Construction project full-process supervision system based on digital and intelligent integration

Abstract

The present invention relates to the field of deep learning technology and proposes a full-process supervision system for construction projects based on digital intelligence fusion, aiming to achieve closed-loop supervision of construction site data perception, state prediction, and intelligent decision-making. The system includes a digital intelligence perception module, a feature extraction module, a digital intelligence-driven prediction module, a graph reasoning decision module, and a construction supervision module, integrating a GCN-LSTM-NF-JD model and a regularized graph reasoning model for multimodal feature prediction and knowledge reasoning, respectively. By collecting and fusing heterogeneous data such as temperature, stress, displacement, image, and engineering information, component state prediction, construction delay scoring, and risk level identification are achieved, and high-risk node warnings and optimal construction path recommendations are generated based on graph neural networks. The system can output construction scheduling tables, on-site intervention instructions, and risk decision-making reports to improve project safety, progress control capabilities, and response efficiency.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a full-process supervision system for construction projects based on digital intelligence integration. Background Art

[0002] With the development of building informatization and smart construction site technology, more and more construction project supervision systems are trying to introduce artificial intelligence algorithms to perceive, analyze and predict construction site conditions in order to improve the level of intelligent project management. However, existing systems still have many limitations in algorithm and model design. First, the existing system time series modeling is structurally rigid. Many supervision systems only use shallow neural networks or traditional sliding window methods to model time series, which cannot capture the complex time series dynamics and nonlinear evolution relationships in the construction process. Even when using sequence models such as LSTM, they often fail to incorporate spatial dependency structures, resulting in prediction lags, large time deviations, and insensitivity to the identification of abnormal evolution trends. Second, existing systems perform well in specific construction sites or sub-scenarios, but because they do not systematically consider data distribution drift and model structure adaptability, their prediction accuracy drops sharply when switching between different projects, construction stages or regions, and they lack sustainable deployment capabilities. Summary of the Invention

[0003] The present invention proposes a full-process supervision system for construction projects based on digital and intelligent fusion, aiming to improve the processing capabilities of construction sites in multi-source data fusion, intelligent state prediction and reasoning decision-making; the system constructs a high-dimensional, multi-modal engineering perception system, and realizes intelligent, structured and visual supervision of the construction process; the system integrates two core models: one is the GCN-LSTM-NF-JD model, which is used to process multi-modal engineering feature data, complete component state prediction, construction time deviation scoring and abnormality level identification; the other is the regularized graph reasoning model, which is used to process cleaned heterogeneous data, perform multi-hop reasoning through graph neural networks and regularization mechanisms, and generate decision results such as construction path recommendations and high-risk node warnings; the above two models run in coordination to realize full-process automated management from data perception, feature extraction to state prediction and intelligent decision-making, providing a highly intelligent, interpretable and real-time responsive supervision system architecture for construction projects, significantly improving the safety, efficiency and risk management capabilities of the construction process.

[0004] The present invention provides a construction project full-process supervision system based on digital intelligence fusion, which includes a digital intelligence perception module, a feature extraction module, a digital intelligence driven prediction module, a graph reasoning decision module and a construction supervision module;

[0005] The digital intelligence perception module collects heterogeneous data from construction sites and performs data denoising, format conversion, and event tagging to generate the original multi-source data set. It uses local weighted regression and isolation forest to remove outliers from the original multi-source data set, then performs missing value filling and data alignment to build a unified spatiotemporal index structure and generate cleaned heterogeneous data.

[0006] The feature extraction module performs principal component analysis and wavelet packet decomposition on numerical sensor data to generate numerical feature data; performs YOLOv8 target detection and DeepSORT target tracking on video image data to generate image feature data; extracts engineering information features of structured engineering information to generate engineering structure feature data; and fuses numerical feature data, image feature data, and engineering structure feature data to generate multimodal engineering feature data.

[0007] The digital intelligence-driven prediction module builds a GCN-LSTM-NF-JD model. This model processes multimodal engineering feature data and generates multidimensional prediction results, including component status prediction, construction time deviation score, and anomaly level labeling. The GCN-LSTM-NF-JD model includes a GCN network, an LSTM network, an autoregressive normalized flow model, and a fully connected layer.

[0008] The graph reasoning decision module builds a regularized graph reasoning model, processes cleaned heterogeneous data through the regularized graph reasoning model, generates decision reasoning results, and combines the decision reasoning results to obtain the optimal construction path recommendation and high-risk node warning list. The regularized graph reasoning model includes a state embedding unit, a graph reasoning unit, and a decision generation unit.

[0009] The construction supervision module combines multi-dimensional prediction results and decision-making reasoning results to intelligently optimize construction plans, resource scheduling and site management strategies, and generate construction optimization schedules, site intervention instructions and visual risk decision reports.

[0010] Furthermore, the process of processing multimodal engineering feature data through the GCN-LSTM-NF-JD model to generate multidimensional prediction results specifically includes the following steps:

[0011] Step S1: The index unit of the multimodal engineering feature data is represented as a component. The multimodal engineering feature data is input into the GCN network to extract the spatial dependency features between the components and form graph-structured spatial embedding data. The graph-structured spatial embedding data is input into the LSTM network in time series to model the historical dependency of the component states and output LSTM temporal encoding features.

[0012] A component is an identifiable engineering object unit with spatial, temporal, and process semantics. It is both a graph node and an analysis object for a time series.

[0013] Step S2: Process the LSTM time series encoding features through the autoregressive normalized flow model to generate the encoded latent variable sequence;

[0014] Step S3: Construct an explicit nonlinear equation system model and use the Jacobi parallel iteration method to perform parallel decoding on the encoded latent variable sequence. To improve computational efficiency and take into account temporal dependencies, a selective parallel processing mechanism is introduced. Traditional serial inference is used in the high-dependency layer of the autoregressive normalized flow model to preserve the causal information order between latent variables. Jacobi parallel decoding is performed in the redundant layer of the autoregressive normalized flow model to accelerate the decoding process, optimize the overall decoding performance and improve timeliness, achieve the optimal compromise between efficiency and accuracy, and obtain the Jacobi decoding tensor.

[0015] Step S4: Generate a multi-dimensional prediction result by processing the Jacobi decoding tensor using a fully connected layer.

[0016] Furthermore, in step S2, the process of the autoregressive normalized flow model processing the LSTM time series coding features specifically includes: performing a reversible chain transformation on the LSTM time series coding features, gradually mapping them to a Gaussian latent variable space with probabilistic interpretation, and generating a latent variable sequence; using a neural network parameterized offset function and scaling function to perform autoregressive control on each dimension of the latent variable in the latent variable sequence, retaining the characteristic timing and structural dependency of the LSTM time series coding features, and generating an encoded latent variable sequence.

[0017] Furthermore, the process of processing cleaned heterogeneous data through the regularized graph reasoning model specifically includes the following:

[0018] The state embedding unit extracts entities and relationships from the cleaned heterogeneous data, constructs a set of knowledge triples, and encodes the knowledge triples into a third-order sparse tensor as the engineering knowledge graph. It also performs feature extraction on the cleaned heterogeneous data to generate a low-dimensional state vector, which is then aligned and mapped with the entities in the engineering knowledge graph to generate a state-aware embedding matrix.

[0019] The graph reasoning unit combines the state-aware embedding matrix, initializes the graph neural network, and uses the engineering knowledge graph as the graph structure to construct a multi-hop message propagation path between nodes and perform multi-hop path propagation. It introduces an intermediate variable regularization mechanism and applies a Frobenius norm penalty term to the intermediate messages in each hop path propagation. It also performs matrix expansion operations on the tensors in all propagation paths, constructs a tensor expansion matrix, and calculates the upper bound constraint of the overlapping nuclear norm based on the tensor expansion matrix, thereby achieving structural compression and correlation enhancement of the reasoning process. It generates regularized node representations and constructs the knowledge reasoning result tensor.

[0020] The decision generation unit generates a decision reasoning result based on the knowledge reasoning result tensor.

[0021] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0022] By introducing the GCN-LSTM-NF-JD joint model, the present invention achieves efficient modeling and intelligent prediction of multimodal engineering features at the construction site, improving the ability to accurately identify component state evolution, construction time deviation, and abnormal behavior. The model fully integrates the spatial structural dependence and temporal dynamic changes between components, which not only enhances the model's perception of construction progress and safety status under complex working conditions, but also improves the reversibility and reasoning efficiency of feature representation through normalized flow and Jacobi decoding mechanisms, solving the problems of prediction response lag and weak expression of temporal information in existing systems, and providing high-precision and interpretable state assessment results for construction sites.

[0023] At the same time, the present invention realizes structured knowledge reasoning of cleaned heterogeneous engineering data by constructing a regularized graph reasoning model, enhancing the system's ability to automatically identify and warn of construction paths and high-risk nodes. Based on the engineering knowledge graph, the model effectively simulates the complex interactive relationships between components through state embedding and multi-hop path propagation mechanisms. The introduced regularization mechanism significantly improves the reasoning stability and result credibility, solving problems such as ineffective path diffusion and large structural noise interference in traditional systems, and providing solid support for path recommendation and risk intervention in complex engineering scenarios.

[0024] Through the collaborative integration of the GCN-LSTM-NF-JD model and the regularized graph reasoning model, the present invention constructs a full-process, intelligent construction project supervision system, realizing a closed-loop management system from data perception, feature fusion, predictive analysis to reasoning and decision-making; the system not only greatly improves the information processing efficiency and predictive response speed of project supervision, but also enhances the structural perception and strategy optimization capabilities of the construction management process, providing a technical foundation for the realization of smart construction sites, unmanned supervision and adaptive control of project risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a module diagram of a construction project full-process supervision system based on digital intelligence integration proposed in the present invention;

[0026] Figure 2 This is the construction scheduling Gantt chart proposed in Example 7.

[0027] Figure 2 In the task list, green represents "completed" tasks, orange represents "in progress" tasks, blue represents "in planning" stages, and red represents "warning delays". DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] Example 1, according to Figure 1 The present invention provides a full-process supervision system for construction projects based on digital-intelligence fusion, which includes a digital-intelligence perception module, a feature extraction module, a digital-intelligence driven prediction module, a graph reasoning decision module, and a construction supervision module.

[0030] The digital intelligence perception module collects heterogeneous data from construction sites and performs data denoising, format conversion, and event tagging to generate an original multi-source dataset. It uses local weighted regression and isolation forest to remove outliers from the original multi-source dataset, then performs missing value completion and data alignment to build a unified spatiotemporal index structure and generate cleaned heterogeneous data. The cleaned heterogeneous data includes numerical sensor data, video image data, and structured engineering information.

[0031] Numerical sensor data includes temperature data, displacement data, stress data, and vibration data;

[0032] Video image data includes construction site monitoring data and drone shooting data;

[0033] Structured engineering information includes BIM model data, construction schedule, and process node data;

[0034] The feature extraction module performs principal component analysis and wavelet packet decomposition on numerical sensor data to generate numerical feature data; performs YOLOv8 target detection and DeepSORT target tracking on video image data to generate image feature data; extracts engineering information features of structured engineering information to generate engineering structure feature data; and fuses numerical feature data, image feature data, and engineering structure feature data to generate multimodal engineering feature data.

[0035] The digital intelligence-driven prediction module builds a GCN-LSTM-NF-JD model. This model processes multimodal engineering feature data and generates multidimensional prediction results, including component status prediction, construction time deviation score, and anomaly level labeling. The GCN-LSTM-NF-JD model includes a GCN network, an LSTM network, an autoregressive normalized flow model, and a fully connected layer.

[0036] PS: (NF: Normalizing Flow, JD: Jacobi Decoding);

[0037] The graph reasoning decision module builds a regularized graph reasoning model, processes cleaned heterogeneous data through the regularized graph reasoning model, generates decision reasoning results, and combines the decision reasoning results to obtain the optimal construction path recommendation and high-risk node warning list. The regularized graph reasoning model includes a state embedding unit, a graph reasoning unit, and a decision generation unit.

[0038] The construction supervision module combines multi-dimensional prediction results and decision-making reasoning results to intelligently optimize construction plans, resource scheduling and site management strategies, and generate construction optimization schedules, site intervention instructions and visual risk decision reports.

[0039] Example 2: This example is based on Example 1. In this example, the process of processing multimodal engineering feature data through the GCN-LSTM-NF-JD model to generate multidimensional prediction results specifically includes the following steps:

[0040] Step S1: The index unit of the multimodal engineering feature data is represented as a component. The multimodal engineering feature data is input into the GCN network to extract the spatial dependency features between the components and form graph-structured spatial embedding data. The graph-structured spatial embedding data is input into the LSTM network in time series to model the historical dependency of the component states and output LSTM temporal encoding features.

[0041] A component is an identifiable engineering object unit with spatial, temporal, and process semantics. It is both a graph node and an analysis object for a time series.

[0042] Step S2: Process the LSTM time series encoding features through the autoregressive normalized flow model to generate the encoded latent variable sequence;

[0043] Step S3: Construct an explicit nonlinear equation system model and use the Jacobi parallel iteration method to perform parallel decoding on the encoded latent variable sequence. To improve computational efficiency and take into account temporal dependencies, a selective parallel processing mechanism is introduced. Traditional serial inference is used in the high-dependency layer of the autoregressive normalized flow model to preserve the causal information order between latent variables. Jacobi parallel decoding is performed in the redundant layer of the autoregressive normalized flow model to accelerate the decoding process, optimize the overall decoding performance and improve timeliness, achieve the optimal compromise between efficiency and accuracy, and obtain the Jacobi decoding tensor. The formula used is as follows:

[0044] Parallel decoding update formula:

[0045] ;

[0046] in, represents the layer index in the autoregressive normalized flow model, represents the Jacobi iteration index, represents the latent variable dimension index, Indicates the In the layer, The latent variable at the position Update value in iterations; Indicates the In the layer, The latent variable value at each position, represents the Hadamard product, represents the exponential function, Indicates the The scaling function of the layer, Represents the distance from the 0th dimension to the All subsets of the elements of the dimension, Indicates the In the iteration, Layer latent variables Before The sub-vector of dimensions, i.e., the pre-order dependent variable; Indicates the The offset function of the layer;

[0047] Constructing an explicit nonlinear equation model means modeling the update process of the encoded latent variable sequence as a solvable nonlinear system consisting of multiple dimensions and multiple levels, where the value of each variable explicitly depends on the previous variable and the previous level variable. The system is solved in parallel and iteratively using the Jacobi method to achieve efficient decoding.

[0048] Step S4: Generate a multi-dimensional prediction result by processing the Jacobi decoding tensor using a fully connected layer.

[0049] Example 3: This example is based on Example 1. In this example, the process of processing multimodal engineering feature data to generate multidimensional prediction results specifically includes the following steps:

[0050] Step E1: The index units of the multimodal engineering feature data are represented as components. The multimodal engineering feature data is input into the GCN network to extract the spatial dependency features between components and form graph-structured spatial embedding data. The graph-structured spatial embedding data is input into the LSTM network in time series to model the historical dependency of component states and output LSTM temporal encoding features.

[0051] Step E2: Process the LSTM time series encoding features through the variational autoencoder to generate a latent variable sequence with a standard normal distribution;

[0052] Step E3: Input the latent variable sequence of the standard normal distribution into the multi-layer perceptron decoder, perform forward propagation, and generate the decoded engineering component state tensor;

[0053] Step E4: Combine the decoded engineering component state tensor to generate a multi-dimensional prediction result.

[0054] Example 4. This example is based on Example 2. In this example, in step S2, the process of the autoregressive normalized flow model processing the LSTM time series coding features specifically includes: performing a reversible chain transformation on the LSTM time series coding features, gradually mapping them to a Gaussian latent variable space with probabilistic interpretation, and generating a latent variable sequence; using a neural network parameterized offset function and scaling function to perform autoregressive control on each dimension of the latent variable in the latent variable sequence, retaining the characteristic timing and structural dependency of the LSTM time series coding features, and generating an encoded latent variable sequence.

[0055] Example 5: This example is based on Example 4. In this example, the process of processing cleaned heterogeneous data using a regularized graph reasoning model specifically includes the following:

[0056] The state embedding unit extracts entities and relationships from the cleaned heterogeneous data, constructs a set of knowledge triples, and encodes the knowledge triples into a third-order sparse tensor as the engineering knowledge graph. It also performs feature extraction on the cleaned heterogeneous data to generate a low-dimensional state vector, which is then aligned and mapped with the entities in the engineering knowledge graph to generate a state-aware embedding matrix.

[0057] The graph reasoning unit combines the state-aware embedding matrix, initializes the graph neural network, uses the engineering knowledge graph as the graph structure, constructs a multi-hop message propagation path between nodes, and performs multi-hop path propagation. It introduces an intermediate variable regularization mechanism, introduces an intermediate variable regularization mechanism in each hop path propagation, applies a Frobenius norm penalty term to the intermediate message, and performs a matrix expansion operation on the tensors in all propagation paths to construct a tensor expansion matrix. Based on the tensor expansion matrix, the upper bound constraint of the overlapping nuclear norm is calculated to achieve structural compression and correlation enhancement of the reasoning process. It generates a regularized node representation and constructs the knowledge reasoning result tensor. The formula used is as follows:

[0058] Message passing and node update formula in multi-hop path propagation:

[0059] ;

[0060] in, Indicates the index of the current propagation layer of the graph neural network, represents the target node, express Neighbor nodes of Representation node In the The new embedding representation of the layer, Represents neighbor nodes In the The state embedding vector of the layer, represents a nonlinear activation function, Representation node The adjacency set of Representation node To Node The weight of the information transmitted, represents the learnable weight matrix;

[0061] Calculate the upper bound constraint formula for the overlapping nuclear norm:

[0062] ;

[0063] in, represents the power exponent parameter, represents the third-order tensor, that is, the tensor expansion form in all propagation paths, represents the first modular expansion of a third-order tensor, represents the second modular expansion of the third-order tensor, represents the third-mode expansion of a third-order tensor; represents the nuclear norm, represents the exponentiation of the nuclear norm of the expanded matrix;

[0064] The decision generation unit generates a decision reasoning result based on the knowledge reasoning result tensor.

[0065] Example 6: This example is based on Example 4. In this example, the process of processing cleaned heterogeneous data specifically includes the following:

[0066] The state embedding unit extracts entities and relationships from the cleaned heterogeneous data, constructs a set of knowledge triples, and encodes the knowledge triples into a third-order sparse tensor as the engineering knowledge graph. It also performs feature extraction on the cleaned heterogeneous data to generate a low-dimensional state vector, which is then aligned and mapped with the entities in the engineering knowledge graph to generate a state-aware embedding matrix.

[0067] The graph reasoning unit combines the state-aware embedding matrix to initialize the graph neural network, uses the engineering knowledge graph as the graph structure, and performs graph embedding calculations. During the graph embedding process, L2 norm regularization constraints are introduced to node features, and the Dropout mechanism is used to prevent overfitting, generate stable node representations, and construct the knowledge reasoning result tensor.

[0068] The decision generation unit generates a decision reasoning result based on the knowledge reasoning result tensor.

[0069] Embodiment 7, according to Figure 2 This embodiment is based on the sixth embodiment. In this embodiment, the construction supervision module combines the multi-dimensional prediction results and decision reasoning results to intelligently optimize the construction plan, resource scheduling and site management strategy, and generates a construction optimization schedule, site intervention instructions and a visual risk decision report;

[0070] In this embodiment, for Project A:

[0071] The multi-dimensional prediction results are shown in Table 1

[0072] Table 1

[0073] ;

[0074] Decision reasoning results:

[0075] High-risk node warning list: Component: C302-B, Status: "Increased risk of structural instability", Related nodes: Construction schedule delay + increased vibration intensity;

[0076] Optimal path recommendation result: The order of steel bar binding → formwork construction → concrete pouring nodes is optimized to A → C → B, avoiding delayed nodes;

[0077] The construction optimization scheduling is shown in Table 2

[0078] Table 2

[0079] ;

[0080] The visual risk decision report generates a construction scheduling Gantt chart, which includes foundation treatment, basement structure construction, above-ground structure construction, C302-B component pouring, main structure acceptance and interior construction preparation.

[0081] The present invention and its embodiments are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A full-process construction project supervision system based on digital-intelligence integration includes a digital-intelligence perception module and a feature extraction module. The digital-intelligence perception module constructs cleaned heterogeneous data, and the feature extraction module extracts features from the cleaned heterogeneous data to generate multimodal engineering feature data. This system is characterized by: The system also includes a digital intelligence driven prediction module and a graph reasoning decision module; The digital intelligence-driven prediction module establishes a GCN-LSTM network. It introduces normalized flow and Jacobi parallel decoding mechanisms to optimize the feature representation and decoding efficiency of the GCN-LSTM network. This results in a GCN-LSTM-NF-JD model, which includes a GCN network, an LSTM network, an autoregressive normalized flow model, and a fully connected layer. The graph reasoning decision module establishes a graph neural network, introduces a state-aware embedding mechanism and an intermediate variable regularization mechanism, optimizes the semantic expression ability of the graph neural network and controls the structural sparsity of the reasoning process, and constructs a regularized graph reasoning model. This regularized graph reasoning model processes cleaned heterogeneous data and generates decision reasoning results. The digital intelligence-driven prediction module processes multimodal engineering feature data through the GCN-LSTM-NF-JD model to generate multidimensional prediction results. The specific steps are as follows: Step S1: Input the multimodal engineering feature data into the GCN network and LSTM network to obtain LSTM temporal encoding features; Step S2: Process the LSTM time series encoding features through the autoregressive normalized flow model to generate the encoded latent variable sequence; Step S3: Construct an explicit nonlinear equation system model and use the Jacobi parallel iteration method to perform parallel decoding on the encoded latent variable sequence; introduce a selective parallel processing mechanism to optimize the overall decoding performance and obtain the Jacobi decoding tensor; Step S4: Use the fully connected layer to process the Jacobi decoding tensor to generate a multi-dimensional prediction result; In step S2, the process of processing LSTM time series coding features with the autoregressive normalized flow model specifically includes: performing a reversible chain transformation on the LSTM time series coding features and mapping them to the Gaussian latent variable space to generate a latent variable sequence; using the offset function and scaling function parameterized by the neural network to perform autoregressive control on each dimension of the latent variable in the latent variable sequence, retaining the characteristic timing and structural dependency of the LSTM time series coding features, and generating an encoded latent variable sequence.

2. The construction project full-process supervision system based on digital intelligence integration according to claim 1 is characterized by: The regularized graph reasoning model includes a state embedding unit, a graph reasoning unit, and a decision generation unit.

3. The construction project full-process supervision system based on digital intelligence integration according to claim 1 is characterized by: The specific selective parallel processing mechanism is as follows: for the highly dependent layers with strong dependencies in the autoregressive normalized flow model, the traditional serial inference method is used to preserve the information order; for the redundant layers with weak dependencies, the Jacobi parallel decoding strategy is used to accelerate the decoding process.

4. The construction project full-process supervision system based on digital intelligence integration according to claim 2 is characterized by: The state embedding unit extracts entities and relationships from cleaned heterogeneous data, constructs an engineering knowledge graph, and generates a state-aware embedding matrix.

5. The construction project full-process supervision system based on digital intelligence integration according to claim 4 is characterized by: The graph reasoning unit combines the state-aware embedding matrix, initializes the graph neural network, and uses the engineering knowledge graph as the graph structure to perform multi-hop path propagation. It introduces an intermediate variable regularization mechanism, applies a Frobenius norm penalty term in each hop path propagation, and performs matrix expansion operations on the tensors in all propagation paths to construct a tensor expansion matrix. Based on the tensor expansion matrix, the upper bound constraint of the overlapping nuclear norm is calculated to obtain the knowledge reasoning result tensor.

6. The construction project full-process supervision system based on digital intelligence integration according to claim 5 is characterized by: The decision generation unit generates a decision reasoning result based on the knowledge reasoning result tensor.

Images