory provisions of industry regulations; S4.3 Using state selection operators and business relationship selection operators, filtering out discrete features required for specific businesses from the global state matrix and global interaction relationship matrix; S4.4 Constructing a business feature judgment mapping function, logically aligning the filtered discrete features with the external benchmark dictionary to generate feature vectors for each business dimension; S4.5 Summarizing all business feature vectors to form a global business matrix, triggering corresponding active intervention commands when the value of a business feature vector exceeds a preset threshold or logical red line.

[0042] The global business matrix construction in step S4.1 specifically involves: parallel processing of construction site operations. A specific business logic (covering dimensions such as security, quality, schedule, and cost) defines the moment. Global business matrix This matrix is ​​represented as a column vector consisting of all business features across the entire scenario, i.e. .

[0043] Each element The representative system is for the first The feature vectors of specific business outputs aim to transform the underlying physical perception features and spatiotemporal interaction relationships into a top-level calculable and evaluable business decision-making language, providing a unified mathematical expression foundation for final management intervention.

[0044] In step S2, the external benchmark dictionary includes the management benchmark dictionary. With regulatory benchmark dictionary Management Benchmark Dictionary It integrates elements such as BIM design geometric datum, project management schedule time nodes, and cost budget control lines; regulatory datum dictionary. It encapsulates mandatory clause parameters from current laws and regulations such as the "Construction Project Management Standard" and the "Safety Technical Regulations for the Use of Construction Machinery," including absolute thresholds for safe distances and protection level standards. By vectorizing these external constraints, it provides legal, compliant, and scientific true values ​​for subsequent business characteristic determination.

[0045] In this embodiment, constructing the business feature determination mapping function specifically involves: constructing a business feature determination mapping function based on multi-operator fusion. And obtain business feature vectors The system utilizes state selection operators Selecting operators based on business relationships From the global state matrix With global relation matrix The discrete features required for specific business operations are extracted and combined with a benchmark dictionary. and Establish a business feature determination mapping function Thus, the time is obtained. Business feature vector As shown in the following formula: This function can automatically generate feature vectors covering various aspects such as personnel compliance and area safety (e.g., not wearing work clothes or entering restricted areas), danger distance early warning (e.g., intrusion monitoring of dangerous areas within hoisting radius), multi-stage building quality defect monitoring (e.g., defect quantification and component geometric deviation identification), progress and production efficiency monitoring (e.g., monitoring of process node lag), dynamic cost accounting and resource statistics (e.g., resource depletion monitoring), and environmental monitoring and facility integrity (e.g., monitoring of damaged protective nets and excessive dust). When the dynamic value output by the judgment function exceeds the preset threshold or logic red line, the system immediately generates a business instruction with management attributes to determine whether to trigger a hardware alarm or resource scheduling instruction at the lower level.

[0046] Output a global business matrix and execute multimodal management intervention measures to form a perception and decision-making closed loop; summarize the feature vectors of each business dimension to form the final global business matrix, and execute differentiated management intervention measures according to specific business scenarios, such as issuing early warning signals and implementing resource adjustment measures, so as to completely open up the closed loop link from physical perception to proactive intervention. Through this standardized decoupling and mapping mechanism based on the "4R" framework, intelligent and proactive closed-loop control of all elements and the entire life cycle of the construction site can be achieved.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A 4R system method for the research and development of AI applications for on-site management of construction projects, characterized by: Includes the following steps: S1, Recognize Everything: Acquire raw images, video data, and multi-source heterogeneous sensor data from the construction site, automatically segment and identify all physical elements on site, including personnel, machinery, materials, and environment, and abstract unstructured physical elements into a discrete generalized entity target matrix to form a standardized data perception base. S2. State Identification: Based on the generalized entity target matrix, high-dimensional feature quantization is performed on the spatiotemporal and semantic attributes of each entity. The time dimension is introduced to realize the evolution tracking of entity state. A diagonal global state matrix containing the heterogeneous state features of each entity at time t is constructed to complete the digital mapping of the physical world. S3. Relationship Identification: Based on the global state matrix, solve the spatial topological relationships and temporal coordination relationships between all entities in the scene, and construct a global interaction relationship matrix with row and column dimensions equivalent to the number of on-site targets, providing computable relationship features for the spatiotemporal interaction of targets. S4. Business Identification: Logically align the global state matrix and global interaction relationship matrix with a pre-built external benchmark dictionary containing a management benchmark dictionary and a regulatory benchmark dictionary to generate a global business matrix covering the dimensions of safety, quality, schedule, and cost. When a business characteristic exceeds a preset threshold, an executable proactive intervention instruction is output, forming a complete closed loop from physical perception to management decision-making. Step S4 specifically includes: S4.1 Construct a global business matrix composed of multiple business feature vectors, where each business feature vector corresponds to a specific engineering management business dimension; S4.2 Establish a multi-source driven external benchmark dictionary, including a management benchmark dictionary that integrates BIM design and project management parameters, and a regulatory benchmark dictionary that encapsulates mandatory provisions of industry regulations; S4.3 Utilize state selection operators and business relationship selection operators to filter out discrete features required for specific businesses from the global state matrix and global interaction relationship matrix; S4.4 Construct a business feature determination mapping function, logically align the filtered discrete features with the external benchmark dictionary, and generate feature vectors for each business dimension; S4.

5. Summarize all business feature vectors to form a global business matrix. When the value of a business feature vector exceeds a preset threshold or logical red line, trigger the corresponding active intervention command.

2. The 4R system method for the research and development of AI applications for on-site management of construction projects according to claim 1, characterized in that: Step S1 specifically includes: S1.1 Construct a generalized entity target matrix that includes four categories of elements: personnel, machinery and vehicles, materials and protective equipment, and environment and static facilities; S1.2 Perform brightness adjustment, contrast adjustment and geometric transformation preprocessing on the original image or video data, and remove highly repetitive or similar frames by calculating the image perception hash feature to form a cleaned construction scene image dataset. S1.3 For targets with regular shapes, the cleaned image is input into the Grounding DINO model for open vocabulary target detection. The DINOv2 model is used to extract foreground visual features and cluster them to generate initial localization anchor points. The bounding box information and the initial localization anchor points are used as joint conditional cues input into the SAM visual basic model to output a high-precision pixel-level instance segmentation mask. S1.

4. Confidence assessment and multi-scale enhanced consistency verification are performed on the instance segmentation results. High-confidence pseudo-labels with confidence scores higher than the preset threshold are selected. Combined with a very small number of manually labeled long-tailed complex samples, the Mask R-CNN architecture is used for semi-supervised iterative training to generate an instance segmentation model with engineering generalization capabilities. S1.5 For targets with irregular shapes, a fully convolutional semantic segmentation network is used to extract global contextual features of the image and perform pixel-level multi-classification prediction to supplement background semantic information. S1.6 Summarize all entity recognition results and background semantic information, and fill them into the generalized entity target matrix.

3. The 4R system method for the research and development of AI applications for on-site management of construction projects according to claim 1, characterized in that: Step S2 specifically includes: S2.1 Construct a diagonal global state matrix at time t, where each diagonal element corresponds to the state feature of an entity, and configure differentiated state attribute domains according to the entity type; S2.

2. Using the three-dimensional manifold space mapping method, the perspective projection transformation relationship between image pixel coordinates and real-world coordinates is established. The manifold constraint of the physical surface of the construction site is introduced to realize the absolute scale recovery of monocular vision and quantify the three-dimensional coordinate domain and geometric dimension parameters of the entity. S2.

3. Embed multi-target cross-frame tracking logic in the large model inference network, introduce multi-dimensional evidence joint modeling to realize cross-frame continuous binding of entity ID, and solve the motion vector features of the entity, the motion vector features including instantaneous motion velocity and acceleration vector; S2.4 For personnel and mechanical entities, microscopic action semantic features are extracted through spatiotemporal graph convolutional neural networks; for material and component entities, life cycle stages are determined based on target appearance evolution features by combining a multi-stage appearance evolution network with ordinal regression judgment operators. S2.

5. By integrating visual tensor data with infrared thermal imaging and ultrasonic multi-source heterogeneous sensing data, the defects quantification parameters of materials and components are extracted, and the closed-loop quantification of the state matrix of all elements is completed.

4. The 4R system method for the research and development of AI applications for on-site management of construction projects according to claim 1, characterized in that: Step S3 specifically includes: S3.1 Construct the global interaction relationship matrix at time t, where each element in the matrix represents the objective spatiotemporal interaction relationship between two entities; S3.2 Define the relation feature extraction operator and relation feature calculation rule, and extract the dimension-reduced feature submatrix from the global state matrix as the calculation parameter; S3.3 Based on the three-dimensional coordinate domain, geometric dimension parameters, and feedforward motion vectors of the entities, construct a three-dimensional dynamic safety envelope surface equation with time parameters as variables, and calculate the precise Euclidean distance and topological interference features between entities. The topological interference features include inclusion, intersection, or separation topological features. S3.4 Extract the action semantic classification variables, life cycle stage variables and corresponding timestamp sequences of entities from the global state matrix, and calculate the absolute coordination time of strongly coupled processes and the leading lag topological logic between the main working entities. S3.5 Fill the spatial relationship features and temporal relationship features into the spatial subarray and temporal subarray of the global interaction relationship matrix, respectively.

5. The 4R system method for the research and development of AI applications for on-site management of construction projects according to claim 3, characterized in that: The three-dimensional manifold space mapping method in step S2.2 is as follows: establish a perspective projection transformation equation including the camera intrinsic matrix, the camera extrinsic matrix, and the axial depth factor; for the ground target, establish the homography transformation of the support surface manifold equation and solve the absolute horizontal coordinates in reverse; for the high-altitude target, establish the elevation benchmark manifold equation of the BIM model design axis and solve the absolute scale factor in reverse; project the point set in the mask in reverse to the world coordinate system to construct the three-dimensional point cloud envelope, and quantize the absolute coordinate domain and geometric dimension parameters of the entity by solving the minimum circumscribed cuboid.

6. The 4R system method for the research and development of AI applications for on-site management of construction projects according to claim 3, characterized in that: The multidimensional evidence joint modeling method in step S2.3 is as follows: the cosine distance of the target appearance features and the motion smoothness based on optical flow field estimation are fused to construct a multidimensional association cost matrix; the Hungarian algorithm is used to perform bipartite graph optimal matching to achieve cross-frame continuous binding of entity IDs; Based on the coordinate changes of entities in consecutive frames, the instantaneous velocity and acceleration vector of the entities are calculated in reverse.

7. The 4R system method for the research and development of AI applications for on-site management of construction projects according to claim 4, characterized in that: The method for calculating the topological interference features in step S3.3 is as follows: perform spatial volume intersection integration on the three-dimensional dynamic safety envelope surface of multiple entities to accurately determine the inclusion, intersection or separation topological relationships between entities; and map the continuous spatial distance features and discrete topological interference features together to the spatial subarray of the global interaction relationship matrix.

8. The 4R system method for the research and development of AI applications for on-site management of construction projects according to claim 1, characterized in that: The business feature determination mapping function can generate six types of business feature vectors: personnel compliance and regional safety, danger distance early warning, building entity quality defect monitoring, progress and production efficiency supervision, dynamic cost accounting and resource statistics, and environmental supervision and facility integrity monitoring.

Images