A building engineering progress risk prediction and evaluation method based on big data analysis

Abstract

The application discloses a kind of based on big data analysis's construction engineering progress risk prediction evaluation method, comprising: S1, synchronous acquisition construction engineering field multi-source heterogeneous big data, space-time alignment and construct five-order progress tensor space;S2, complete denoising and pass through Riemann manifold learning projection to low-dimensional feature space, calculate geodesic trajectory and dynamic risk potential field;S3, construct heterogeneous progress correlation diagram and map manifold characteristics to vertex attribute;S4, decouple noise and management factors using counterfactual learning mechanism, mine causal characteristics;S5, use the improved neural stochastic differential equation of introducing heterogeneous graph neural network drift term guide mechanism to construct differential equation, solve risk evolution trajectory;S6, matching knowledge graph generates comprehensive risk evaluation report.The application realizes the accurate prediction of risk evolution trend, improves the accuracy of complex engineering progress risk control.

Description

Technical Field

[0001] This invention relates to the field of construction project management and intelligent risk prediction, and in particular to a method for predicting and evaluating the progress risk of construction projects based on big data analysis. Background Technology

[0002] In recent years, big data analytics in construction engineering, combining deep learning and complex system modeling techniques, has been widely applied in fields such as construction monitoring, resource scheduling, and safety management, becoming a core technological means to achieve intelligent engineering management. However, in practical applications, construction project schedule risk prediction scenarios face numerous challenges, including complex and ever-changing construction environments, strong heterogeneity of multi-source data, and complex risk transmission mechanisms. The effectiveness of existing prediction technologies remains constrained by various factors.

[0003] Currently, most risk prediction methods rely on single-modal data sources for modeling, making it difficult to fully utilize the complementary characteristics of multi-source information such as on-site sensor data, design model information, text reports, and video materials. This results in a lack of comprehensiveness in the perception of risk status. Some systems only use static time series analysis or simple statistical regression models, ignoring potential environmental noise interference and the extraction of key causal factors. This makes the models highly susceptible to being misled by external non-critical factors (such as short-term weather fluctuations) and capturing false correlations, limiting the accuracy and robustness of the models in identifying real risks. At the same time, the prediction process often lacks in-depth analysis of risk propagation mechanisms, making it difficult to clearly demonstrate the transmission chain of risks between different construction elements to managers, affecting the credibility of early warning results and their value in decision-making guidance.

[0004] Furthermore, existing continuous-time evolution prediction models typically assume that the deterministic trend parameters are fixed functions or rely solely on self-fitting from historical states when setting risk change trends. This fails to dynamically integrate the complex process logic dependencies and resource supply and demand constraints in engineering projects into the adjustment process of trend parameters. Consequently, the risk evolution direction output by the model contradicts the actual physical logic of construction, making it difficult to accurately depict the cascading propagation law of risks in complex networks and seriously affecting the practical value of the model in real engineering scenarios.

[0005] Therefore, how to provide a method for predicting and evaluating the progress risk of construction projects based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for predicting and evaluating the progress risk of construction projects based on big data analysis. This invention fully integrates key steps such as fifth-order progress tensor construction, Riemannian manifold learning, causal decoupling using heterogeneous graph transformers, improved neural stochastic differential equation prediction, and knowledge graph matching. It constructs an intelligent risk prediction process with data spatiotemporal alignment, manifold geometric feature extraction, graph topology-guided drift term constraints, and causal intervention mechanisms. This invention captures the nonlinear evolution characteristics of progress states through Riemannian manifold learning. In particular, it introduces a drift term guidance mechanism based on heterogeneous graph neural networks into the improved neural stochastic differential equations, dynamically correcting drift term parameters using graph structures to accurately characterize the deterministic trend of risk cascading transmission. This invention possesses advantages such as deeper feature extraction, evolutionary trends conforming to actual physical logic, prediction dimensions including probability and path, and interpretable causal mechanisms. It can significantly improve the accuracy and timeliness of risk prediction in complex construction environments, effectively solving problems such as one-sided risk state representation, evolutionary trend modeling detached from topological constraints, and insufficient decision-making basis in existing methods.

[0007] A method for predicting and evaluating the progress risk of construction projects based on big data analysis, according to an embodiment of the present invention, includes the following steps:

[0008] S1. Simultaneously collect multi-source heterogeneous big data from the construction site, perform spatiotemporal alignment and tensor processing, and construct a fifth-order progress tensor space.

[0009] S2. Complete and denoise the missing values ​​and noise in the fifth-order progress tensor space. Project the processed high-dimensional tensor data to the low-dimensional manifold feature space through Riemannian manifold learning. Calculate the geodesic trajectory of the current progress state in the low-dimensional manifold feature space and construct a dynamic risk potential energy field to generate a manifold geometric feature set.

[0010] S3. Abstract the process nodes, resource nodes, and environmental nodes in the construction project into vertices of a heterogeneous graph, and abstract the logical dependencies between processes, the supply and demand of resources, and the environmental constraints into edges of the graph. Construct a heterogeneous progress association graph and map the manifold geometric feature set to the vertex attributes of the heterogeneous progress association graph.

[0011] S4. Input the heterogeneous schedule correlation graph into a pre-built heterogeneous graph transformer with a causal intervention mechanism, introduce a counterfactual learning mechanism as a causal attention unit, and decouple environmental noise and internal management factors by comparing the difference between the actual observed state and the risk-free counterfactual state, deeply explore the core causal features that lead to schedule risks, and output node feature codes.

[0012] S5. Input the node feature encoding into the improved neural stochastic differential equation, introduce the drift term guidance mechanism of the heterogeneous graph neural network, utilize the topological constraints of the heterogeneous progress association graph to constrain the risk evolution trend, construct the differential dynamic equation, solve the evolution trajectory of the risk within the predetermined future time window, and output the dynamic evolution prediction result.

[0013] S6. Match the dynamic evolution prediction results with historical risk cases in the preset knowledge graph library to identify the underlying causes of the current risk pattern and similar historical response strategies, and generate a comprehensive risk assessment report.

[0014] Optionally, S1 specifically includes:

[0015] S11. By synchronously collecting raw data through IoT sensors deployed at the construction site, BIM modeling software interface and monitoring cameras, the construction site's IoT sensor time series data, BIM model progress data, construction log text and monitoring video stream are obtained.

[0016] S12. Perform spatiotemporal alignment processing, using the geographic information in the BIM model progress data as the spatial reference and the timestamp of the IoT sensor time series data as the time reference, map each frame of the monitoring video stream and the construction log text to a unified time coordinate system and spatial coordinate system.

[0017] S13. Perform tensor filling operation on the spatiotemporally aligned IoT sensor time series data, BIM model progress data, construction log text and monitoring video stream, and map the data of different modalities to the corresponding index positions of the fifth-order tensor to generate initial tensor data.

[0018] S14. Define the five dimensions of the initial tensor data as time, space, resources, process and environment, and construct a fifth-order schedule tensor space.

[0019] Optionally, S2 specifically includes:

[0020] S21. First, expand the fifth-order schedule tensor space into five two-dimensional matrices along the five dimensions of time, space, resources, process and environment. Calculate the product of each expanded two-dimensional matrix with its own transpose matrix to obtain the covariance matrix of each matrix. Solve for the eigenvalues ​​and eigenvectors of the covariance matrix. Arrange the eigenvectors in columns to obtain the factor matrix. Take the square root of the eigenvalues ​​to obtain the singular values.

[0021] S22. Perform tensor multiplication on the vectors in the five-dimensional factor matrices and multiply them with the singular value diagonal matrix to obtain the tensor kernel matrix. Sort all singular values ​​according to their numerical values, retain the first preset number of main singular values, and reconstruct them by performing tensor multiplication on the factor matrix to generate denoised and completed tensor data.

[0022] S23. Preset the parameters of the radial basis kernel function. Using any two data points in the tensor data, calculate the difference between the coordinate values ​​of these two data points. Square the difference and sum them to obtain the squared distance. Use the exponential function to calculate the value with half the negative number of the squared distance as the exponent as the element of the kernel function matrix. Calculate the eigenvalues ​​and eigenvectors of the kernel function matrix. Take the square root of the eigenvalues ​​to obtain the reciprocal and multiply it by the product of the eigenvector matrix and the kernel function matrix. Use the column vectors in the result matrix as the feature coordinates of the projected manifold.

[0023] S24. Read the manifold feature coordinates of the current time node and the manifold feature coordinates of the previous time node in the low-dimensional manifold feature space and calculate the coordinate difference. Use Euclidean distance as the tangent space distance between the two points, calculate the logarithm of the ratio of the tangent space distance to the hyperbolic tangent function, multiply the tangent space direction vector by the logarithm, and calculate the geodesic vector from the previous time node to the current time node. Connect the geodesic vectors of these two nodes to form the geodesic trajectory of the current progress state in the low-dimensional manifold feature space.

[0024] S25. Preset the value of the standard progress reference coordinate point, read the coordinate of each manifold feature in the low-dimensional manifold feature space, calculate the difference between the manifold feature and the standard progress reference coordinate point in each dimension, and sum the squared values ​​to obtain the squared distance value. Multiply the squared distance value by negative half and take the negative number to obtain the potential energy value corresponding to the manifold feature coordinate point. Combine all potential energy values ​​to construct a dynamic risk potential energy field.

[0025] S26. Extract three adjacent trajectory points on the geodesic trajectory. Subtract the coordinates of the previous point from the coordinates of the middle point to obtain a first-order difference vector. Subtract the coordinates of the middle point from the coordinates of the next point to obtain another first-order difference vector. Calculate the vector product of the two first-order difference vectors and take the magnitude. Calculate the cube of the magnitude of the first-order difference vector. Divide the magnitude of the vector product by the cube of the magnitude of the first-order difference vector to obtain the curvature parameter.

[0026] S27. Read the potential energy value of the current coordinate point in the dynamic risk potential energy field and take the absolute value. Multiply the curvature parameter by the preset first weight coefficient to obtain the first weighting term. Multiply the absolute value of the potential energy value by the preset second weight coefficient to obtain the second weighting term and add it to the first weighting term to obtain the progress deviation value. Combine the progress deviation values ​​of all points to generate a manifold geometric feature set.

[0027] Optionally, S3 specifically includes:

[0028] S31. Read the BIM model progress data containing geographic information, extract the task name and number in the construction plan as process nodes, read the IoT sensor time series data, extract the unique identifier of mechanical equipment and materials as resource nodes, extract the unique identifier of environmental monitoring equipment as environmental nodes, and summarize the process nodes, resource nodes and environmental nodes to construct a set of vertices of a heterogeneous graph.

[0029] S32. Extract the preceding and following logical relationships of process nodes in the BIM model progress data. Extract the planned start and end times of two process nodes and calculate the difference. If it is greater than or equal to zero, it is determined that there is a dependency relationship. Connect the two process nodes with preceding and following dependencies with directed line segments to construct process logical dependency edges.

[0030] S33. Extract records of material delivery from resource nodes to process nodes from IoT sensor time-series data, calculate the material delivery quantity, and when the delivery quantity is greater than zero, establish a supply and demand connection line between the corresponding resource node and process node to construct a resource supply and demand relationship edge.

[0031] S34. Extract the overlapping information of environmental nodes and process nodes in time and space, compare the monitoring values ​​of environmental nodes with the preset monitoring thresholds of process nodes, establish constraint connection lines when the preset monitoring thresholds are exceeded, construct environmental constraint relationship edges, summarize the process logic dependency edges, resource supply and demand relationship edges and environmental constraint relationship edges, and construct a heterogeneous progress association graph.

[0032] S35. Perform the mapping of manifold geometric feature set to the vertex attributes of the heterogeneous progress graph. Read the manifold geometric feature set and assign the feature values ​​in the manifold geometric feature set as vertex attributes to the vertices in the heterogeneous progress graph according to the timestamp and spatial position of the vertices.

[0033] Optionally, S4 specifically includes:

[0034] S41. Pre-construct a heterogeneous graph transformer containing a sequentially connected heterogeneous graph attention layer, causal intervention layer, and feature output layer. Input the heterogeneous progress association graph into the heterogeneous graph attention layer, read the connection edges between vertices in the heterogeneous progress association graph, and for each connection edge, calculate the dot product of the feature vector of the source node and the feature vector of the target node and divide it by the square root of the dimension to obtain the attention score. Then input the score into the exponential function to calculate the normalization coefficient. Multiply the normalization coefficient by the feature vector of the source node to obtain the weighted node features.

[0035] S42. Construct a causal attention unit, preset a risk-free counterfactual state, replace the monitoring values ​​of environmental nodes in the heterogeneous progress association graph with zero, and set the supply and demand values ​​of resource nodes to the theoretical maximum value to generate a risk-free counterfactual heterogeneous graph.

[0036] S43. Input the weighted node features into the causal intervention layer, introduce the counterfactual learning mechanism, read the actual observed state features of the heterogeneous progress association graph, calculate the vector difference between the actual observed state features and the risk-free counterfactual heterogeneous graph features, and use it as the causal effect estimate. In the causal intervention layer, set the causal effect estimate of the environmental nodes as a noise interference term and remove it from the actual observed state features to obtain the pure node features after removing environmental noise.

[0037] S44. Calculate the variance of the pure node features, divide the pure node features by the variance for standardization, and then perform corresponding operations with the preset linear transformation matrix and bias vector to deeply mine the core causal features that lead to schedule risks.

[0038] S45. Input the core causal features into the feature output layer and multiply them with the preset weight matrix to obtain the contribution value. Normalize the contribution value to obtain the weight value within the interval. Multiply the weight value with the core causal features to generate the node feature code and output it.

[0039] Optionally, S5 specifically includes:

[0040] S51. Construct an improved neural stochastic differential equation, define the differential dynamics equation to include drift term, diffusion term and Brownian motion term, preset the number of layers and neurons of the drift term neural network, initialize the weight matrix and bias vector in the drift term neural network to random values, and input the node feature encoding into the improved neural stochastic differential equation in the order of time steps.

[0041] S52. Introduce a drift term guidance mechanism based on heterogeneous graph neural network, read the adjacency matrix of each node in the heterogeneous progress association graph, extract the neighbor node number directly connected to each node, retrieve the corresponding feature vector in the node feature encoding containing causal weight according to the neighbor node number, sum and average the retrieved neighbor node feature vectors to generate the initial graph embedding vector.

[0042] S53. Input the initial graph embedding vector into a two-layer pre-defined fully connected neural network, perform matrix multiplication, calculate the product of the initial graph embedding vector and the first layer pre-defined weight matrix, and superimpose the first layer bias vector. After processing by the Tanh activation function, input it into the second layer, calculate the product with the second layer pre-defined weight matrix, and superimpose the second layer bias vector to output the high-dimensional graph embedding vector.

[0043] S54. Input the high-dimensional graph embedding vector into the feedforward layer of the drift term neural network, read the original weight matrix to be updated of the drift term neural network, multiply the high-dimensional graph embedding vector with the weight matrix to be updated element by element, use the matrix value after multiplication to replace the parameter value in the weight matrix to be updated, calculate the product of the replaced weight matrix and the node feature encoding containing causal weights input in S51 and add the bias term, use the topological structure of the heterogeneous progress association graph to constrain the deterministic trend of risk evolution, and generate the guiding drift term value.

[0044] S55. Read the standard normal distribution random number generator, generate a random noise array with the same dimension as the node feature encoding of the current time step input, multiply the random noise array with the preset noise coefficient of the diffusion term, calculate the diffusion fluctuation value, perform addition operation on the value of the guiding drift term and the diffusion fluctuation value, and obtain the derivative dynamic value of the risk of the current time step changing with continuous time.

[0045] S56. Set the end time point of the predetermined future time window, construct a differential dynamic equation describing the cascading transmission of risk in the heterogeneous progress relationship diagram, use the state value of the current time step and the dynamic value of the derivative to calculate the predicted state value of the next time step, repeat the calculation steps until the accumulated time step reaches the end time point of the predetermined future time window, and output all intermediate calculation results as the future risk state sequence.

[0046] S57. Connect the values ​​in the future risk state sequence with a broken line according to the time sequence, draw the evolution trajectory of risk over time, compare each value in the evolution trajectory with the preset risk threshold one by one, and mark the time point when the value first exceeds the preset risk threshold as the key risk time point.

[0047] S58. Based on the key risk time points, locate the process node where the risk occurs in the heterogeneous schedule association diagram, read the directed edge connection relationship in the heterogeneous schedule association diagram, start from the located process node, backtrack along the opposite direction of the directed edge to search for the parent node, arrange all the parent nodes on the path in the connection order, and generate the risk transmission path.

[0048] S59. The proportion of data points in the future risk state sequence whose values ​​exceed the preset risk threshold to the total number of data points is used as the probability of risk occurrence. The planned working hours of all process nodes on the risk transmission path are read and summed to obtain the total time. The total time is subtracted from the standard working hours to calculate the delay time. The probability of risk occurrence, the risk transmission path and the delay time are combined to output the dynamic evolution prediction result.

[0049] Optionally, S56 specifically includes:

[0050] S561. Read the system configuration parameters, extract the total duration of the preset future time window, divide the total duration by the preset single-step time step, calculate the total number of iterations to be executed, and store the value of the total number of iterations as the loop termination condition.

[0051] S562. Construct a differential dynamic equation describing the cascading transmission of risk in a heterogeneous progress correlation diagram. Set the node feature encoding containing causal weights as the initial state variable of the equation, set the dynamic value of the derivative as the rate of change parameter of the state variable, read the state value of the node feature encoding at the current time step, multiply it with the dynamic value of the derivative, multiply the product result with the single-step time step size, and calculate the deterministic drift increment.

[0052] S563. Call the random number generation function to generate random numbers that follow a standard normal distribution, and multiply them by the preset diffusion term coefficient and the square root of the single-step time step to obtain the random diffusion increment. Perform an addition operation between the deterministic drift increment and the random diffusion increment, and add the calculation result to the state value of the node feature encoding at the current time step to obtain the predicted state value for the next time step.

[0053] S564. Update the predicted state value of the next time step to the state value of the node feature encoding of the new current time step, accumulate the single-step time step length to the total running time, and record the predicted state value to the result storage list. Determine whether the total running time has reached the predetermined future time window termination time point. If not, return to step S562 to continue the calculation; otherwise, terminate the loop calculation process. Read all the predicted state values ​​recorded in the result storage list, combine them in chronological order, and output the future risk state sequence.

[0054] Optionally, S6 specifically includes: establishing a knowledge graph database of construction engineering risks, performing subgraph matching and semantic retrieval between the dynamic evolution prediction results and historical risk cases in the knowledge graph database, identifying the underlying causes of the current risk and historically similar response strategies, and combining a comprehensive risk assessment report that includes the probability of risk occurrence, risk transmission path and delay duration, underlying causes and historically similar response strategies.

[0055] The beneficial effects of this invention are:

[0056] This invention addresses the challenges of multi-source heterogeneity, high noise levels, and difficulty in feature extraction in construction engineering data by constructing a fifth-order schedule tensor space and deploying Riemannian manifold learning. It employs tensor decomposition for data completion and denoising, and combines potential field theory and geodesic trajectory extraction to extract manifold geometric feature sets, generating feature vectors that reflect the essence of schedule deviations. These feature vectors are mapped to a heterogeneous schedule correlation graph, and counterfactual learning is used to decouple environmental noise and internal management factors, outputting node feature codes containing causal weights. These node feature codes are then input into an improved neural stochastic differential equation, specifically introducing a drift term guidance mechanism based on a heterogeneous graph neural network. The drift term parameters are dynamically corrected using the graph topology, and the derivative dynamics are calculated using stochastic diffusion to solve for the evolution trajectory of risk within a predetermined window, outputting prediction results including probability of occurrence, transmission path, and delay duration. Furthermore, knowledge graph matching is used to identify deep-seated causes and corresponding strategies. Ultimately, this achieves geometric representation, causal decoupling, and dynamic evolution prediction of construction engineering schedule risks, effectively improving the comprehensiveness of risk status perception, the physical logic of evolution trend prediction, and the interpretability of results. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of a construction project schedule risk prediction and evaluation method based on big data analysis proposed in this invention;

[0059] Figure 2 This is a flowchart of the progress state manifold geometric feature extraction and dynamic risk potential field construction based on Riemannian manifold learning proposed in this invention.

[0060] Figure 3 This is a flowchart of the improved neural stochastic differential equation risk evolution trajectory prediction based on the drift term guidance mechanism of heterogeneous graph neural network proposed in this invention. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0062] refer to Figures 1-3 A method for predicting and evaluating the progress risk of construction projects based on big data analysis includes the following steps:

[0063] S1. Synchronously collect multi-source heterogeneous big data from the construction site. The multi-source heterogeneous big data includes time-series data from IoT sensors at the construction site, BIM model progress data containing geographic information, unstructured construction log text, and monitoring video streams. Then, perform spatiotemporal alignment and tensor processing to construct a fifth-order progress tensor space containing five-dimensional information of time, space, resources, process, and environment.

[0064] S2. Use tensor decomposition algorithm to complete and denoise missing values ​​and noise in the fifth-order progress tensor space. Project the processed high-dimensional tensor data to the low-dimensional manifold feature space through Riemannian manifold learning. Calculate the geodesic trajectory of the current progress status in the low-dimensional manifold feature space and construct a dynamic risk potential energy field based on potential energy field theory to generate a manifold geometric feature set that reflects the degree of progress deviation.

[0065] S3. Abstract the process nodes, resource nodes, and environmental nodes in the construction project into vertices of a heterogeneous graph, and abstract the logical dependencies between processes, the supply and demand of resources, and the environmental constraints into edges of the graph. Construct a dynamically evolving heterogeneous progress association graph, and map the manifold geometric feature set to the vertex attributes of the heterogeneous progress association graph.

[0066] S4. Input the heterogeneous schedule correlation graph into a pre-built heterogeneous graph transformer with a causal intervention mechanism, introduce a counterfactual learning mechanism as a causal attention unit, and decouple environmental noise and internal management factors by comparing the difference between the actual observed state and the risk-free counterfactual state, deeply explore the core causal features that lead to schedule risks, and output node feature codes containing causal weights.

[0067] S5. Input the node feature encoding into the improved neural stochastic differential equation, introduce the drift term guidance mechanism of the heterogeneous graph neural network, use the topological constraints of the heterogeneous progress association graph to constrain the risk evolution trend, construct the differential dynamic equation, solve the evolution trajectory of the risk within the predetermined future time window, and output the dynamic evolution prediction results including the probability of risk occurrence, transmission path and delay duration.

[0068] S6. Match the dynamic evolution prediction results with historical risk cases in the preset knowledge graph library to identify the underlying causes of the current risk pattern and similar historical response strategies, and generate a comprehensive risk assessment report.

[0069] This invention significantly improves the accuracy and timeliness of construction project schedule risk prediction. By constructing a fifth-order schedule tensor space, it achieves unified modeling of multi-source heterogeneous big data, effectively solving the problems of missing data and noise interference at the construction site. Utilizing Riemannian manifold learning to capture geodesic trajectories and dynamic risk potential fields, it can keenly perceive subtle deviations in schedule status. Combining heterogeneous graph neural networks and counterfactual learning mechanisms, it can accurately decouple environmental noise and internal management factors, uncover the core causal features leading to risks, and significantly improve the interpretability of feature encoding. An improved neural stochastic differential equation introducing a heterogeneous graph neural network drift term guidance mechanism can dynamically constrain risk evolution trends using graph topology, thereby accurately predicting risk trajectories within future windows. Finally, by matching knowledge graphs to generate comprehensive evaluation reports, it provides managers with scientific decision-making basis. This solution exhibits stronger robustness in complex construction environments, achieving dynamic quantification and early warning of schedule risks.

[0070] In this embodiment, S1 specifically includes:

[0071] S11. By synchronously collecting raw data through IoT sensors deployed at the construction site, BIM modeling software interface and monitoring cameras, the construction site can obtain IoT sensor time series data, BIM model progress data containing geographic information, unstructured construction log text and monitoring video stream.

[0072] S12. Perform spatiotemporal alignment processing, using the geographic information in the BIM model progress data as the spatial reference and the timestamp of the IoT sensor time series data as the time reference, to map each frame of the monitoring video stream and the unstructured construction log text to a unified time coordinate system and spatial coordinate system.

[0073] S13. Perform tensor filling operations on the spatiotemporally aligned IoT sensor time series data, BIM model progress data containing geographic information, unstructured construction log text, and monitoring video stream, mapping the data of different modalities to the corresponding index positions of the fifth-order tensor to generate initial tensor data.

[0074] S14. Define the five dimensions of the initial tensor data as time, space, resources, process and environment, and construct a fifth-order progress tensor space containing the five dimensions of time, space, resources, process and environment.

[0075] In this embodiment, S2 specifically includes:

[0076] S21. First, expand the fifth-order schedule tensor space into five two-dimensional matrices along the five dimensions of time, space, resources, process and environment. Calculate the product of each expanded two-dimensional matrix with its own transpose matrix to obtain the covariance matrix of each matrix. Solve for the eigenvalues ​​and eigenvectors of the covariance matrix. Arrange the eigenvectors in columns to obtain the factor matrix. Take the square root of the eigenvalues ​​to obtain the singular values.

[0077] S22. Perform tensor multiplication on the vectors in the five-dimensional factor matrices and multiply them with the singular value diagonal matrix to obtain the tensor kernel matrix. Sort all singular values ​​according to their numerical values, retain the first preset number of main singular values, and reconstruct them by performing tensor multiplication on the factor matrix to generate denoised and completed tensor data.

[0078] S23. Preset the parameters of the radial basis kernel function. Using any two data points in the tensor data, calculate the difference between the coordinate values ​​of these two data points. Square the difference and sum them to obtain the squared distance. Use the exponential function to calculate the value with half the negative number of the squared distance as the exponent as the element of the kernel function matrix. Calculate the eigenvalues ​​and eigenvectors of the kernel function matrix. Take the square root of the eigenvalues ​​to obtain the reciprocal and multiply it by the product of the eigenvector matrix and the kernel function matrix. Use the column vectors in the result matrix as the feature coordinates of the projected manifold.

[0079] S24. Read the manifold feature coordinates of the current time node and the manifold feature coordinates of the previous time node in the low-dimensional manifold feature space and calculate the coordinate difference. Use Euclidean distance as the tangent space distance between the two points, calculate the logarithm of the ratio of the tangent space distance to the hyperbolic tangent function, multiply the tangent space direction vector by the logarithm, and calculate the geodesic vector from the previous time node to the current time node. Connect the geodesic vectors of these two nodes to form the geodesic trajectory of the current progress state in the low-dimensional manifold feature space.

[0080] S25. Preset the value of the standard progress reference coordinate point, read the coordinate of each manifold feature in the low-dimensional manifold feature space, calculate the difference between the manifold feature and the standard progress reference coordinate point in each dimension, and sum the squared values ​​to obtain the squared distance value. Multiply the squared distance value by negative half and take the negative number to obtain the potential energy value corresponding to the manifold feature coordinate point. Combine all potential energy values ​​to construct a dynamic risk potential energy field.

[0081] S26. Extract three adjacent trajectory points on the geodesic trajectory. Subtract the coordinates of the previous point from the coordinates of the middle point to obtain a first-order difference vector. Subtract the coordinates of the middle point from the coordinates of the next point to obtain another first-order difference vector. Calculate the vector product of the two first-order difference vectors and take the magnitude. Calculate the cube of the magnitude of the first-order difference vector. Divide the magnitude of the vector product by the cube of the magnitude of the first-order difference vector to obtain the curvature parameter.

[0082] S27. Read the potential energy value of the current coordinate point in the dynamic risk potential energy field and take the absolute value. Multiply the curvature parameter by the preset first weight coefficient to obtain the first weighting term. Multiply the absolute value of the potential energy value by the preset second weight coefficient to obtain the second weighting term and add it to the first weighting term to obtain the progress deviation value. Combine the progress deviation values ​​of all points to generate a manifold geometric feature set.

[0083] This implementation employs an innovative technique combining high-order tensor decomposition and Riemannian manifold learning, offering significant advantages over traditional principal component analysis and linear regression methods in processing multi-source heterogeneous data from construction engineering projects. Traditional methods struggle to effectively handle the spatiotemporal correlations of high-dimensional data and have limitations in nonlinear feature extraction, easily leading to the loss of risk features. This invention, however, expands and reconstructs along five dimensions using a tensor decomposition algorithm, accurately capturing the complex intrinsic dependencies between data. While achieving efficient denoising and completion, it preserves the original data's structure and details to the greatest extent possible. Riemannian manifold learning projects high-dimensional data into a low-dimensional space, quantifying progress deviation through geodesic trajectories and dynamic risk potential fields, effectively overcoming the problem of measurement bias in Euclidean space and enabling the sensitive identification of subtle nonlinear changes in progress status. Combined with comprehensive analysis of curvature parameters, the anti-interference capability and accuracy of feature representation are significantly improved, providing high-manifold geometric support for subsequent risk prediction and enhancing the model's robustness and interpretability in complex construction environments.

[0084] In this embodiment, S3 specifically includes:

[0085] S31. Read the BIM model progress data containing geographic information, extract the task name and number in the construction plan as process nodes, read the IoT sensor time series data, extract the unique identifier of mechanical equipment and materials as resource nodes, extract the unique identifier of environmental monitoring equipment as environmental nodes, and summarize the process nodes, resource nodes and environmental nodes to construct a set of vertices of a heterogeneous graph.

[0086] S32. Extract the preceding and following logical relationships of process nodes in the BIM model progress data. Extract the planned start and end times of two process nodes and calculate the difference. If it is greater than or equal to zero, it is determined that there is a dependency relationship. Connect the two process nodes with preceding and following dependencies with directed line segments to construct process logical dependency edges.

[0087] S33. Extract records of material delivery from resource nodes to process nodes from IoT sensor time-series data, calculate the material delivery quantity, and when the delivery quantity is greater than zero, establish a supply and demand connection line between the corresponding resource node and process node to construct a resource supply and demand relationship edge.

[0088] S34. Extract the overlapping information of environmental nodes and process nodes in time and space, compare the monitoring values ​​of environmental nodes with the preset monitoring thresholds of process nodes, establish constraint connection lines when the preset monitoring thresholds are exceeded, construct environmental constraint relationship edges, summarize the process logic dependency edges, resource supply and demand relationship edges and environmental constraint relationship edges, and construct a dynamically evolving heterogeneous progress relationship graph.

[0089] S35. Perform the mapping of manifold geometric feature set to the vertex attributes of the heterogeneous progress graph. Read the manifold geometric feature set and assign the feature values ​​in the manifold geometric feature set as vertex attributes to the vertices in the heterogeneous progress graph according to the timestamp and spatial position of the vertices.

[0090] In this embodiment, S4 specifically includes:

[0091] S41. A heterogeneous graph transformer containing a sequentially connected heterogeneous graph attention layer, causal intervention layer, and feature output layer is pre-constructed. The dynamically evolving heterogeneous progress association graph is input into the heterogeneous graph attention layer. The connection edges between vertices in the heterogeneous progress association graph are read. For each connection edge, the dot product of the feature vector of the source node and the feature vector of the target node is calculated and divided by the square root of the dimension to obtain the attention score. Then, the normalization coefficient is calculated by inputting the exponential function. The normalization coefficient is multiplied by the feature vector of the source node to obtain the weighted node features.

[0092] S42. Construct a causal attention unit, preset a risk-free counterfactual state, replace the monitoring values ​​of environmental nodes in the heterogeneous progress association graph with zero, and set the supply and demand values ​​of resource nodes to the theoretical maximum value to generate a risk-free counterfactual heterogeneous graph.

[0093] S43. Input the weighted node features into the causal intervention layer, introduce the counterfactual learning mechanism, read the actual observed state features of the heterogeneous progress association graph, calculate the vector difference between the actual observed state features and the risk-free counterfactual heterogeneous graph features, and use it as the causal effect estimate. In the causal intervention layer, set the causal effect estimate of the environmental nodes as a noise interference term and remove it from the actual observed state features to obtain the pure node features after removing environmental noise.

[0094] S44. Calculate the variance of the pure node features, divide the pure node features by the variance for standardization, and then perform corresponding operations with the preset linear transformation matrix and bias vector to deeply mine the core causal features that lead to schedule risks.

[0095] S45. Input the core causal features into the feature output layer and multiply them with the preset weight matrix to obtain the contribution value. Normalize the contribution value to obtain the weight value within the interval. Multiply the weight value with the core causal features to generate node feature codes containing causal weights and output them.

[0096] In this embodiment, S5 specifically includes:

[0097] S51. Construct an improved neural stochastic differential equation, define the differential dynamics equation to include drift term, diffusion term and Brownian motion term, preset the number of layers and neurons of the drift term neural network, initialize the weight matrix and bias vector in the drift term neural network to random values, and input the node feature encoding containing causal weights into the improved neural stochastic differential equation in the order of time steps.

[0098] S52. Introduce a drift term guidance mechanism based on heterogeneous graph neural network, read the adjacency matrix of each node in the heterogeneous progress association graph, extract the neighbor node number directly connected to each node, retrieve the corresponding feature vector in the node feature encoding containing causal weight according to the neighbor node number, sum and average the retrieved neighbor node feature vectors to generate an initial graph embedding vector containing graph structure constraints.

[0099] S53. Input the initial graph embedding vector into a two-layer pre-defined fully connected neural network, perform matrix multiplication, calculate the product of the initial graph embedding vector and the first layer pre-defined weight matrix, and superimpose the first layer bias vector. After processing by the Tanh activation function, input it into the second layer, calculate the product with the second layer pre-defined weight matrix, and superimpose the second layer bias vector to output the high-dimensional graph embedding vector.

[0100] S54. Input the high-dimensional graph embedding vector into the feedforward layer of the drift term neural network, read the original weight matrix to be updated of the drift term neural network, multiply the high-dimensional graph embedding vector with the weight matrix to be updated element by element, use the matrix value after multiplication to replace the parameter value in the weight matrix to be updated, calculate the product of the replaced weight matrix and the node feature encoding containing causal weights input in S51 and add the bias term, use the topological structure of the heterogeneous progress association graph to constrain the deterministic trend of risk evolution, and generate deterministic guiding drift term values;

[0101] S55. Read the standard normal distribution random number generator, generate a random noise array with the same dimension as the node feature encoding of the current time step input, multiply the random noise array with the preset noise coefficient of the diffusion term, calculate the diffusion fluctuation value, perform addition operation on the value of the guiding drift term and the diffusion fluctuation value, and obtain the derivative dynamic value of the risk of the current time step changing with continuous time.

[0102] S56. Set the end time point of the predetermined future time window, construct a differential dynamic equation describing the cascading transmission of risk in the heterogeneous progress relationship diagram, use the state value of the current time step and the dynamic value of the derivative to calculate the predicted state value of the next time step, repeat the calculation steps until the accumulated time step reaches the end time point of the predetermined future time window, and output all intermediate calculation results as the future risk state sequence.

[0103] S57. Connect the values ​​in the future risk state sequence with a broken line according to the time sequence, draw the evolution trajectory of risk over time, compare each value in the evolution trajectory with the preset risk threshold one by one, and mark the time point when the value first exceeds the preset risk threshold as the key risk time point.

[0104] S58. Based on the key risk time points, locate the process node where the risk occurs in the heterogeneous schedule association diagram, read the directed edge connection relationship in the heterogeneous schedule association diagram, start from the located process node, backtrack along the opposite direction of the directed edge to search for the parent node, arrange all the parent nodes on the path in the connection order, and generate the risk transmission path.

[0105] S59. The proportion of data points in the future risk state sequence whose values ​​exceed the preset risk threshold to the total number of data points is used as the probability of risk occurrence. The planned working hours of all process nodes on the risk transmission path are read and summed to obtain the total time. The total time is subtracted from the standard working hours to calculate the delay time. The probability of risk occurrence, the risk transmission path and the delay time are combined to output the dynamic evolution prediction result.

[0106] This invention introduces an improved neural stochastic differential equation combined with a heterogeneous graph-guided mechanism to achieve dynamic evolution trajectory prediction and cascade transmission analysis of construction project schedule risks. Node features containing causal weights are encoded and input into the differential equation. A high-dimensional graph embedding vector is generated by aggregating neighbor features using an adjacency matrix, and the drift term weight parameters are corrected. The deterministic trend of risk evolution is constrained by the graph topology, while Brownian motion terms describe random fluctuations, solving for the risk state sequence within future time windows. This invention can accurately capture the dynamic propagation patterns and key nodes of risks in the process network. The calculated risk evolution trajectory, combined with threshold comparison, effectively locates risk time points and transmission paths, significantly improving the prediction accuracy of delay duration and probability of occurrence.

[0107] The improved neural stochastic differential equation of this invention shares common ground with other neural stochastic differential equations in terms of basic architecture. Both follow the differential dynamics framework of stochastic processes and define a differential equation structure that includes drift terms, diffusion terms, and Brownian motion terms. Both utilize neural networks to parameterize the drift term function, calculate the derivative of risk over continuous time using numerical solution methods, and employ random noise multiplied by preset coefficients to describe the uncertainty fluctuations of the system.

[0108] The difference lies in that this invention breaks the closed nature of relying solely on historical node features for drift term calculation, and introduces a drift term guidance mechanism based on heterogeneous graph neural networks. Neural stochastic differential equations directly interact with the currently input feature vector and the weight matrix to be updated, ignoring topological constraints between nodes; while this invention, in steps S52 and S53, aggregates neighbor node features by reading the adjacency matrix and generates a high-dimensional graph embedding vector through a fully connected layer. Then, in step S54, this vector is multiplied element-wise with the weight matrix of the drift term neural network, directly utilizing the graph structure to correct the parameter space of the drift term.

[0109] The beneficial effect of the improvement lies in the fact that, by introducing a heterogeneous graph embedding to guide the drift term, the neural stochastic differential equation can utilize the topological structure of the heterogeneous schedule association graph to dynamically constrain the deterministic trend of risk evolution, effectively solving the problem that traditional methods cannot capture the logical dependencies and resource coupling between processes. Embedding the global prior knowledge of the graph into the dynamic evolution of the differential equation means that risk prediction is no longer limited to the historical trajectory of a single node, but rather propagates along the cascading paths of the association graph. This significantly improves the modeling accuracy of the systemic risk propagation laws in complex engineering projects and enhances the rationality and reliability of the prediction results in dynamic evolution scenarios.

[0110] In this embodiment, S56 specifically includes:

[0111] S561. Read the system configuration parameters, extract the total duration of the preset future time window, divide the total duration by the preset single-step time step, calculate the total number of iterations to be executed, and store the value of the total number of iterations as the loop termination condition.

[0112] S562. Construct a differential dynamic equation describing the cascading transmission of risk in a heterogeneous progress correlation diagram. Set the node feature encoding containing causal weights as the initial state variable of the equation, set the dynamic value of the derivative as the rate of change parameter of the state variable, read the state value of the node feature encoding at the current time step, multiply it with the dynamic value of the derivative, multiply the product result with the single-step time step size, and calculate the deterministic drift increment.

[0113] S563. Call the random number generation function to generate random numbers that follow a standard normal distribution, and multiply them by the preset diffusion term coefficient and the square root of the single-step time step to obtain the random diffusion increment. Perform an addition operation between the deterministic drift increment and the random diffusion increment, and add the calculation result to the state value of the node feature encoding at the current time step to obtain the predicted state value for the next time step.

[0114] S564. Update the predicted state value of the next time step to the state value of the node feature encoding of the new current time step, accumulate the single-step time step length to the total running time, and record the predicted state value to the result storage list. Determine whether the total running time has reached the predetermined future time window termination time point. If not, return to step S562 to continue the calculation; otherwise, terminate the loop calculation process. Read all the predicted state values ​​recorded in the result storage list, combine them in chronological order, and output the future risk state sequence.

[0115] In this embodiment, S6 specifically includes: establishing a knowledge graph library of construction engineering risks, performing subgraph matching and semantic retrieval on the dynamic evolution prediction results and historical risk cases in the knowledge graph library, identifying the underlying causes of the current risk and historically similar response strategies, and combining the probability of risk occurrence, risk transmission path and delay duration, underlying causes and historically similar response strategies into a comprehensive risk assessment report.

[0116] Example 1: To verify the feasibility of this invention in practice, it was applied to a schedule risk prediction and control platform for a key highway reconstruction and expansion project in a certain province. The project is 146 kilometers long, involving the expansion of a two-way eight-lane highway and the reconstruction of numerous bridges, tunnels, and interchanges, with a total investment of 23 billion yuan. The project spans three prefecture-level cities, includes eight specialized sections, more than 120 construction teams, and over 8,300 frontline personnel. The complex geological conditions along the route and the need for continuous traffic flow necessitate frequent traffic diversions. This results in traditional methods relying on manual weekly reports and static Gantt charts suffering from information lag and weak predictive capabilities, making it difficult to address the cascading effects of schedule delays.

[0117] In practical applications, the method of this invention first accesses the five-dimensional data source of the project management system, including daily plans and actual completions in the time dimension, construction section distribution in the spatial dimension, material supply and equipment configuration in the resource dimension, critical paths and operational logic in the process dimension, and meteorological conditions and traffic control information in the environmental dimension. The system constructs the above heterogeneous data into a fifth-order schedule tensor space, and achieves adaptive completion and denoising of missing data and noise through tensor decomposition and reconstruction. On this basis, this invention uses the Riemannian manifold learning method to project the high-dimensional tensor data onto a low-dimensional manifold feature space, captures the evolution path of the schedule state through geodesic trajectories, constructs a dynamic risk potential energy field, quantifies the deviation between the current schedule state and the baseline plan, and calculates curvature parameters to identify schedule inflection points.

[0118] Subsequently, this invention maps the manifold geometric feature set to a heterogeneous schedule correlation graph. The nodes in the graph encompass process nodes, resource nodes, segment nodes, and external environment nodes, while edges represent relationships such as process dependencies, resource constraints, and spatiotemporal correlations. Through a counterfactual learning mechanism, the system separates schedule fluctuations caused by environmental noise (such as severe weather or unforeseen events) and schedule deviations caused by internal management factors (such as improper resource allocation or process coordination failures) in the latent space. Based on this, this invention constructs an improved neural stochastic differential equation. Its core innovation lies in introducing a drift term guidance mechanism based on a heterogeneous graph neural network. This mechanism dynamically corrects the drift term parameters using the graph topology, embedding the updated node features of the heterogeneous graph as guidance information into the drift term, while retaining the stochastic diffusion term to describe uncertainty fluctuations. The model solves the differential equation within a predetermined prediction window, outputting the evolution trajectory of risk over time, including key prediction results such as the probability of risk occurrence, transmission path, and expected delay duration. The table below compares the performance of this invention with traditional methods in the task of predicting schedule risks at key engineering nodes:

[0119] Table 1. Performance Comparison Data Between the Invention and Traditional Methods

[0120]

[0121] Based on the comparison data in the table above, it can be seen that the risk prediction method based on Riemannian manifold learning and improved neural stochastic differential equations proposed in this invention has shown significant performance advantages over traditional methods in the key node management of highway reconstruction and expansion projects. In particular, it has achieved a qualitative leap in key indicators such as early warning accuracy, prediction lead time, and control of false alarms and missed alarms.

[0122] In terms of early warning accuracy, this invention maintains a high level of over 94% across all seven key stages, while the average accuracy of traditional methods is only around 70%. For example, in the "main bridge cantilever construction stage," the accuracy of traditional methods is only 68.9% due to the difficulty in handling complex process connections, while this invention, through multi-source data fusion and manifold geometric feature extraction, improves the accuracy to 95.8%, significantly enhancing the ability to identify high-risk processes.

[0123] Regarding predictive timeliness, this invention utilizes manifold trajectories to capture early trends, extending the average lead time for predictions from less than 5 days to approximately 12 days. For example, in the "asphalt pavement paving" scenario, this invention significantly extends the warning time from 6.1 days to 13.5 days, providing ample window space for resource allocation. The control of false alarm and missed alarm rates is also remarkably effective. This invention controls the false alarm rate to around 3.1% and the missed alarm rate to within 1.8%, compared to the traditional method's approximately 11% false alarm and nearly 10% missed alarm. This greatly reduces ineffective interventions and risk omissions, significantly improving the precision of project schedule management and the reliability of decision-making.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting and evaluating the progress risk of construction projects based on big data analysis, characterized in that, Includes the following steps: S1. Simultaneously collect multi-source heterogeneous big data from the construction site, perform spatiotemporal alignment and tensor processing, and construct a fifth-order progress tensor space. S2. Complete and denoise the missing values ​​and noise in the fifth-order progress tensor space. Project the processed high-dimensional tensor data to the low-dimensional manifold feature space through Riemannian manifold learning. Calculate the geodesic trajectory of the current progress state in the low-dimensional manifold feature space and construct a dynamic risk potential energy field to generate a manifold geometric feature set. S3. Abstract the process nodes, resource nodes, and environmental nodes in the construction project into vertices of a heterogeneous graph, and abstract the logical dependencies between processes, the supply and demand of resources, and the environmental constraints into edges of the graph. Construct a heterogeneous progress association graph and map the manifold geometric feature set to the vertex attributes of the heterogeneous progress association graph. S4. Input the heterogeneous schedule correlation graph into a pre-built heterogeneous graph transformer with a causal intervention mechanism, introduce a counterfactual learning mechanism as a causal attention unit, and decouple environmental noise and internal management factors by comparing the difference between the actual observed state and the risk-free counterfactual state, deeply explore the core causal features that lead to schedule risks, and output node feature codes. S5. Input the node feature encoding into the improved neural stochastic differential equation, introduce the drift term guidance mechanism of the heterogeneous graph neural network, utilize the topological constraints of the heterogeneous progress association graph to constrain the risk evolution trend, construct the differential dynamic equation, solve the evolution trajectory of the risk within the predetermined future time window, and output the dynamic evolution prediction result. S6. Match the dynamic evolution prediction results with historical risk cases in the preset knowledge graph library to identify the underlying causes of the current risk pattern and similar historical response strategies, and generate a comprehensive risk assessment report.

2. The method for predicting and evaluating the progress risk of construction projects based on big data analysis according to claim 1, characterized in that, S1 specifically includes: S11. By synchronously collecting raw data through IoT sensors deployed at the construction site, BIM modeling software interface and monitoring cameras, the construction site's IoT sensor time series data, BIM model progress data, construction log text and monitoring video stream are obtained. S12. Perform spatiotemporal alignment processing, using the geographic information in the BIM model progress data as the spatial reference and the timestamp of the IoT sensor time series data as the time reference, map each frame of the monitoring video stream and the construction log text to a unified time coordinate system and spatial coordinate system. S13. Perform tensor filling operation on the spatiotemporally aligned IoT sensor time series data, BIM model progress data, construction log text and monitoring video stream, and map the data of different modalities to the corresponding index positions of the fifth-order tensor to generate initial tensor data. S14. Define the five dimensions of the initial tensor data as time, space, resources, process and environment, and construct a fifth-order schedule tensor space.

3. The method for predicting and evaluating the progress risk of construction projects based on big data analysis according to claim 1, characterized in that, S2 specifically includes: S21. First, expand the fifth-order schedule tensor space into five two-dimensional matrices along the five dimensions of time, space, resources, process and environment. Calculate the product of each expanded two-dimensional matrix with its own transpose matrix to obtain the covariance matrix of each matrix. Solve for the eigenvalues ​​and eigenvectors of the covariance matrix. Arrange the eigenvectors in columns to obtain the factor matrix. Take the square root of the eigenvalues ​​to obtain the singular values. S22. Perform tensor multiplication on the vectors in the five-dimensional factor matrices and multiply them with the singular value diagonal matrix to obtain the tensor kernel matrix. Sort all singular values ​​according to their numerical values, retain the first preset number of main singular values, and reconstruct them by performing tensor multiplication on the factor matrix to generate denoised and completed tensor data. S23. Preset the parameters of the radial basis kernel function. Using any two data points in the tensor data, calculate the difference between the coordinate values ​​of these two data points. Square the difference and sum them to obtain the squared distance. Use the exponential function to calculate the value with half the negative number of the squared distance as the exponent as the element of the kernel function matrix. Calculate the eigenvalues ​​and eigenvectors of the kernel function matrix. Take the square root of the eigenvalues ​​to obtain the reciprocal and multiply it by the product of the eigenvector matrix and the kernel function matrix. Use the column vectors in the result matrix as the feature coordinates of the projected manifold. S24. Read the manifold feature coordinates of the current time node and the manifold feature coordinates of the previous time node in the low-dimensional manifold feature space and calculate the coordinate difference. Use Euclidean distance as the tangent space distance between the two points, calculate the logarithm of the ratio of the tangent space distance to the hyperbolic tangent function, multiply the tangent space direction vector by the logarithm, and calculate the geodesic vector from the previous time node to the current time node. Connect the geodesic vectors of these two nodes to form the geodesic trajectory of the current progress state in the low-dimensional manifold feature space. S25. Preset the value of the standard progress reference coordinate point, read the coordinate of each manifold feature in the low-dimensional manifold feature space, calculate the difference between the manifold feature and the standard progress reference coordinate point in each dimension, and sum the squared values ​​to obtain the squared distance value. Multiply the squared distance value by negative half and take the negative number to obtain the potential energy value corresponding to the manifold feature coordinate point. Combine all potential energy values ​​to construct a dynamic risk potential energy field. S26. Extract three adjacent trajectory points on the geodesic trajectory. Subtract the coordinates of the previous point from the coordinates of the middle point to obtain a first-order difference vector. Subtract the coordinates of the middle point from the coordinates of the next point to obtain another first-order difference vector. Calculate the vector product of the two first-order difference vectors and take the magnitude. Calculate the cube of the magnitude of the first-order difference vector. Divide the magnitude of the vector product by the cube of the magnitude of the first-order difference vector to obtain the curvature parameter. S27. Read the potential energy value of the current coordinate point in the dynamic risk potential energy field and take the absolute value. Multiply the curvature parameter by the preset first weight coefficient to obtain the first weighting term. Multiply the absolute value of the potential energy value by the preset second weight coefficient to obtain the second weighting term and add it to the first weighting term to obtain the progress deviation value. Combine the progress deviation values ​​of all points to generate a manifold geometric feature set.

4. The method for predicting and evaluating the progress risk of construction projects based on big data analysis according to claim 1, characterized in that, S3 specifically includes: S31. Read the BIM model progress data containing geographic information, extract the task name and number in the construction plan as process nodes, read the IoT sensor time series data, extract the unique identifier of mechanical equipment and materials as resource nodes, extract the unique identifier of environmental monitoring equipment as environmental nodes, and summarize the process nodes, resource nodes and environmental nodes to construct a set of vertices of a heterogeneous graph. S32. Extract the preceding and following logical relationships of process nodes in the BIM model progress data. Extract the planned start and end times of two process nodes and calculate the difference. If it is greater than or equal to zero, it is determined that there is a dependency relationship. Connect the two process nodes with preceding and following dependencies with directed line segments to construct process logical dependency edges. S33. Extract records of material delivery from resource nodes to process nodes from IoT sensor time-series data, calculate the material delivery quantity, and when the delivery quantity is greater than zero, establish a supply and demand connection line between the corresponding resource node and process node to construct a resource supply and demand relationship edge. S34. Extract the overlapping information of environmental nodes and process nodes in time and space, compare the monitoring values ​​of environmental nodes with the preset monitoring thresholds of process nodes, establish constraint connection lines when the preset monitoring thresholds are exceeded, construct environmental constraint relationship edges, summarize the process logic dependency edges, resource supply and demand relationship edges and environmental constraint relationship edges, and construct a heterogeneous progress association graph. S35. Perform the mapping of manifold geometric feature set to the vertex attributes of the heterogeneous progress graph. Read the manifold geometric feature set and assign the feature values ​​in the manifold geometric feature set as vertex attributes to the vertices in the heterogeneous progress graph according to the timestamp and spatial position of the vertices.

5. The method for predicting and evaluating the progress risk of construction projects based on big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Pre-construct a heterogeneous graph transformer containing a sequentially connected heterogeneous graph attention layer, causal intervention layer, and feature output layer. Input the heterogeneous progress association graph into the heterogeneous graph attention layer, read the connection edges between vertices in the heterogeneous progress association graph, and for each connection edge, calculate the dot product of the feature vector of the source node and the feature vector of the target node and divide it by the square root of the dimension to obtain the attention score. Then input the score into the exponential function to calculate the normalization coefficient. Multiply the normalization coefficient by the feature vector of the source node to obtain the weighted node features. S42. Construct a causal attention unit, preset a risk-free counterfactual state, replace the monitoring values ​​of environmental nodes in the heterogeneous progress association graph with zero, and set the supply and demand values ​​of resource nodes to the theoretical maximum value to generate a risk-free counterfactual heterogeneous graph. S43. Input the weighted node features into the causal intervention layer, introduce the counterfactual learning mechanism, read the actual observed state features of the heterogeneous progress association graph, calculate the vector difference between the actual observed state features and the risk-free counterfactual heterogeneous graph features, and use it as the causal effect estimate. In the causal intervention layer, set the causal effect estimate of the environmental nodes as a noise interference term and remove it from the actual observed state features to obtain the pure node features after removing environmental noise. S44. Calculate the variance of the pure node features, divide the pure node features by the variance for standardization, and then perform corresponding operations with the preset linear transformation matrix and bias vector to deeply mine the core causal features that lead to schedule risks. S45. Input the core causal features into the feature output layer and multiply them with the preset weight matrix to obtain the contribution value. Normalize the contribution value to obtain the weight value within the interval. Multiply the weight value with the core causal features to generate the node feature code and output it.

6. The method for predicting and evaluating the progress risk of construction projects based on big data analysis according to claim 1, characterized in that, S5 specifically includes: S51. Construct an improved neural stochastic differential equation, define the differential dynamics equation to include drift term, diffusion term and Brownian motion term, preset the number of layers and neurons of the drift term neural network, initialize the weight matrix and bias vector in the drift term neural network to random values, and input the node feature encoding into the improved neural stochastic differential equation in the order of time steps. S52. Introduce a drift term guidance mechanism based on heterogeneous graph neural network, read the adjacency matrix of each node in the heterogeneous progress association graph, extract the neighbor node number directly connected to each node, retrieve the corresponding feature vector in the node feature encoding containing causal weight according to the neighbor node number, sum and average the retrieved neighbor node feature vectors to generate the initial graph embedding vector. S53. Input the initial graph embedding vector into a two-layer pre-defined fully connected neural network, perform matrix multiplication, calculate the product of the initial graph embedding vector and the first layer pre-defined weight matrix, and superimpose the first layer bias vector. After processing by the Tanh activation function, input it into the second layer, calculate the product with the second layer pre-defined weight matrix, and superimpose the second layer bias vector to output the high-dimensional graph embedding vector. S54. Input the high-dimensional graph embedding vector into the feedforward layer of the drift term neural network, read the original weight matrix to be updated of the drift term neural network, multiply the high-dimensional graph embedding vector with the weight matrix to be updated element by element, use the matrix value after multiplication to replace the parameter value in the weight matrix to be updated, calculate the product of the replaced weight matrix and the node feature encoding containing causal weights input in S51 and add the bias term, use the topological structure of the heterogeneous progress association graph to constrain the deterministic trend of risk evolution, and generate the guiding drift term value. S55. Read the standard normal distribution random number generator, generate a random noise array with the same dimension as the node feature encoding of the current time step input, multiply the random noise array with the preset noise coefficient of the diffusion term, calculate the diffusion fluctuation value, perform addition operation on the value of the guiding drift term and the diffusion fluctuation value, and obtain the derivative dynamic value of the risk of the current time step changing with continuous time. S56. Set the end time point of the predetermined future time window, construct a differential dynamic equation describing the cascading transmission of risk in the heterogeneous progress relationship diagram, use the state value of the current time step and the dynamic value of the derivative to calculate the predicted state value of the next time step, repeat the calculation steps until the accumulated time step reaches the end time point of the predetermined future time window, and output all intermediate calculation results as the future risk state sequence. S57. Connect the values ​​in the future risk state sequence with a broken line according to the time sequence, draw the evolution trajectory of risk over time, compare each value in the evolution trajectory with the preset risk threshold one by one, and mark the time point when the value first exceeds the preset risk threshold as the key risk time point. S58. Based on the key risk time points, locate the process node where the risk occurs in the heterogeneous schedule association diagram, read the directed edge connection relationship in the heterogeneous schedule association diagram, start from the located process node, backtrack along the opposite direction of the directed edge to search for the parent node, arrange all the parent nodes on the path in the connection order, and generate the risk transmission path. S59. The proportion of data points in the future risk state sequence whose values ​​exceed the preset risk threshold to the total number of data points is used as the probability of risk occurrence. The planned working hours of all process nodes on the risk transmission path are read and summed to obtain the total time. The total time is subtracted from the standard working hours to calculate the delay time. The probability of risk occurrence, the risk transmission path and the delay time are combined to output the dynamic evolution prediction result.

7. The method for predicting and evaluating the progress risk of construction projects based on big data analysis according to claim 6, characterized in that, S56 specifically includes: S561. Read the system configuration parameters, extract the total duration of the preset future time window, divide the total duration by the preset single-step time step, calculate the total number of iterations to be executed, and store the value of the total number of iterations as the loop termination condition. S562. Construct a differential dynamic equation describing the cascading transmission of risk in a heterogeneous progress correlation diagram. Set the node feature encoding containing causal weights as the initial state variable of the equation, set the dynamic value of the derivative as the rate of change parameter of the state variable, read the state value of the node feature encoding at the current time step, multiply it with the dynamic value of the derivative, multiply the product result with the single-step time step size, and calculate the deterministic drift increment. S563. Call the random number generation function to generate random numbers that follow a standard normal distribution, and multiply them by the preset diffusion term coefficient and the square root of the single-step time step to obtain the random diffusion increment. Perform an addition operation between the deterministic drift increment and the random diffusion increment, and add the calculation result to the state value of the node feature encoding at the current time step to obtain the predicted state value for the next time step. S564. Update the predicted state value of the next time step to the state value of the node feature encoding of the new current time step, accumulate the single-step time step length to the total running time, and record the predicted state value to the result storage list. Determine whether the total running time has reached the predetermined future time window termination time point. If not, return to step S562 to continue the calculation; otherwise, terminate the loop calculation process. Read all the predicted state values ​​recorded in the result storage list, combine them in chronological order, and output the future risk state sequence.

8. The method for predicting and evaluating the progress risk of construction projects based on big data analysis according to claim 1, characterized in that, S6 specifically includes: establishing a knowledge graph database of construction engineering risks, performing subgraph matching and semantic retrieval between the dynamic evolution prediction results and historical risk cases in the knowledge graph database, identifying the underlying causes of current risks and historically similar response strategies, and combining the probability of risk occurrence, risk transmission path and delay duration, underlying causes and historically similar response strategies into a comprehensive risk assessment report.

Images