A multi-dimensional intelligent management method and system for whole-process cost

Abstract

The application discloses a kind of multi-dimension intelligent management methods and systems of whole-process cost, method includes: according to the heterogeneous cost data of engineering planning, design, construction, completion each stage, generate the structured cost data cube of space-time correlation;Based on the structured cost data cube, output dynamic cost prediction curve and deviation sensitive node;According to the dynamic cost prediction curve, utilize the BIM / CIM collaborative platform of blockchain empowerment to carry out multi-party task allocation optimization, generate the collaborative instruction set of smart contract coding;Based on the collaborative instruction set and real-time engineering data flow, output risk probability matrix and early warning signal;According to the risk probability matrix, adopt multi-objective optimization algorithm to generate anti-interference decision scheme set, output optimal cost control strategy after through digital twin simulation verification.Using the embodiment of the application, the cost data of each stage of engineering can be efficiently integrated, dynamic cost prediction, optimization collaborative decision and effective risk management are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cost management, and particularly relates to a multi-dimensional intelligent management method and system for whole-process cost. BACKGROUND

[0002] In the field of construction engineering management, cost control and cost management are always the key factors affecting the success of a project. With the increasing complexity and scale of engineering projects, whole-process cost management is facing unprecedented challenges. In each stage of an engineering project, from planning, design to construction and completion, the cost data involved are often heterogeneous and scattered. How to efficiently integrate and manage these data to achieve scientific cost prediction and control has become a problem to be solved in the industry development.

[0003] Traditional cost management methods mainly rely on manual collection and analysis of data, which can easily lead to information lag, inaccurate cost prediction and other problems. At the same time, the lack of cross-stage collaborative management makes the data island phenomenon increasingly serious in different stages, and the information exchange between different participants (such as the owner, the design party, the construction party, etc.) is not smooth, which leads to untimely and inaccurate decision-making. In addition, due to the influence of market price fluctuations and construction progress changes and other factors, traditional cost management often cannot realize dynamic adjustment, so that the project is prone to overruns in the execution process, causing huge economic losses. SUMMARY

[0004] The purpose of the present application is to provide a multi-dimensional intelligent management method and system for whole-process cost, which can efficiently integrate the cost data of each stage of the project, realize dynamic cost prediction, optimize collaborative decision-making and effective risk management, and solve the problems in the prior art.

[0005] One embodiment of the present application provides a multi-dimensional intelligent management method for whole-process cost, which comprises:

[0006] According to the heterogeneous cost data of each stage of the project planning, design, construction and completion, cross-stage feature alignment is performed through a data fusion engine driven by federated learning to generate a structured cost data cube with spatio-temporal correlation, wherein the data fusion engine adopts a differential privacy protection mechanism to eliminate data islands;

[0007] Based on the structured cost data cube, a spatio-temporal graph neural network prediction model is constructed in combination with market price fluctuation trends and construction progress change factors to output a dynamic cost prediction curve and a deviation sensitive node, wherein the spatio-temporal graph neural network prediction model captures the associated influence of potential cost overrun risk factors through an attention mechanism;

[0008] According to the dynamic cost prediction curve, multi-party task allocation optimization is performed by using a BIM / CIM collaborative platform enabled by a blockchain, and a collaborative instruction set encoded by a smart contract is generated, wherein the BIM / CIM collaborative platform realizes private protection type data synchronization of the owner, the design party and the construction party through zero-knowledge proof;

[0009] Based on the collaborative instruction set and real-time engineering data flow, a potential risk path is analyzed by a dynamic knowledge graph engine, and a risk probability matrix and an early warning signal are output, wherein the dynamic knowledge graph engine realizes risk evolution simulation by fusing an industry knowledge base and a reinforcement learning strategy;

[0010] According to the risk probability matrix, a multi-objective optimization algorithm is used to generate an anti-interference decision scheme set, and an optimal cost control strategy is output after digital twin simulation verification, wherein the optimal cost control strategy synchronously drives a three-dimensional risk heat map and a decision path deduction animation in a visual interface.

[0011] Optionally, according to the heterogeneous construction cost data of the planning, design, construction and completion stages, cross-stage feature alignment is performed by a federated learning driven data fusion engine to generate a spatio-temporal related structured cost data cube, wherein the data fusion engine uses a differential privacy protection mechanism to eliminate data islands, including:

[0012] According to the BIM model parameters of the planning stage, the bill of quantities of the design stage, the progress log of the construction stage and the settlement documents of the completion stage, a four-dimensional tensor containing time stamp, spatial coordinate and cost attribute is generated by field semantic mapping through a heterogeneous data adapter, and a standardized feature vector set is output;

[0013] Based on the feature vector set, a cross-stage alignment algorithm under the federated learning framework is used to optimize the local model parameters of each participant through projection gradient descent, and distributed feature space alignment is completed under the differential privacy constraint to generate a cross-stage joint feature matrix;

[0014] The joint feature matrix is input into a spatio-temporal correlation engine, and the coupling relationship of engineering progress-cost consumption-geographical distribution is captured through a three-dimensional convolution kernel to build a six-dimensional cost data cube containing time sequence correlation, spatial topology correlation and cost transmission correlation, and output a structured data cube with a privacy protection label.

[0015] Optionally, based on the structured cost data cube, a spatio-temporal graph neural network prediction model is constructed by combining market price fluctuation trend and construction progress change factor to output a dynamic cost prediction curve and a bias sensitive node, wherein the spatio-temporal graph neural network prediction model captures the associated influence of potential cost overrun risk factors through an attention mechanism, including:

[0016] According to the time series cost data in the structured data cube, the cost fluctuation characteristics are extracted through a multi-scale sliding window, and a time dimension dynamic embedding vector is constructed in combination with an external market price index;

[0017] ‌Based on the construction progress change factor, the spatial dependence relationship between construction nodes is modeled through a graph attention mechanism to generate a spatial topological weight matrix;

[0018] ‌The time dynamic embedding vector and the spatial topological weight matrix are input into a spatio-temporal graph convolution layer, the spatio-temporal features are fused through a bidirectional gated recurrent unit, and a correlation influence intensity distribution map of the potential cost overrun risk factor is output;

[0019] ‌According to the risk influence intensity distribution map, a dynamic cost prediction curve with confidence annotation is generated by using a dynamic threshold segmentation algorithm to identify deviation sensitive nodes.

[0020] Optionally, according to the dynamic cost prediction curve, a BIM / CIM collaborative platform enabled by a blockchain is used for multi-party task allocation optimization to generate a collaborative instruction set encoded by a smart contract, wherein the BIM / CIM collaborative platform realizes private protection type data synchronization of the owner, the designer and the constructor through zero-knowledge proof, including:

[0021] According to the dynamic cost prediction curve, the private data such as the owner's budget constraint, the designer's drawing change and the constructor's resource scheduling are homomorphically encrypted through a zero-knowledge proof protocol to generate a verifiable ciphertext data set;

[0022] ‌Based on the ciphertext data set, a multi-party secure computing decomposition collaborative task is adopted, and a task allocation matrix with weight constraints is output by solving an optimal task allocation scheme under the multi-party interest balance through a Hungarian algorithm;

[0023] ‌The task allocation matrix is encoded into a smart contract, the contract triggering condition is defined through an event-driven state machine, and a smart contract template containing an execution logic chain is generated;

[0024] ‌A lightweight blockchain node is deployed on the BIM / CIM collaborative platform, private protection type data synchronization across participants is realized through a sharding consensus mechanism, and a collaborative instruction hash chain with a timestamp is generated;

[0025] ‌According to the real-time engineering data stream, the off-chain data authenticity is verified through an oracle, the smart contract execution state is dynamically updated, and a traceable collaborative instruction set is output.

[0026] Optionally, based on the collaborative instruction set and the real-time engineering data stream, a dynamic knowledge graph engine is used to analyze potential risk paths to output a risk probability matrix and an early warning signal, wherein the dynamic knowledge graph engine realizes risk evolution simulation by combining an industry knowledge base and a reinforcement learning strategy, including:

[0027] According to the cooperative instruction set, risk semantic fragments in engineering contract clauses and construction specification documents are extracted by a domain adaptive pre-training model to generate risk knowledge triples with probability labels;

[0028] ‌Based on the risk knowledge triples, in combination with abnormal events in real-time engineering data streams, a risk propagation path topology graph is constructed by a time series graph convolution network;

[0029] ‌The risk propagation path is input into a reinforcement learning environment, and a double deep Q network is used to simulate the risk evolution process to generate a risk probability transition matrix containing cascade effect intensity;

[0030] ‌Based on the risk probability transition matrix, a backward propagation importance sampling algorithm is used to identify key risk paths and output early warning signals with time window constraints;

[0031] ‌The risk model parameters are compressed by knowledge distillation technology and deployed to edge computing nodes for real-time risk monitoring to generate a lightweight risk evolution simulator.

[0032] Optionally, according to the risk probability matrix, a multi-objective optimization algorithm is used to generate a set of anti-interference decision schemes, and an optimal cost control strategy is output after digital twin simulation verification, wherein the optimal cost control strategy synchronously drives the three-dimensional risk heat map and decision path deduction animation in the visual interface, including:

[0033] According to the risk probability matrix, the NSGA-III algorithm is used to simultaneously optimize the cost control rate, the engineering quality compliance rate, and the risk avoidance coefficient, and an initial Pareto frontier containing 300-500 solutions is generated;

[0034] ‌The Pareto solution set is input into a digital twin engine, and the construction process under different decision schemes is simulated by a physics engine to generate a multi-dimensional simulation data set containing cost deviation, quality defects, and risk outbreak points;

[0035] ‌Based on the simulation data set, a fuzzy comprehensive evaluation method is used to calculate the fitness scores of each scheme, and a gradient boosting decision tree is used to select the optimal cost control strategy;

[0036] ‌The optimal strategy is fused with the BIM model, a three-dimensional risk heat map is rendered by ray tracing technology, and a decision path deduction animation is generated, and a visual decision support interface is output.

[0037] Optionally, based on the construction progress change factor, a graph attention mechanism is used to model the spatial dependency between construction nodes to generate a spatial topology weight matrix, including:

[0038] According to the node delay data in the construction progress change factor, the spatio-temporal correlation features between the construction nodes are extracted through the spatio-temporal graph convolution layer to generate an initial dependency graph with timestamps;

[0039] Based on the initial dependency graph, the spatial correlation weight between nodes is calculated by combining the geographic coordinate data in the BIM model and using a multi-modal attention mechanism, wherein the physical distance between nodes, the construction process dependency degree and the resource flow direction are fused to output a multi-modal attention score matrix;

[0040] The multi-modal attention score matrix is input into a dynamic residual connection network, and the implicit dependency relationship in the historical construction log is fused through a gated recurrent unit to generate a spatio-temporal enhanced node dependency strength distribution graph;

[0041] Based on the dependency strength distribution graph, a graph pruning algorithm with differential privacy protection is used to remove weak dependency edges with a confidence lower than a threshold, and a spatial topology weight matrix with a hierarchical structure is generated through spectral clustering to output a construction node spatial topology with a security protection label.

[0042] Optionally, based on the ciphertext data set, a multi-party secure computing decomposition collaborative task is used to solve the optimal task allocation scheme under the multi-party interest balance through the Hungarian algorithm to output a task allocation matrix with weight constraints, including:

[0043] According to the homomorphic encryption ciphertext data set, a collaborative task is generated through a secure multi-party computing protocol with privacy protection, wherein the task parameters are divided into multiple shadow shares by using secret sharing technology;

[0044] Based on the task demand vector, a multi-party interest game model is constructed, and an initial task allocation scheme is calculated through a Nash equilibrium solving algorithm, wherein the model constraint conditions include the upper limit of the owner's budget, the resource capacity of the construction party and the design party's change tolerance;

[0045] The initial allocation scheme is input into the Hungarian algorithm, and a dynamic weight adjustment strategy is introduced to optimize the conflict task matching, wherein the dynamic weight is updated adaptively according to the real-time supply chain risk index, and an intermediate matrix of task allocation with elastic constraints is output;

[0046] The fairness of the task allocation intermediate matrix is verified through zero-knowledge proof, and a verifiable allocation proof is generated by using the zk-SNARK protocol, and finally a task allocation matrix with weight constraints that meets the multi-party interest balance and cannot be tampered with is output.

[0047] Specifically, based on the simulation data set, the fitness score of each scheme is calculated by using the fuzzy comprehensive evaluation method, and the optimal cost control strategy is selected by using the gradient boosting decision tree, including:

[0048] According to the multi-dimensional engineering parameters in the simulation data set, the cost bias rate, the quality defect index, and the risk outbreak probability are adaptively weighted by a dynamic weight generator based on Monte Carlo sampling, and a dynamic evaluation weight vector with a confidence interval is generated;

[0049] Based on the dynamic evaluation weight vector, a fuzzy comprehensive evaluation algorithm is used to calculate the fitness score of the scheme, the fuzzy comprehensive evaluation algorithm introduces a time decay factor to correct the contribution of historical simulation data, and maps the high-dimensional nonlinear pattern through a radial basis function kernel, outputs a fitness score matrix with a timeliness label;

[0050] According to the fitness score matrix, a multi-modal feature fusion channel is constructed, the hyperplane curvature feature, the decision variable correlation feature and the simulation process stability feature of the Pareto solution set are extracted, and a unified dimension enhanced feature vector set is generated through a heterogeneous feature embedding network;

[0051] The enhanced feature vector set is input into a gradient boosting decision tree model, a multi-objective splitting criterion is used to optimize the tree structure growth direction, and a cost-sensitive learning strategy is used to enhance the search ability of the high-dimensional non-convex solution space, and the global ranking result of the optimal cost control strategy is output.

[0052] According to the global ranking result, combined with the physical constraint verification of the digital twin engine, the strategy parameters are iteratively optimized through the back propagation correction mechanism, and the optimal cost control strategy execution sequence that meets the construction feasibility boundary is generated.

[0053] Another embodiment of the present application provides a multi-dimensional intelligent management system for whole-process cost, which comprises:

[0054] An alignment module is configured to perform cross-stage feature alignment through a federated learning driven data fusion engine based on heterogeneous cost data at various stages of engineering planning, design, construction and completion, and generate a structured cost data cube with spatiotemporal correlation, wherein the data fusion engine uses a differential privacy protection mechanism to eliminate data islands.

[0055] A construction module is configured to construct a spatiotemporal graph neural network prediction model based on the structured cost data cube, combined with market price fluctuation trends and construction progress change factors, and output a dynamic cost prediction curve and a bias sensitive node, wherein the spatiotemporal graph neural network prediction model captures the associated influence of potential cost overrun risk factors through an attention mechanism.

[0056] An optimization module is configured to perform multi-party task allocation optimization using a BIM / CIM collaborative platform empowered by blockchain based on the dynamic cost prediction curve, and generate a collaborative instruction set encoded by a smart contract, wherein the BIM / CIM collaborative platform realizes private protection type data synchronization of the owner, the designer and the constructor through zero-knowledge proof.

[0057] a parsing module configured to parse potential risk paths through a dynamic knowledge graph engine based on the collaborative instruction set and real-time engineering data flow, and output a risk probability matrix and an early warning signal, wherein the dynamic knowledge graph engine fuses an industry knowledge base and a reinforcement learning strategy to realize risk evolution simulation;

[0058] an output module configured to generate an anti-interference decision scheme set using a multi-objective optimization algorithm according to the risk probability matrix, and output an optimal cost control strategy after verification through digital twin simulation, wherein the optimal cost control strategy synchronously drives a three-dimensional risk heat map and a decision path deduction animation in a visual interface.

[0059] Compared with the prior art, the multi-dimensional intelligent management method for whole-process cost provided by the application can generate a structured cost data cube associated with time and space according to heterogeneous cost data at various stages of engineering planning, design, construction and completion; based on the structured cost data cube, a dynamic cost prediction curve and a deviation sensitive node are output; according to the dynamic cost prediction curve, multi-party task allocation optimization is performed using a BIM / CIM collaborative platform empowered by a blockchain to generate a collaborative instruction set encoded by a smart contract; based on the collaborative instruction set and real-time engineering data flow, a risk probability matrix and an early warning signal are output; according to the risk probability matrix, an anti-interference decision scheme set is generated using a multi-objective optimization algorithm, and an optimal cost control strategy is output after verification through digital twin simulation, so that the cost data at various stages of engineering can be efficiently integrated, dynamic cost prediction, optimized collaborative decision and effective risk management can be realized. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A hardware structure block diagram of a computer terminal of the multi-dimensional intelligent management method for whole-process cost provided by the embodiment of the application is provided.

[0061] Figure 2 A flowchart of the multi-dimensional intelligent management method for whole-process cost provided by the embodiment of the application is provided.

[0062] Figure 3 A structure diagram of the multi-dimensional intelligent management system for whole-process cost provided by the embodiment of the application is provided. DETAILED DESCRIPTION

[0063] The embodiments described below with reference to the drawings are exemplary and are only used to explain the application, and cannot be explained as a limitation of the application.

[0064] The embodiment of the application first provides a multi-dimensional intelligent management method for whole-process cost, which can be applied to electronic devices such as computer terminals, specifically, common computers and the like.

[0065] The following will be described in detail taking the computer terminal as an example. Figure 1 A hardware structure block diagram of a computer terminal of a multi-dimensional intelligent management method for whole-process cost provided by an embodiment of the present application is shown in the figure. Figure 1 As shown in the figure, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0066] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions which, when executed, can cause the processor to execute any multi-dimensional intelligent management method for whole-process cost.

[0067] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0068] The internal memory provides an environment for the execution of the computer program in the non-volatile storage medium, which, when executed by the processor, can cause the processor to execute any multi-dimensional intelligent management method for whole-process cost.

[0069] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0070] It should be understood that the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0071] Referring to Figure 2 The embodiment of the present application provides a multi-dimensional intelligent management method for whole-process cost, which can include the following steps:

[0072] S201, according to the heterogeneous construction cost data of the planning, design, construction and completion stages, cross-stage feature alignment is performed through a federated learning driven data fusion engine to generate a spatiotemporally correlated structured cost data cube, wherein the data fusion engine adopts a differential privacy protection mechanism to eliminate data islands;

[0073] Specifically, according to the BIM model parameters of the planning stage, the bill of quantities of the design stage, the progress log of the construction stage, and the settlement documents of the completion stage, field semantic mapping is performed through a heterogeneous data adapter to generate a four-dimensional tensor containing a timestamp, a spatial coordinate, and a cost attribute, and output a standardized feature vector set;

[0074] The heterogeneous data adapter adopts a modular design and configures special parsers for data features of different stages:

[0075] BIM model parameter parsing: Extract component attributes (such as concrete strength grade C30, steel specification HRB400) from IFC (Industry Foundation Class) files, associate geometric data with material attributes through an IFC-to-RDF converter, and generate semantic triples with spatial coordinates (x, y, z). For example, a certain beam component is parsed as: (Component ID-2038, Type: Frame Beam, Coordinates (12.5, 8.2, 3.0), Concrete Volume 2.8 m³).

[0076] Bill of quantities processing: Use NLP technology (such as BERT-CRF model) to identify key fields (project code, unit, quantity) in the list text and match them with BIM component IDs. For example: (Project Code 010502003, Rectangular Column, m³, 15.6) → mapped to BIM component ID-2041.

[0077] Progress log integration: Parse the semi-structured text of the construction log (such as "2023-05-20, 3-layer beam slab pouring completed 80%"), extract the timestamp (20230520), progress value (80%), and associated component ID through regular expressions.

[0078] Settlement document standardization: Use OCR (Optical Character Recognition, accuracy ≥98%) to recognize the amount and visa number in the scanned documents and align them with the timestamp of the progress log.

[0079] Four-dimensional tensor construction:

[0080] Dimension 1 (Time): Unix timestamp (e.g. 1684540800);

[0081] Dimension 2 (Space): Three-dimensional coordinates (x, y, z) with a precision of 0.01 meters;

[0082] Dimension 3 (Cost): Sub-item cost (material cost, labor cost, etc.), unit Yuan;

[0083] Dimension 4 (Attribute): Classification code (e.g., 010101 - earthwork engineering).

[0084] The final output of the standardized feature vector set contains about 500,000 records, each record is a 128-dimensional vector (time 8 dimensions + space 24 dimensions + cost 32 dimensions + attribute 64 dimensions), stored in Apache Parquet format for efficient query.

[0085] ‌Based on the feature vector set, a cross-stage alignment algorithm under the federated learning framework is used to optimize the local model parameters of each participant through projected gradient descent, complete distributed feature space alignment under differential privacy constraints, and generate a cross-stage joint feature matrix;

[0086] The federated learning system includes three types of participant nodes:

[0087] Owner node: holds planning phase BIM data;

[0088] Design institute node: stores design phase bill of quantities;

[0089] Construction unit node: maintains construction / completion data.

[0090] Cross-stage alignment process:

[0091] Local feature extraction: each node uses the ResNet-50 model (without the last fully connected layer) to extract local features, outputting a 512-dimensional feature vector. For example, the design institute node maps the bill item "010101001 Excavation 100m³" to the vector [0.12, -0.05,..., 0.33].

[0092] Differential privacy protection: add Laplace noise (ε=0.5, sensitivity Δf=1.0) to the features before uploading, ensuring that individual data points cannot be distinguished. The noise amount calculation formula is: noise = Laplace(0, Δf / ε), the actual noise value range added is [-0.2, 0.2].

[0093] Projected gradient descent:

[0094] Initialize the global projection matrix W ∈ R^(512×256);

[0095] In each iteration, each node calculates the local gradient ∇W;

[0096] After gradient aggregation, update W ← W - η·∇W (learning rate η=0.01).

[0097] Feature space alignment: After 100 iterations, the cosine similarity of each node feature in the 256-dimensional shared space is above 0.85 (initial value 0.3).

[0098] Joint feature matrix generation: The aligned features are concatenated by timestamp to form a 2000x256 matrix (2000 key time points, 256-dimensional features per time point), and reduced to 64 dimensions through principal component analysis (PCA) to improve computational efficiency.

[0099] ‌The joint feature matrix is input into the spatio-temporal correlation engine, which captures the coupling relationship between project progress, cost consumption, and geographic distribution through a three-dimensional convolution kernel, constructs a six-dimensional cost data cube containing time series correlation, spatial topology correlation, and cost transmission correlation, and outputs a structured data cube with privacy protection labels.

[0100] The spatio-temporal correlation engine adopts a three-layer processing architecture:

[0101] Time series correlation layer:

[0102] Use 1D convolution kernel (width 7, step 1) to scan the time axis to extract weekly and monthly patterns. For example, detect cost peaks on the 25th of each month (related to progress payment).

[0103] Output time feature map (dimension: 50x64, 50 time segments).

[0104] Spatial topology correlation layer:

[0105] Use 3D convolution kernel (3x3x3) to process BIM coordinate data to identify spatial dependencies. For example, find that concrete overruns for underground 2nd floor columns will affect steel consumption for the 1st floor beams above.

[0106] Output spatial feature map (dimension: 30x30x30x32, corresponding to 30m x 30m x 30m spatial grid).

[0107] Cost transmission correlation layer:

[0108] Model the cost flow between sub-projects through graph attention network (GAT), with edge weights reflecting the transfer ratio of materials / labor. For example, the weight from steel processing (010516) to concrete pouring (010502) is 0.7.

[0109] Output graph embedding matrix (dimension: 100x16, 100 sub-projects).

[0110] Six-dimensional data cube construction:

[0111] Dimension 1-2: Time (year-month, e.g. 2023-05);

[0112] Dimension 3-5: Space (x, y, z, grid precision 1 meter);

[0113] Dimension 6: Cost transmission path (e.g., 010501→010502→010503).

[0114] Each cell stores the multi-dimensional feature splicing result (total dimension = 50+32+16=98), and generates a privacy protection tag through HMAC-SHA256. For example: (202305, (12,8,3), 010501→010502): [0.12,...,0.45], tag=0x3a7d...

[0115] The cube supports OLAP (Online Analytical Processing) queries such as "Structural engineering cost transmission path analysis in the area of coordinates (12,8,3) in May 2023", with a response time <200ms.

[0116] This step integrates heterogeneous data (such as BIM parameters, bill of quantities, progress logs, settlement documents) in the engineering life cycle (planning, design, construction, completion) through federated learning technology without sharing raw data, aligns cross-stage features using the projected gradient descent algorithm, and models the coupling relationship between time, space, and cost using a three-dimensional convolution kernel. Finally, a six-dimensional structured data cube is generated. The differential privacy mechanism injects noise during feature alignment to ensure the privacy of each participant's (such as the owner and the construction party) data, breaking the data silo problem in traditional cost management and achieving deep integration of cross-stage data; privacy protection technology enhances multi-party collaboration trust, providing a high-dimensional and high-precision data foundation for subsequent prediction and decision-making.

[0117] S202, based on the structured cost data cube, combining market price fluctuation trend and construction progress change factor, constructing a spatio-temporal graph neural network prediction model, outputting dynamic cost prediction curve and deviation sensitive node, wherein the spatio-temporal graph neural network prediction model captures the associated influence of potential cost overrun risk factors through attention mechanism;

[0118] Specifically, according to the time series cost data in the structured data cube, cost fluctuation features can be extracted through a multi-scale sliding window, and a time dimension dynamic embedding vector can be constructed by combining external market price indexes;

[0119] The structured cost data cube contains cost time series data of each stage of the project (such as daily material cost, labor cost). The multi-scale sliding window is designed in three levels:

[0120] Short-term window (7 days): Capture sudden cost fluctuations (such as a 5% daily increase in steel prices), use first-order difference to extract change rate features;

[0121] Mid-term window (30 days): Identify periodic fluctuations (e.g., cash flow tightness due to end-of-month settlements) by extracting dominant frequency components using Fast Fourier Transform (FFT);

[0122] Long-term window (90 days): Analyze trend changes (e.g., cost escalation due to inflation) by fitting a slope using linear regression.

[0123] External market price index integration:

[0124] Data sources: Access real-time data streams such as threaded steel futures prices (code RB), cement price indices (CEMPI), etc.;

[0125] Feature fusion: Align external indices with engineering cost data on the time axis using Dynamic Time Warping (DTW) algorithm to eliminate the impact of market data delays. For example, when steel market prices are delayed by 3 days affecting engineering procurement, DTW automatically shifts market data forward by 3 days to match.

[0126] Dynamic embedding vector generation:

[0127] Vector dimension: 128 dimensions, including short-term fluctuation intensity (16 dimensions), mid-term cycle phase (32 dimensions), long-term trend coefficient (16 dimensions), and external market correlation (64 dimensions);

[0128] Example: A weekly dynamic vector may be encoded as [0.35, -0.12,..., 0.08], indicating a 35% short-term steel cost increase, but a lower long-term trend than expected.

[0129] ‌Based on the construction progress change factor, the spatial dependence relationship between construction nodes is modeled through a graph attention mechanism to generate a spatial topology weight matrix;

[0130] Specifically, based on the construction progress change factor, the spatial dependence relationship between construction nodes is modeled through a graph attention mechanism to generate a spatial topology weight matrix, which can extract the spatio-temporal correlation features between construction nodes through a spatio-temporal graph convolution layer according to the node delay data in the construction progress change factor, and generate an initial dependency graph with timestamps;

[0131] The construction progress change factor includes the delay days of each node (such as "foundation pouring" and "steel structure installation") (such as node A delayed for 3 days), and the process dependence relationship (such as node B must be started after node A is completed). The spatio-temporal graph convolution layer is composed of spatial graph convolution (Spatial GCN) and temporal gate convolution (Temporal GRU), and the specific process is as follows:

[0132] Space graph construction: Take construction nodes as graph vertices, physical adjacency (such as node distance < 50 meters) and process dependency as edges, and construct an initial adjacency matrix. For example, node A and node B have process dependency, and the edge weight is set to 1.0; if only physically adjacent but no process relationship, the weight is set to 0.3.

[0133] Time feature extraction: Slide the time window (window size 7 days) on the delay data of each node (such as [3 days, 0 days, 5 days]), extract the time sequence features through GRU (hidden unit 64), and output the time embedding vector with dimension 128.

[0134] Space-time convolution fusion: Use two layers of GCN (output dimension 64 for each layer) to process spatial relationships, input the time embedding output by GRU as node features, and generate space-time fusion features through ReLU activation function. For example, the feature vector of node A combines its own delay of 3 days and the progress influence of its adjacent node B.

[0135] Dependency graph generation: Calculate the cosine similarity between node features, and establish a dependency edge for node pairs with similarity > 0.7, and finally generate an initial dependency relationship graph with timestamp. Each edge in the graph is labeled with dependency strength (0~1) and the latest update time (such as 2023-10-05 14:00).

[0136] Based on the initial dependency relationship graph, combined with the geographic coordinate data in the BIM model, a multi-modal attention mechanism is used to calculate the spatial correlation weight between nodes, which integrates node physical distance, construction process dependency and resource flow direction three channel features, and outputs a multi-modal attention score matrix;

[0137] The multi-modal attention mechanism processes different modal features through three independent attention heads, followed by a feature fusion layer:

[0138] Physical distance attention head:

[0139] Input: Euclidean distance between nodes (such as 35 meters from node A to B), normalized to [0, 1];

[0140] Calculation: Map to query (Query) and key (Key) through a fully connected layer (input 1D, output 16D), and get the attention score after dot product and scaling. For example, the score of nodes with a distance < 20 meters is improved to 0.9.

[0141] Process dependency attention head:

[0142] Input: Process dependency strength (from step 1 dependency strength value);

[0143] Calculation: Multi-head attention (4 heads) is used, each head maps dependency strength to a 16-dimensional vector, and cross-attention is calculated. For example, one head of the strong dependency (0.8) node pair reaches 0.95.

[0144] Resource flow direction attention head:

[0145] Input: Resource flow direction (e.g. "concrete pump truck from node A to B" is encoded as a direction vector);

[0146] Calculation: Direction encoding matrix (8 dimensions) is used to represent flow direction, and directional features are extracted through bidirectional GRU (hidden units 32) to calculate attention scores.

[0147] Feature fusion: Three head attention scores are concatenated (3x16=48 dimensions), fused through a fully connected layer (output 16 dimensions), and then normalized through Softmax to generate a multi-modal attention score matrix. For example, the final attention score of node A and B is 0.88, reflecting their strong spatial correlation.

[0148] Input the multi-modal attention score matrix into the dynamic residual connection network, and fuse the implicit dependency relationships in the historical construction log through the gated recurrent unit to generate a spatio-temporal enhanced node dependency strength distribution map;

[0149] The dynamic residual connection network is composed of residual blocks and GRU modules:

[0150] Residual block processing:

[0151] Input: Multi-modal attention score matrix (N x N, N is the number of nodes);

[0152] Operation: Two convolutional layers (kernel size 3x3, channel number 64→128), residual skip connection preserves original features;

[0153] Output: 128-channel enhanced attention feature map.

[0154] Historical log fusion:

[0155] Data: Past 30 days of construction logs (e.g. "node C delayed for 2 days due to material shortage"), encoded as a time series vector (dimension 64);

[0156] GRU processing: Bidirectional GRU (hidden units 64) is used to extract time series dependency features, outputting a 128-dimensional implicit relationship vector.

[0157] Feature concatenation and gating:

[0158] Concatenate the residual feature map and GRU output to get a 256-dimensional joint feature; control information flow through the gating mechanism (Sigmoid function).

[0159] Dependency strength distribution generation: Full connection layer mapping is performed on the gated output (256→1 dimension), and the dependency strength value (0~1) is converted through the Sigmoid function to generate the spatiotemporal enhanced distribution. For example, the dependency strength of nodes A and B is increased from 0.8 to 0.92, reflecting the influence of their frequent joint action in historical logs.

[0160] Based on the dependency strength distribution, a graph pruning algorithm with differential privacy protection is used to remove weak dependency edges with confidence lower than the threshold, and a spatial topology weight matrix with hierarchical structure is generated through spectral clustering, and the construction node spatial topology with security protection label is output.

[0161] Differential privacy pruning:

[0162] Noise injection: Add Laplace noise (noise scale ϵ=0.1) to the dependency strength, and adjust the strength value: strength′=strength+Laplace(0,1 / ϵ).

[0163] Threshold filtering: Remove edges with strength <0.4 (such as node C-D strength 0.35 removed), and keep high-confidence dependency relationships.

[0164] Spectral clustering layering:

[0165] Constructing Laplacian matrix: L=D−W, where D is the degree matrix and W is the pruned weight matrix;

[0166] Eigenvalue decomposition: Calculate the first three smallest eigenvectors of L to form a low-dimensional embedding (3D);

[0167] K-means clustering: Divide the nodes into 5 clusters (such as "basic construction cluster" "renovation cluster"), and the weighted average of the dependency strength within the cluster is used as the hierarchical weight.

[0168] Topology matrix generation:

[0169] Intra-level weight: The weight between nodes in the same cluster is increased by 20% (such as from 0.8→0.96);

[0170] Inter-level weight: The weight of cross-cluster nodes is reduced by 30% (such as from 0.6→0.42);

[0171] Security label addition: Label each edge with privacy budget consumption (such as ϵ=0.05) and cluster attribution label.

[0172] Example output:

[0173] Nodes A, B, and C belong to the "structure construction cluster", and the internal weight is 0.92~0.95;

[0174] Node D, E belong to "Device Installation Cluster", with a cross-cluster weight of 0.4;

[0175] All edges are annotated with a differential privacy parameter ϵ = 0.1, meeting the ISO 27001 data security standard.

[0176] ‌The time dynamic embedding vector and the spatial topology weight matrix are input into the spatio-temporal graph convolution layer. The spatio-temporal features are fused through a bidirectional gated recurrent unit, and the correlation influence intensity distribution map of the latent over-reach risk factor is output.

[0177] Customized design of spatio-temporal graph convolution (ST-GCN):

[0178] Spatial convolution:

[0179] Neighborhood aggregation based on topology weight matrix, each node aggregates its 3-hop neighbor features;

[0180] Use Chebyshev polynomial to approximate the graph convolution kernel (order K = 3) to reduce computational complexity.

[0181] Temporal convolution:

[0182] Use dilation causal convolution (dilation = 2) along the time axis to capture long-term temporal dependencies;

[0183] Convolution kernel size = 5, step = 1, output time dimension compressed to 1 / 3 of the original length.

[0184] Improvement of bidirectional gated recurrent unit (Bi-GRU):

[0185] Forward GRU: learn the influence of historical cost data on the current situation (such as last month's steel price rise leading to current budget tightness);

[0186] Reverse GRU: predict the reverse constraints of future risks on current decisions (such as anticipating labor shortage in three months and need to reserve in advance);

[0187] Gating mechanism: control information forgetting and updating through sigmoid function (output range 0~1). For example, when detecting abnormal fluctuations, the forgetting gate value drops to 0.2, forcing to retain historical stable patterns.

[0188] Correlation influence intensity distribution map generation:

[0189] Heat map generation: map the risk influence intensity (scalar value) of each node to the corresponding location of the BIM model;

[0190] Dynamic threshold segmentation:

[0191] Red area (intensity > 0.8): immediate intervention is required (such as "main structure node" with a risk value of 0.91 due to steel price rise);

[0192] Yellow zone (0.5~0.8): Need monitoring (e.g. "curtain wall installation" risk value 0.63);

[0193] Green zone (<0.5): Safe zone.

[0194] ‌According to the risk intensity distribution map, the deviation sensitive node is identified by using a dynamic threshold segmentation algorithm, and a dynamic cost prediction curve with confidence annotation is generated.

[0195] Dynamic threshold segmentation algorithm:

[0196] Initial threshold setting:

[0197] Based on historical engineering data statistics, take the 90th percentile (e.g. 0.75) of the risk intensity distribution as the baseline threshold;

[0198] Adaptive adjustment:

[0199] When more than 5 nodes break through the threshold for 3 consecutive days, the threshold is automatically lowered by 10% (e.g. to 0.675);

[0200] When no node breaks through for 7 consecutive days, the threshold is raised by 5%.

[0201] Sensitive node identification:

[0202] Feature extraction: calculate for each node over the threshold:

[0203] Risk transmission speed (change rate of risk intensity of adjacent nodes);

[0204] Resource sensitivity (cost increase caused by unit resource shortage of this node);

[0205] Decision tree classification: use C4.5 algorithm to classify nodes into:

[0206] Key sensitive nodes (need to be handled within 48 hours);

[0207] General sensitive nodes (need to be monitored weekly).

[0208] Dynamic cost prediction curve generation:

[0209] Baseline prediction: generate the cost trajectory without intervention based on the ARIMA model (p=2, d=1, q=1);

[0210] Risk correction: use the risk intensity of sensitive nodes as a correction factor to adjust the predicted value by ridge regression (λ=0.5);

[0211] Confidence calculation:

[0212] Use Bootstrap sampling to generate 100 groups of prediction sequences;

[0213] Compute 95% confidence intervals, e.g. "Q3 predicted cost = 285 million ± 12 million (confidence 92%)".

[0214] Visualize the output:

[0215] Curve annotation: annotate sensitive node names and response suggestions at key time points (e.g. milestone nodes);

[0216] Interactive functions: click on sensitive nodes to drill down and view associated risk factor details.

[0217] This step combines structured data cubes with external market indices, construction change factors, and builds a spatio-temporal graph neural network model. The model extracts time dimension cost fluctuation characteristics through multi-scale sliding window, and analyzes the spatial dependence relationship between construction nodes using graph attention mechanism, finally outputs dynamic cost prediction curve and key risk nodes (such as material price sensitive period, process delay transmission path), accurately predicts the dynamic changes of engineering cost, identifies the "fragile nodes" that may cause cost overrun; through spatio-temporal correlation analysis, it reveals the implicit risks (such as the cascading effect of supply chain delay on multiple processes), and supports early intervention.

[0218] S203, according to the dynamic cost prediction curve, using the BIM / CIM collaborative platform enabled by blockchain to optimize multi-party task allocation, generating a collaborative instruction set encoded by smart contract, wherein the BIM / CIM collaborative platform realizes the privacy protection type data synchronization of the owner, the designer and the constructor through zero-knowledge proof;

[0219] Specifically, according to the dynamic cost prediction curve, the privacy data of the owner's budget constraint, the designer's drawing change, the constructor's resource scheduling, etc. can be homomorphic encrypted through zero-knowledge proof protocol to generate a verifiable ciphertext data set;

[0220] Zero-knowledge proof protocol (Zero-Knowledge Proof, ZKP) is used in this stage to realize the secure verification of multi-party privacy data. Taking zk-SNARK (zk-SNARK) as an example, the sensitive data of the owner's budget constraint (such as total budget 5000 million), the designer's drawing change record (such as structural modification times 3), the constructor's resource scheduling plan (such as daily supply of concrete 200 tons) are first encrypted by Paillier homomorphic encryption algorithm to generate ciphertext form. For example, the owner's budget plaintext 5000 million is encrypted to ciphertext E(5000)=0x3a7d...f2c1, ensuring that the data cannot be reversed in the transmission and calculation process.

[0221] Implementation details of homomorphic encryption:

[0222] Key generation: 2048-bit RSA key pair is used, the public key is used for encryption, and the private key is kept by the data owner;

[0223] Encryption operation: scalar encryption is performed on numerical data (such as budget amount), and block encryption is performed on text data (such as design change description);

[0224] Verifiability: proof parameters (CRS, common reference string) are generated by the Groth16 protocol of zero-knowledge proof to ensure that the encryption process meets the protocol specification.

[0225] Generation of verifiable ciphertext data set:

[0226] Data sharding: the encrypted data is split into multiple shadow shards according to the participants (owner, designer, and constructor), for example, the owner data shards are [S1, S2, S3], which are stored in different blockchain nodes;

[0227] Proof generation: each participant uses zk-SNARK to generate a proof, for example, the owner proves that the budget ciphertext E(5000) indeed corresponds to the plaintext 50 million yuan, and does not exceed the preset range (such as 100 million upper limit);

[0228] Verification aggregation: the verification node verifies the validity of all shards through elliptic curve pairing (EC Pairing) to finally generate a globally verifiable ciphertext set VSet = {E(Budget), E(Design), E(Resource)}.

[0229] Based on the ciphertext data set, the collaborative task is decomposed by multi-party secure computation, and the optimal task allocation scheme under the multi-party interest balance is solved by the Hungarian algorithm, and a task allocation matrix with weight constraint is output;

[0230] Specifically, based on the ciphertext data set, the collaborative task is decomposed by multi-party secure computation, and the optimal task allocation scheme under the multi-party interest balance is solved by the Hungarian algorithm, and a task allocation matrix with weight constraint is output, which can generate a task demand vector with privacy protection according to the ciphertext data set of homomorphic encryption through secure multi-party computation protocol, wherein the task parameters are divided into multiple shadow shares by using secret sharing technology;

[0231] In the engineering collaborative task allocation, the demand data (such as budget, resource capacity, and change tolerance) of the owner, designer, and constructor is converted into a ciphertext data set by homomorphic encryption (such as Paillier algorithm). For example, the owner's budget constraint "50 million yuan" is encrypted as ciphertext C1, the designer's drawing change limit "3 times" is encrypted as C2, and the constructor's equipment scheduling capacity "20 sets / day" is encrypted as C3.

[0232] Secure Multi-Party Computation (MPC) protocol uses Shamir secret sharing technology to decompose task parameters:

[0233] Parameter partitioning: each ciphertext parameter is partitioned into 5 shadow shares (threshold k=3), for example, the budget ciphertext C1 is split into shares S 11 , S 12 , …, S 15 , at least 3 shares are required to recover;

[0234] Distributed storage: shares are stored in 5 independent nodes (such as blockchain shard nodes) to ensure that single point failure cannot leak data;

[0235] Task demand vector generation: each participant collaborates through MPC protocol to generate task demand vector without exposing plaintext. For example, when calculating "total budget ≥ equipment procurement cost + labor cost", the ciphertext is directly operated through the additive homomorphism property, and the output is an encrypted Boolean value (1 represents satisfaction, 0 represents dissatisfaction).

[0236] The final generated task demand vector contains the following privacy protection fields:

[0237] Demand 1: Equipment procurement ciphertext budget (threshold 3 / 5);

[0238] Demand 2: Design change frequency limit (threshold 2 / 5);

[0239] Demand 3: Construction resource elasticity coefficient (range 0.5~1.2).

[0240] Based on the task demand vector, a multi-party interest game model is constructed, and the initial task allocation scheme is calculated through Nash equilibrium algorithm, wherein the model constraint conditions include the upper limit of the owner's budget, the resource capacity of the construction party, and the tolerance of the design party to change;

[0241] The multi-party interest game model models the owner, the design party, and the construction party as three game participants, and the objective functions are:

[0242] Owner: Minimize total cost (objective function f1 = Σ(equipment cost + labor cost));

[0243] Design party: Minimize design change frequency (f2 = Σ(drawing modification frequency));

[0244] Construction party: Maximize resource utilization rate (f3 = equipment utilization rate × work efficiency).

[0245] Constraint conditions:

[0246] Owner's budget ≤ 50 million yuan;

[0247] Design change ≤ 3 times;

[0248] Construction equipment scheduling ≤ 20 units / day.

[0249] Nash equilibrium solution using Alternating Direction Method of Multipliers (ADMM):

[0250] Variable splitting: Global variables (e.g., equipment allocation plan) are decomposed into local copies, and each participant optimizes their own objectives while considering consistency with other parties' copies.

[0251] Iterative update: In each iteration, each party alternately updates local variables and coordinates conflicts through Lagrange multipliers. For example, the owner proposes an equipment procurement plan in the tth round, and the construction party adjusts the equipment scheduling plan in the t+1th round to match the procurement plan.

[0252] Convergence criterion: Terminate when the objective function change rate is <0.1% or the maximum number of iterations (e.g., 100) is reached.

[0253] An example of the initial task allocation plan is as follows:

[0254] Equipment procurement: 10 excavators (budget 12 million yuan);

[0255] Design changes: 2 times (involving structural reinforcement);

[0256] Resource scheduling: Equipment utilization rate 85%, work efficiency 0.9.

[0257] Input the initial allocation plan into the Hungarian algorithm, and introduce a dynamic weight adjustment strategy to optimize the matching of conflicting tasks. The dynamic weight is updated adaptively according to the real-time supply chain risk index, and the intermediate matrix of the task allocation with elastic constraints is output.

[0258] The Hungarian algorithm is used to solve the bipartite graph matching problem of tasks and resources. When constructing the cost matrix, dynamic weight adjustment is introduced:

[0259] Cost matrix definition: Rows represent tasks (e.g., earth excavation, structure pouring), columns represent resources (equipment, personnel), and matrix elements Cᵢⱼ represent the cost of task i using resource j.

[0260] Dynamic weight:

[0261] Supply chain risk index (0~100, calculated by logistics delay rate, raw material price fluctuation, etc.): When the index >70, the equipment scheduling cost weight increases by 30%.

[0262] Elastic constraints: The resource capacity of the construction party allows ±10% fluctuation, which is reflected in the expansion of the resource column in the matrix (e.g., "excavator" is split into "excavator-normal" and "excavator-elastic").

[0263] Hungarian algorithm optimization process:

[0264] Initial matching: find the minimal set of lines that cover all rows and columns with zero elements;

[0265] Adjust weights: update the cost matrix according to the real-time risk index, for example, when the risk index rises to 75, adjust the "earth excavation - excavator" cost from 100 to 130;

[0266] Iterative optimization: find the optimal matching by augmented path search, time complexity O(n³), support up to 1000 task nodes.

[0267] For example, the output task allocation intermediate matrix contains the following information as shown in Table 1:

[0268] Table 1

[0269]

[0270] The optimal matching is: earth excavation → excavator - flexible (cost 150), structure pouring → crane (200), pipeline laying → excavator - normal (180).

[0271] Verify the fairness of the task allocation intermediate matrix through zero-knowledge proof, generate a verifiable allocation proof using zk-SNARK protocol, and finally output a weighted constraint task allocation matrix that is tamper-proof and satisfies the balance of interests of multiple parties.

[0272] Zero-knowledge proof (zk-SNARK) is used to verify the fairness of the task allocation scheme, ensuring no malicious manipulation:

[0273] Arithmetic circuit construction: encode the allocation logic as a circuit, for example, verify "total cost ≤ budget" can be represented as a gate circuit: Σ (allocation selection × cost) ≤ 5000;

[0274] Trusted setup: generate proof key (pk) and verification key (vk), avoid single-point trust problem through multi-party computation ceremony (such as Tau ceremony);

[0275] Proof generation: input the task allocation intermediate matrix, generate proof π (about 1 KB), the proof allocation satisfies all constraints and does not leak sensitive data;

[0276] Verification: any verifier can confirm the validity of the scheme in milliseconds using vk and π, without knowing the specific allocation details.

[0277] The final output of the weighted constraint task allocation matrix contains:

[0278] Resource allocation details: encrypted and stored in the blockchain (such as Hyperledger Fabric);

[0279] Dynamic weight log: record the timing relationship between supply chain risk index and weight adjustment;

[0280] zk-SNARK proof: ensure that the scheme is tamper-proof and auditable.

[0281] For example, after the zk-SNARK proof of a certain allocation scheme is automatically verified by the smart contract, the following operations are triggered:

[0282] The owner releases the first installment to the smart contract escrow account;

[0283] The designer uploads the new version of the BIM model to the collaborative platform;

[0284] The construction party schedules equipment according to the matrix, and real-time data is stored on the chain.

[0285] Encode the task allocation matrix into a smart contract, define the contract trigger conditions through an event-driven state machine, and generate a smart contract template containing an execution logic chain;

[0286] The smart contract is coded in Solidity language and implemented automatically based on an event-driven architecture. Taking the "foundation pouring" task as an example, the contract logic includes:

[0287] State definition:

[0288] State 0: Task pending allocation;

[0289] State 1: Construction party confirms resources;

[0290] State 2: Owner pays advance payment;

[0291] State 3: Task execution;

[0292] State 4: Acceptance completed.

[0293] Event trigger:

[0294] Event_ResourceConfirm: Construction party uploads resource preparation proof (such as concrete detection report);

[0295] Event_Payment: Owner pays 30% advance payment through on-chain transfer;

[0296] Event_ProgressUpdate: Construction party submits daily progress (such as pouring volume 100m³).

[0297] Execution logic chain:

[0298] When Event_ResourceConfirm is triggered and the signature is verified, the state changes from 0 to 1;

[0299] When Event_Payment detects that the amount ≥ 30% of the budget, the state changes from 1→2;

[0300] When Event_ProgressUpdate accumulates ≥ 80% completion, the state changes from 3→4.

[0301] Smart contract template generation: automatically generate contract instances according to different task parameters (such as budget, duration), and deploy them to the blockchain network.

[0302] ‌Deploy lightweight blockchain nodes on the BIM / CIM collaboration platform, realize cross-participant privacy-protected data synchronization through sharding consensus mechanism, and generate timestamped collaboration instruction hash chain;

[0303] Lightweight blockchain nodes are optimized based on Hyperledger Fabric architecture, and each participant (owner, designer, constructor) deploys a Peer node to form a consortium chain. Sharding consensus uses the PBFT (practical Byzantine fault tolerance) algorithm, which divides the network into multiple shards (such as shard 1 handles design data, and shard 2 handles construction data), each shard contains 3 nodes (tolerant to 1 faulty node).

[0304] Data synchronization process:

[0305] Transaction proposal: participants submit data update requests (such as designers uploading new version BIM models), and generate transaction proposals TxProposal;

[0306] Intra-shard consensus: shard nodes reach agreement through PBFT three-round voting (Pre-prepare, Prepare, Commit), for example, shard 1 nodes A / B / C all agree that model version V2.3 is valid;

[0307] Cross-shard synchronization: through atomic commit protocol (Atomic Commit) to ensure the integrity of cross-shard transactions. For example, design model updates need to be synchronized to the construction shard, if any shard refuses, then roll back;

[0308] Hash chain generation: each transaction is packaged as a block, the block header contains the previous hash, timestamp (UNIX millisecond level), and Merkle root. For example, block #7821 records "design change V2.3", hash 0x5b9e...d4a3, timestamp 1630000000000.

[0309] Privacy protection mechanism:

[0310] Channel isolation: sensitive data (such as budget) is transmitted through private channels, only relevant participants can decrypt;

[0311] Data desensitization: Non-critical fields (such as log IDs) are hashed to avoid exposing business details.

[0312] ‌According to real-time engineering data flow, off-chain data authenticity is verified by the Oracle, the execution state of the smart contract is dynamically updated, and a traceable collaborative instruction set is output.

[0313] The Oracle serves as a bridge for off-chain data, collecting real-time data from IoT devices (such as concrete humidity sensors) and ERP systems (such as resource inventory databases), and transmitting it to the blockchain through TLS (Transport Layer Security) encryption. Taking the "concrete pouring progress" as an example:

[0314] Data collection: sensors upload humidity data every 5 minutes (e.g., humidity 72%), and ERP systems update inventory every hour (e.g., remaining concrete 500 tons);

[0315] Oracle verification: using Chainlink Oracle nodes, cross-verification is performed through multiple data sources (e.g., comparing the average of 3 sensors) to ensure data authenticity;

[0316] Smart contract triggering: when humidity ≥ 70% and inventory ≥ 200 tons, trigger Event_ProgressUpdate to update the task status to "in execution";

[0317] Instruction set generation: all state change records generate instruction sequences in chronological order, for example: [

[0319] {Time: 1630000000, Cmd: "StartFoundation", Params: {Concrete: 200t}},

[0320] {Time: 1630003600, Cmd: "PauseDueToRain", Params: {Delay: 2h}},

[0321] {Time: 1630010800, Cmd: "ResumeWork"}

[0322] ]。

[0323] Traceability guarantee:

[0324] Full-link audit: each instruction is associated with a blockchain transaction hash, which can be traced back to the original data;

[0325] Digital signature: participants sign key instructions (e.g., construction party leader signature 0x3a7d...), ensuring traceability of responsibility.

[0326] Based on the cost prediction results, the zero-knowledge proof is used for homomorphic encryption of multi-party private data (such as owner budget, design change requirement, construction resource plan), the task allocation scheme is optimized through the Hungarian algorithm, and the intelligent contract is coded for automatic execution. The block chain sharding consensus mechanism ensures that the instruction set is synchronized and tamper-proof in the BIM / CIM platform, and realizes efficient cooperation of multiple parties on the premise of protecting commercial secrets; the intelligent contract automatically triggers the task instruction (such as material procurement, process adjustment), reduces the delay and error of human communication.

[0327] S204, based on the collaborative instruction set and real-time engineering data flow, the potential risk path is analyzed through the dynamic knowledge graph engine, and the risk probability matrix and early warning signal are output, wherein the dynamic knowledge graph engine integrates industry knowledge base and reinforcement learning strategy to realize risk evolution simulation;

[0328] Specifically, according to the collaborative instruction set, the risk semantic fragments in the engineering contract clauses and construction specification documents can be extracted through the domain adaptive pre-training model to generate risk knowledge triples with probability annotation;

[0329] The domain adaptive pre-training model adopts the RoBERTa model (Robustly optimized BERT model) based on the Transformer architecture, and performs secondary pre-training through engineering field professional corpus to adapt to engineering risk semantic understanding. The model input is the text paragraph of contract clauses and construction specifications (such as "concrete strength grade shall not be lower than C30"), and the output is the risk entity and its associated relationship.

[0330] Data preprocessing and training process:

[0331] Entity annotation: BIO (Begin-Inside-Outside) annotation system is used to define risk entity types (such as "material risk", "schedule risk", "safety risk") and relationships (such as "cause", "association"). For example, in the sentence "steel specification does not meet the standard may cause structural instability", "steel specification does not meet the standard" is marked as material risk (B-MAT, I-MAT), "structural instability" is marked as safety risk (B-SAFE, I-SAFE), and the relationship is "cause".

[0332] Domain adaptive training: based on the general Chinese corpus (such as WikiZh), engineering field text (100,000) is loaded, pre-training is performed using dynamic mask strategy (mask probability 15%), learning rate is set to 2e-5, batch size is 32, and training is performed for 3 cycles.

[0333] Probability labeling: When the model outputs an entity relationship, it generates a confidence probability (0~1) through the Softmax layer. For example, the confidence of "rebar specification not up to standard → leads to → structural instability" is 0.92.

[0334] Risk knowledge triplets generation:

[0335] Format: (head entity, relationship, tail entity, probability, time window). For example: (concrete strength deficiency, triggers, crack propagation risk, 0.85, [t0+7d, t0+14d]).

[0336] Time window derivation: Based on historical engineering data statistics, such as concrete strength problems usually trigger cracks 7~14 days after construction.

[0337] ‌Based on risk knowledge triplets, combined with abnormal events in real-time engineering data streams, a risk propagation path topology graph is constructed through a time-series graph convolutional network.

[0338] The time-series graph convolutional network (T-GCN) is composed of alternating time convolutional layers (TCN) and graph convolutional layers (GCN), which are used to model the propagation of risks in time and space dimensions.

[0339] Data processing and model construction:

[0340] Abnormal event detection: Real-time engineering data streams (such as sensor-monitored concrete humidity exceeding standards, progress delay alarms) extract features through sliding windows (window size 1 hour, step 10 minutes), and use the Isolation Forest algorithm to detect anomalies (anomaly score >0.6 is considered abnormal).

[0341] Graph initialization: The initial graph is constructed with risk knowledge triplets as nodes and edges, with node attributes including risk type, probability, and time window; edge weights are confidence.

[0342] Time-series graph convolution:

[0343] Time convolution layer: Dilated causal convolution (dilation factor 2, convolution kernel size 3) is used to capture the delayed effects of risks over time (such as material defects triggering progress risks after 3 days).

[0344] Graph convolution layer: Aggregate features of adjacent nodes, such as "concrete strength deficiency" node aggregating "cement grade error" and "improper curing" features.

[0345] Dynamic edge weight update: Adjust edge weights according to real-time abnormal events. For example, when "single-day progress delay ≥8%" is detected, the "schedule risk → cost overrun" edge weight is increased from 0.7 to 0.9.

[0346] Risk propagation path topology graph output:

[0347] Nodes: 200~500 risk points (depending on project scale);

[0348] Edges: Cascading paths (e.g. A→B→C) highlighted in red, independent risks in gray;

[0349] Topology attributes: Each path is annotated with propagation delay (e.g. "Material procurement delay → risk of downtime: 2 days").

[0350] ‌Risk propagation paths are input into a reinforcement learning environment, which simulates risk evolution using a double deep Q network (DDQN) to generate a risk probability transition matrix that includes the strength of cascading effects;

[0351] The reinforcement learning environment is customized based on the OpenAI Gym framework, and the agent learns risk intervention strategies through a double deep Q network (DDQN).

[0352] Environment settings:

[0353] State space: 300-dimensional vector including probabilities of each risk node, project progress, and resource reserves;

[0354] Action space: 20 intervention measures (e.g. "Increase concrete detection frequency", "Emergency purchase of steel bars");

[0355] Reward function: R = 0.5 × (1−risk probability) + 0.3 × progress compliance rate + 0.2 × cost saving rate.

[0356] DDQN training process:

[0357] Network structure: Both the main network and the target network are 3-layer fully connected (512-256-128) with ReLU activation, learning rate 0.001, and experience replay buffer capacity 10,000;

[0358] Training process: Sample 100 experiences (state-action-reward-new state) per iteration, update main network parameters, and synchronize target network every 100 steps;

[0359] Risk evolution simulation: The agent tries different actions in the virtual environment, such as selecting "Enable backup supplier" to reduce the "material shortage" risk probability from 0.8 to 0.3, but increase the cost by 5%.

[0360] Risk probability transition matrix generation:

[0361] Matrix dimensions: N × N (N is the number of risk nodes), element Pij represents the probability of risk i triggering risk j;

[0362] Cascade strength calculation: Statistically calculate the path activation frequency through Monte Carlo simulation (1000 times). For example, the cascade strength of the path "design change → schedule delay → cost overrun" is 0.75.

[0363] ‌Based on the risk probability transition matrix, the backpropagation importance sampling algorithm is used to identify the key risk path, and the early warning signal with time window constraint is output.

[0364] The backpropagation importance sampling (BP-IS) algorithm locates the critical path through gradient backpropagation and combines importance sampling to reduce computational complexity.

[0365] Algorithm steps:

[0366] Gradient calculation: Take the total risk probability as the objective function to calculate the gradient of each path. For example, the gradient of path A→B is ∂P_total / ∂P_AB=0.6, P_total is the total risk probability, and P_AB is the risk probability of path A→B, indicating that it has a greater impact on the overall risk.

[0367] Importance sampling: Assign sampling weights according to the absolute value of the gradient, and preferentially explore high-impact paths (such as the sampling probability of the top 10% of paths is increased to 50%);

[0368] Key path screening: Define the criticality index CI=gradient×frequency / cost, and screen paths with CI>0.5.

[0369] Early warning signal generation:

[0370] Signal format: JSON structure, containing risk path, expected outbreak time, and recommended measures. For example:

[0371] {

[0372] "risk_path": ["concrete strength deficiency", "structural crack", "acceptance failure"],

[0373] "time_window": ["2023-10-05", "2023-10-12"],

[0374] "action": "immediately conduct third-party strength detection"

[0375] }。

[0376] Dynamic push: Real-time transmission to management personnel mobile terminals through 5G messages (5G MSG), with a delay of <200ms.

[0377] ‌Through knowledge distillation technology to compress risk model parameters, deploy to edge computing nodes to realize real-time risk monitoring, and generate lightweight risk evolution simulator.

[0378] Knowledge distillation adopts a teacher-student model architecture, compressing the complex T-GCN+DDQN model (teacher) into a lightweight simulator (student).

[0379] Distillation process:

[0380] Teacher model output: prediction results on historical risk cases (such as risk probability, propagation path);

[0381] Student model construction: adopts MobileNetV3 small architecture (parameter quantity 1M), input dimension compressed to 64;

[0382] Loss function: combines cross-entropy (student output and true label) and KL divergence (student and teacher output distribution similarity), weight ratio 6:4;

[0383] Training configuration: learning rate 0.01, batch size 64, training 50 rounds.

[0384] Edge deployment and optimization:

[0385] Hardware adaptation: deployed to Huawei Atlas 500 edge intelligent small station (computing power 16TOPS), memory occupancy <500MB;

[0386] Real-time inference: input real-time data (such as "today's progress delay 10%"), output risk warning within 1 second;

[0387] Dynamic update: synchronize new knowledge of cloud teacher model every week, incrementally update student model parameters.

[0388] Lightweight simulator performance:

[0389] Accuracy retention: compared with the teacher model, the student model risk prediction accuracy decreases ≤3% (F1-score from 0.91 to 0.88);

[0390] Latency optimization: single inference time from 2.1 seconds to 0.3 seconds, meeting the real-time requirements of engineering sites.

[0391] Dynamic knowledge graph engine extracts risk knowledge triples (such as "material shortage → delay of construction period → penalty for breach of contract") from engineering contracts and construction specifications, combines real-time data (such as logistics abnormal events) to construct risk propagation path topology, and simulates risk evolution process through reinforcement learning, outputs probabilistic risk matrix (such as the probability of cascading risk outbreak) and early warning signals (such as 3 weeks in advance warning of concrete supply interruption), combines industry experience and real-time data to quantify risk transmission path; through simulation and deduction, realize risk "predictive management", support to develop targeted prevention and control strategies.

[0392] S205, according to the risk probability matrix, a multi-objective optimization algorithm is used to generate an anti-interference decision scheme set, and an optimal cost control strategy is output after digital twin simulation verification, wherein the optimal cost control strategy synchronously drives the three-dimensional risk heat map and the decision path deduction animation in the visual interface.

[0393] Specifically, according to the risk probability matrix, the NSGA-III algorithm is used to simultaneously optimize the cost control rate, the engineering quality compliance rate, and the risk avoidance coefficient, and an initial Pareto frontier containing 300-500 solutions is generated.

[0394] NSGA-III (the third generation of non-dominated sorting genetic algorithm) is used to find the Pareto optimal solution set in multi-objective optimization problems. In this step, the optimization objectives are the cost control rate (objective 1), the engineering quality compliance rate (objective 2), and the risk avoidance coefficient (objective 3).

[0395] Parameter setting and population initialization:

[0396] Population size: set to 500 to cover a wider solution space;

[0397] Variable dimension: contains 10 decision variables such as material procurement cost (0-100% of budget), construction period (10-365 days), and risk response resource allocation ratio (0-100%);

[0398] Reference point generation: use the Das-Dennis method to generate uniformly distributed reference points, with a quantity of 100 to ensure uniform distribution of the solution set in the objective space.

[0399] Genetic operation:

[0400] Crossover: use simulated binary crossover (SBX) with a crossover probability of 0.9 and a distribution index η = 20, for example, parent solutions A (cost control rate 70%, quality compliance rate 85%) and B (cost control rate 65%, quality compliance rate 90%) are crossed to generate child solution C (68%, 88%);

[0401] Mutation: polynomial mutation with a mutation probability of 0.1, for example, the risk avoidance coefficient of a solution is mutated from 60% to 55%.

[0402] Non-dominated sorting and selection:

[0403] Non-dominated layering: divide the population solutions into multiple non-dominated layers, for example, the first layer contains 50 solutions and the second layer contains 80 solutions;

[0404] Reference point association: map the solutions to the reference points, select the solutions belonging to the reference points that are closest and not fully occupied to ensure diversity.

[0405] Example: An optimization generates 480 solutions, with the Pareto frontier containing:

[0406] Solution A: Cost control rate 82%, quality compliance rate 92%, risk avoidance coefficient 75%;

[0407] Solution B: Cost control rate 78%, quality compliance rate 95%, risk avoidance coefficient 80%.

[0408] These solutions are ensured to converge by the elite preservation strategy of NSGA-III (top 10% solutions are preserved each generation).

[0409] ‌The Pareto solution set is input into the digital twin engine, which simulates the construction process under different decision schemes through the physics engine, generating a multi-dimensional simulation dataset containing cost deviations, quality defects, and risk outbreak points.

[0410] The digital twin engine builds a virtual construction environment based on BIM (Building Information Modeling) and real-time sensor data, and the physics engine uses NVIDIA PhysX to achieve high-precision dynamic simulation.

[0411] Simulation scenario construction:

[0412] Model import: Convert the BIM model to USD (Universal Scene Description) format, containing 2000+ components such as structural members and mechanical and electrical pipelines.

[0413] Parameter binding: Map decision variables in the Pareto solution (such as material cost) to model attributes (such as concrete strength grade C30→C35).

[0414] Physical simulation process:

[0415] Time step: Set to 1 day, simulate a 365-day construction period.

[0416] Event triggering: When a node risk value in the risk probability matrix >0.7, trigger an emergency resource scheduling event.

[0417] Data collection: Record daily cost deviations (actual cost - budget), quality defects (such as wall flatness error >3mm), and risk outbreak points (such as supply chain interruption times).

[0418] Example: Simulation of solution A shows that on day 120, cost deviation +12% due to steel price increase, on day 200, concrete strength deficiency is found in a certain area (quality defect marker), on day 280, logistics is interrupted due to typhoon (risk outbreak).

[0419] Dataset generation:

[0420] Dimension definition: including time series cost curve (365 points), quality defect distribution heatmap (by floor partition), risk event spatio-temporal coordinates (latitude + longitude + time);

[0421] Data storage: Parquet columnar storage is adopted, with compression ratio up to 5:1, supporting fast query.

[0422] Based on the simulation dataset, the fuzzy comprehensive evaluation method is used to calculate the fitness score of each solution, and the gradient boosting decision tree is used to select the optimal cost control strategy.

[0423] The fuzzy comprehensive evaluation method quantifies the satisfaction of each target through the membership function, and calculates the comprehensive score combined with the weight.

[0424] Membership function design:

[0425] Cost control rate: trapezoidal function, ideal interval [80%, 100%], membership degree 0 when lower than 70%;

[0426] Quality compliance rate: Gaussian function, mean 95%, standard deviation 2%;

[0427] Risk aversion coefficient: S-shaped function, threshold 60% (membership degree drops sharply when lower than 60%).

[0428] Weight allocation:

[0429] Owner preference: cost control (weight 0.5), quality (0.3), risk (0.2);

[0430] Construction party preference: quality (0.5), risk (0.3), cost (0.2).

[0431] The most important weight is determined by AHP (analytic hierarchy process) as cost 0.4, quality 0.4, and risk 0.2.

[0432] Gradient boosting decision tree (GBDT) screening:

[0433] Feature engineering: input features include 15-dimensional indicators such as cost deviation mean, maximum quality defect level, and risk event frequency;

[0434] Model parameters: number of trees 100, learning rate 0.1, maximum depth 6;

[0435] Training and prediction: use 80% data for training and 20% for validation, output fitness score of each solution (0~100 points).

[0436] Example: solution A fitness score 92 (cost 88, quality 90, risk 85), solution B score 89 (cost 85, quality 95, risk 80), finally select solution A as the optimal strategy.

[0437] Additionally, based on the simulation dataset, the fitness score of each scheme is calculated using the fuzzy comprehensive evaluation method, and the optimal cost control strategy is selected through gradient boosting decision trees, which can include:

[0438] According to the multi-dimensional engineering parameters in the simulation dataset, the adaptive weight distribution of the cost deviation rate, the quality defect index, and the risk outbreak probability is realized through the dynamic weight generator based on Monte Carlo sampling, and the dynamic evaluation weight vector with confidence interval is generated;

[0439] Based on the engineering parameter dataset generated by digital twin simulation (containing 12-dimensional parameters such as cost deviation rate, quality defect index, and risk outbreak probability), a dynamic weight generator is constructed. First, the Monte Carlo method is used to randomly sample the priority of the three targets: set a weight combination generated by each sampling (such as cost deviation rate weight w1=0.4, quality defect index w2=0.3, and risk outbreak probability w3=0.3), and perform 1000 iterations of sampling. In each iteration, the fitness score of the current weight combination is calculated according to the success criteria in the historical engineering case library (such as cost deviation rate ≤5%, quality defect index ≤0.2, and risk outbreak probability ≤10%), and the effective weight interval that meets the constraint condition is selected. For example, through statistical distribution, it is found that when w1∈[0.35,0.45], w2∈[0.25,0.35], and w3∈[0.2,0.3], the fitness score confidence interval (95% confidence level) is [82.3,89.5]. Finally, the dynamic evaluation weight vector (w1=0.42±0.03, w2=0.28±0.02, w3=0.30±0.02) is generated, and the confidence labels of each weight (such as the confidence of the cost deviation rate weight 0.92) are marked.

[0440] Based on the dynamic evaluation weight vector, the fuzzy comprehensive evaluation algorithm is used to calculate the fitness score of the scheme, which introduces a time decay factor to correct the contribution of historical simulation data, and maps high-dimensional nonlinear patterns through a radial basis function kernel, outputting a fitness score matrix with timeliness labels;

[0441] ‌1. Time decay factor and fuzzy comprehensive evaluation‌

[0442] In calculating the fitness score, a time decay factor (decay factor λ = 0.85) is designed to dynamically adjust the contribution of historical data according to the time series characteristics of the simulation data set. For example, the simulation data weight at the current time t is 1.0, while the data weight at t-1 is λ = 0.85, and at t-2 is λ² = 0.72, and so on, to ensure that the latest data has a greater impact on the evaluation results. At the same time, fuzzy membership functions (such as the membership function of the cost deviation rate defined as a "low deviation" trapezoidal function with threshold interval [0, 5%]) are used to fuzzify the three objectives. Combine the dynamic weight vector with the fuzzy membership degree to calculate the fitness score of each scheme. For example, a certain scheme has a cost deviation rate of 3.2% (membership degree 0.86), a quality defect index of 0.15 (membership degree 0.92), and a risk outbreak probability of 8% (membership degree 0.78), and its comprehensive score is 0.42 x 0.86 + 0.28 x 0.92 + 0.30 x 0.78 = 0.847.

[0443] 2. Radial basis function kernel mapping and timeliness label

[0444] To handle high-dimensional nonlinear data (such as the complex relationship between construction process parameters and cost deviation), a radial basis function (RBF) kernel (parameter gamma = 0.1) is used to map the 12-dimensional engineering parameters. Through the kernel function, the original data is projected into a high-dimensional space to extract nonlinear relationship features (such as the U-shaped association between concrete strength and quality defects). In the final generated fitness score matrix, each scheme is labeled with a timeliness label (such as "70% of the data in the past week"), which is used for subsequent decision screening.

[0445] According to the fitness score matrix, a multi-modal feature fusion channel is constructed to extract the hyperplane curvature feature of the Pareto solution set, the decision variable correlation degree feature, and the simulation process stability feature, and a unified dimension of enhanced feature vector set is generated through a heterogeneous feature embedding network;

[0446] ‌1. Multi-modal feature extraction and fusion‌

[0447] ‌Hyperplane curvature feature: For the Pareto solution set (300-500 solutions) generated by the NSGA-III algorithm, principal component analysis (PCA) is used to reduce the dimension to 3-dimensional space, and the curvature radius of the solution set frontier is calculated (such as the average curvature radius R = 12.5). The greater the curvature, the higher the diversity of the solution set, and the more complex the decision space.

[0448] ‌Correlation degree feature: Analyze the correlation between cost control strategy parameters (e.g., material procurement cycle, construction team quantity) and the three objectives using Pearson correlation coefficient. For example, it is found that the procurement cycle has a strong positive correlation with the cost deviation rate (r=0.62) and a weak negative correlation with the quality defect rate (r=-0.33).

[0449] ‌Simulation process stability feature: Calculate the variance of the results of 10 repeated experiments of the same strategy in digital twin simulation (e.g., cost deviation variance σ²=0.8). The smaller the variance, the stronger the robustness of the strategy.

[0450] ‌2. Heterogeneous feature embedding network

[0451] Construct a heterogeneous network containing convolutional layers (extracting spatial features), LSTM layers (extracting temporal features), and fully connected layers (fusing features). Input the curvature feature (dimension 3), correlation degree feature (dimension 6), and stability feature (dimension 1) into the network, and generate a unified 128-dimensional enhanced feature vector through weight sharing mechanisms (e.g., convolution kernel size=3×3, stride=1). For example, the feature vector of a certain strategy is represented as [curvature=0.12, procurement cycle correlation=0.62, stability=0.92,...].

[0452] Input the enhanced feature vector set into the gradient boosting decision tree model, use multi-objective splitting criteria to optimize the tree structure growth direction, and use cost-sensitive learning strategies to enhance the search ability of the high-dimensional non-convex solution space, output the global ranking result of the optimal cost control strategy;

[0453] 1. Gradient boosting decision tree (GBDT) modeling

[0454] Set the GBDT model parameters: maximum tree depth 5 (to prevent overfitting), learning rate lr=0.05 (to balance training speed and accuracy), and iteration number n_estimators=200. Use multi-objective splitting criteria to optimize the weighted information gain of the three objectives (cost, quality, and risk) at each node split. For example, when a node is split according to "procurement cycle ≤7 days", the information gain of the left child is cost objective gain 0.15 + quality objective gain 0.08 + risk objective gain 0.06 =0.29, and the right child is 0.21, so the left child is selected as the splitting direction.

[0455] 2. Cost-sensitive learning and global ranking

[0456] Introduce a cost-sensitive learning strategy, impose higher penalty weights (penalty = 3.0) on misclassified samples with high-cost bias (e.g., bias rate > 10%), and set the penalty weight to 1.0 for low-risk samples (risk outbreak probability < 5%). After model training, output the global ranking results of the strategy (e.g., strategy A score 92.3, strategy B score 88.7), and label the pros and cons of each strategy (e.g., "cost control priority" for strategy A, "risk aversion priority" for strategy B).

[0457] According to the global ranking results, combined with the physical constraint verification of the digital twin engine, the strategy parameters are iteratively optimized through the backpropagation correction mechanism to generate the optimal cost control strategy execution sequence that meets the construction feasibility boundary.

[0458] ‌1. Physical constraint verification and backpropagation correction‌

[0459] Input the top 10 strategies into the digital twin engine, simulate the construction process based on the physical engine (e.g., NVIDIA PhysX), and verify whether the strategy meets the actual constraints (e.g., tower crane load capacity ≤ 50 tons, single-day concrete pouring volume ≤ 300 m³). For example, a strategy is marked as unfeasible due to tower crane overload (simulation shows that 55 tons need to be lifted). Through the backpropagation mechanism, the constraint violation signal is fed back to the GBDT model, the feature weight is adjusted (e.g., increase the "tower crane usage rate" weight by 20%), and the strategy ranking is regenerated.

[0460] 2.‌Execution sequence generation and visualization‌

[0461] Finally, 3 groups of feasible strategies (e.g., strategy A, strategy D, strategy G) are selected, and the execution sequence is generated according to the priority:

[0462] First stage (1-30 days): execute strategy A (cost bias rate controlled at 3.5%);

[0463] Second stage (31-60 days): switch to strategy D (respond to unexpected risk events);

[0464] Third stage (61-90 days): enable strategy G (ensure completion acceptance quality).

[0465] In the BIM collaboration platform, render three-dimensional deduction animations through Unreal Engine to show the cost curve changes of each stage strategy, the risk heat map diffusion process, and the trigger logic of key decision nodes.

[0466] ‌Fuse the optimal strategy with the BIM model, render a three-dimensional risk heat map through ray tracing technology, and generate a decision path deduction animation to output a visual decision support interface.

[0467] The visualization system is developed based on Unity Engine, integrating BIM models and real-time data streams for dynamic rendering and interaction.

[0468] Three-dimensional risk heat map rendering:

[0469] Data mapping: risk values (0~1) in the risk probability matrix are mapped to a color gradient (green→yellow→red);

[0470] Ray tracing: real-time calculation of light reflection using NVIDIA RTX 6000 GPU, such as high-risk areas (>0.8) displayed in deep red with a flashing special effect;

[0471] Interactive functions: clicking on the heat map area can pop up detailed risk analysis (such as "underground second floor support structure risk value 0.78").

[0472] Decision path deduction animation:

[0473] Timeline control: animation length 5 minutes, compressed to show 365 days of construction process;

[0474] Key event annotation: such as when the cost exceeds the budget at 120 days, the screen focuses on the procurement module and pops up optimization suggestions ("enable backup suppliers");

[0475] Multi-view switching: supports global overhead view, first-person inspection, component section view, etc.

[0476] Example: in the deduction animation, the user can observe the following process:

[0477] 0-100 days: mainly green heat, stable cost control;

[0478] 120 days: some areas turn yellow, indicating the impact of steel price rise;

[0479] 200 days: local red, showing concrete defects, system automatically triggers quality rectification instructions;

[0480] 365 days: overall return to green, indicating project successful acceptance.

[0481] Output interface:

[0482] Main panel: left side is three-dimensional heat map, right side is fitness score ranking;

[0483] Console: supports parameter adjustment (such as modifying weights and re-running optimization);

[0484] Report export: generates PDF / Excel format decision analysis report.

[0485] The NSGA-III algorithm is used to balance multiple objectives such as cost control rate, quality compliance rate, and risk aversion coefficient, generating hundreds of candidate solutions. A digital twin engine simulates the performance of each solution in a virtual construction environment (e.g., cost deviation, quality defects), ultimately selecting the optimal strategy. A visual interface uses a 3D heatmap to locate high-risk areas and animations to demonstrate the effects of decision implementation, providing a globally optimal solution under complex constraints and avoiding the local optimum trap caused by single-objective optimization. Simulation verification and visualization enhance the credibility of the decision, helping managers intuitively understand the impact of the strategy.

[0486] As can be seen, a spatiotemporally correlated structured cost data cube is generated based on heterogeneous cost data from each stage of engineering planning, design, construction, and completion. Based on this structured cost data cube, a dynamic cost prediction curve and deviation-sensitive nodes are output. According to the dynamic cost prediction curve, a blockchain-enabled BIM / CIM collaborative platform is used to optimize multi-party task allocation and generate a collaborative instruction set encoded with smart contracts. Based on the collaborative instruction set and real-time engineering data stream, a risk probability matrix and early warning signals are output. Based on the risk probability matrix, a multi-objective optimization algorithm is used to generate an anti-interference decision scheme set. After verification through digital twin simulation, the optimal cost control strategy is output. This enables efficient integration of cost data from each stage of the project, achieving dynamic cost prediction, optimized collaborative decision-making, and effective risk management.

[0487] Another embodiment of the present invention provides a multi-dimensional intelligent management system for the entire process cost estimation, see [link to relevant documentation]. Figure 3 The system may include:

[0488] Alignment module 301 is used to perform cross-stage feature alignment based on heterogeneous cost data from each stage of engineering planning, design, construction, and completion, and generate a spatiotemporally correlated structured cost data cube through a federated learning-driven data fusion engine. The data fusion engine uses a differential privacy protection mechanism to eliminate data silos.

[0489] The construction module 302 is used to construct a spatiotemporal graph neural network prediction model based on the structured cost data cube, combined with market price fluctuation trends and construction progress change factors, and output dynamic cost prediction curves and deviation-sensitive nodes. The spatiotemporal graph neural network prediction model captures the correlation effects of potential cost overrun risk factors through an attention mechanism.

[0490] The optimization module 303 is used to optimize the multi-party task allocation based on the dynamic cost prediction curve using a blockchain-enabled BIM / CIM collaborative platform, and generate a collaborative instruction set encoded by smart contracts. The BIM / CIM collaborative platform achieves privacy-protected data synchronization among owners, designers, and contractors through zero-knowledge proofs.

[0491] The analysis module 304 is configured to analyze potential risk paths based on the collaborative instruction set and real-time engineering data flow through a dynamic knowledge graph engine, output a risk probability matrix and an early warning signal, wherein the dynamic knowledge graph engine fuses an industry knowledge base and a reinforcement learning strategy to realize risk evolution simulation.

[0492] The output module 305 is configured to generate an anti-interference decision scheme set by using a multi-objective optimization algorithm according to the risk probability matrix, and output an optimal cost control strategy after digital twin simulation verification, wherein the optimal cost control strategy synchronously drives a three-dimensional risk heat map and a decision path deduction animation in a visual interface.

[0493] It can be seen that the time-space related structured cost data cube is generated according to the heterogeneous cost data of the engineering planning, design, construction and completion stages; the dynamic cost prediction curve and the deviation sensitive node are output based on the structured cost data cube; the multi-party task allocation optimization is performed by using the BIM / CIM collaborative platform enabled by the blockchain according to the dynamic cost prediction curve, the collaborative instruction set encoded by the smart contract is generated; the risk probability matrix and the early warning signal are output based on the collaborative instruction set and the real-time engineering data flow; the anti-interference decision scheme set is generated by using the multi-objective optimization algorithm according to the risk probability matrix, and the optimal cost control strategy is output after the digital twin simulation verification, so that the cost data of each stage of the engineering can be efficiently integrated, the dynamic cost prediction, the optimized collaborative decision and the effective risk management can be realized.

[0494] The above describes the structure, features and effects of the present application in detail according to the embodiments shown in the drawings. The above description is only the preferred embodiments of the present application, but the present application is not limited to the embodiments shown in the drawings. Any changes or modifications made according to the concept of the present application, or equivalent embodiments with equivalent changes, are still within the scope of the present application.

Claims

1. A multi-dimensional intelligent management method for the entire process of cost estimation, characterized in that, The method includes: Based on the heterogeneous cost data of each stage of engineering planning, design, construction and completion, a data fusion engine driven by federated learning is used to perform cross-stage feature alignment to generate a spatiotemporally correlated structured cost data cube. The data fusion engine adopts a differential privacy protection mechanism to eliminate data silos. Based on the structured cost data cube, and combined with market price fluctuation trends and construction progress change factors, a spatiotemporal graph neural network prediction model is constructed to output a dynamic cost prediction curve and deviation-sensitive nodes. The spatiotemporal graph neural network prediction model captures the correlation effects of potential cost overrun risk factors through an attention mechanism. Based on the dynamic cost prediction curve, the blockchain-enabled BIM / CIM collaborative platform is used to optimize the allocation of tasks among multiple parties and generate a collaborative instruction set coded by smart contracts. The BIM / CIM collaborative platform achieves privacy-preserving data synchronization among owners, designers, and contractors through zero-knowledge proofs. Based on the collaborative instruction set and real-time engineering data stream, a dynamic knowledge graph engine is used to analyze potential risk paths and output a risk probability matrix and early warning signals. The dynamic knowledge graph engine integrates industry knowledge base and reinforcement learning strategies to simulate risk evolution. Based on the risk probability matrix, a set of anti-interference decision schemes is generated using a multi-objective optimization algorithm. After verification by digital twin simulation, the optimal cost control strategy is output. The optimal cost control strategy synchronously drives the three-dimensional risk heat map and decision path deduction animation in the visualization interface. The process of obtaining the structured cost data cube includes: Based on the BIM model parameters in the engineering planning stage, the bill of quantities in the design stage, the progress log in the construction stage, and the settlement documents in the completion stage, a heterogeneous data adapter is used to perform semantic mapping of fields to generate a four-dimensional tensor containing timestamps, spatial coordinates, and cost attributes, and output a standardized feature vector set. Based on the feature vector set, a cross-stage alignment algorithm under the federated learning framework is adopted. The local model parameters of each participant are optimized by projective gradient descent, and the distributed feature space alignment is completed under differential privacy constraints to generate a cross-stage joint feature matrix. The joint feature matrix is ​​input into the spatiotemporal correlation engine, which captures the coupling relationship between project progress, cost consumption and geographical distribution through three-dimensional convolution kernels. A six-dimensional cost data cube containing time series correlation, spatial topological correlation and cost transmission correlation is constructed, and a structured data cube with privacy protection label is output. Construction schedule change factors include the number of delay days for each node and the process dependencies.

2. The method according to claim 1, characterized in that, Based on the structured cost data cube, and combined with market price fluctuation trends and construction schedule change factors, a spatiotemporal neural network prediction model is constructed to output a dynamic cost prediction curve and deviation-sensitive nodes. The spatiotemporal neural network prediction model captures the correlation effects of potential cost overrun risk factors through an attention mechanism, including: Based on the time-series cost data in the structured data cube, cost fluctuation features are extracted through a multi-scale sliding window, and a dynamic embedding vector with time dimension is constructed by combining it with the external market price index. Based on the construction schedule change factor, the spatial dependency relationship between construction nodes is modeled through a graph attention mechanism to generate a spatial topology weight matrix; The temporal dynamic embedding vector and spatial topological weight matrix are input into the spatiotemporal graph convolutional layer. The spatiotemporal features are fused through bidirectional gated recurrent units to output a distribution map of the correlation influence intensity of potential overspending risk factors. Based on the risk impact intensity distribution map, a dynamic threshold segmentation algorithm is used to identify deviation-sensitive nodes and generate dynamic cost prediction curves with confidence level labels.

3. The method according to claim 2, characterized in that, The process involves optimizing multi-party task allocation using a blockchain-enabled BIM / CIM collaborative platform based on the dynamic cost prediction curve, generating a collaborative instruction set coded with smart contracts. The BIM / CIM collaborative platform achieves privacy-preserving data synchronization among owners, designers, and contractors through zero-knowledge proofs, including: Based on the dynamic cost prediction curve, the owner's budget constraint privacy data, the designer's drawing change privacy data, and the construction party's resource scheduling privacy data are homomorphically encrypted using a zero-knowledge proof protocol to generate a verifiable ciphertext dataset. Based on the encrypted dataset, a multi-party secure computation is used to decompose the collaborative task. The Hungarian algorithm is used to solve the optimal task allocation scheme under the balance of interests of multiple parties, and the task allocation matrix with weight constraints is output. The task allocation matrix is ​​encoded into a smart contract, and the contract triggering conditions are defined through an event-driven state machine to generate a smart contract template containing an execution logic chain. Deploy lightweight blockchain nodes on the BIM / CIM collaboration platform to achieve privacy-preserving data synchronization across participants through a sharding consensus mechanism, and generate a timestamped collaborative instruction hash chain; Based on real-time engineering data streams, the authenticity of off-chain data is verified through oracles, the execution status of smart contracts is dynamically updated, and a traceable set of collaborative instructions is output.

4. The method according to claim 3, characterized in that, Based on the collaborative instruction set and real-time engineering data stream, a dynamic knowledge graph engine analyzes potential risk paths and outputs a risk probability matrix and early warning signals. The dynamic knowledge graph engine integrates an industry knowledge base and reinforcement learning strategies to simulate risk evolution, including: Based on the collaborative instruction set, risk semantic fragments from engineering contract terms and construction specification documents are extracted through a domain adaptive pre-trained model to generate risk knowledge triples with probability annotations. Based on risk knowledge triples and combined with abnormal events in real-time engineering data streams, a risk propagation path topology graph is constructed using a temporal graph convolutional network. By inputting the risk propagation path into a reinforcement learning environment, the risk evolution process is simulated through a dual-deep Q-network, generating a risk probability transition matrix that includes the intensity of cascading effects. Based on the risk probability transition matrix, a backpropagation importance sampling algorithm is used to identify key risk paths and output an early warning signal with time window constraints. By compressing risk model parameters through knowledge distillation technology and deploying them to edge computing nodes, real-time risk monitoring is achieved, generating a lightweight risk evolution simulator.

5. The method according to claim 4, characterized in that, The process involves generating an anti-interference decision scheme set using a multi-objective optimization algorithm based on the risk probability matrix, verifying it through digital twin simulation, and then outputting the optimal cost control strategy. This optimal cost control strategy synchronously drives the 3D risk heatmap and decision path deduction animation in the visualization interface, including: Based on the risk probability matrix, the three objectives of cost control rate, engineering quality compliance rate and risk aversion coefficient are simultaneously optimized using the NSGA-III algorithm to generate an initial Pareto front containing 300-500 sets of solutions. Input the Pareto solution set into the digital twin engine, and simulate the construction process under different decision schemes through the physics engine to generate a multi-dimensional simulation dataset containing cost deviations, quality defects, and risk outbreak points. Based on the simulation dataset, the fitness score of each scheme is calculated using the fuzzy comprehensive evaluation method, and the optimal cost control strategy is selected by gradient boosting decision tree. By integrating the optimal strategy with the BIM model, a 3D risk heat map is rendered using ray tracing technology, and a decision path deduction animation is generated, outputting a visual decision support interface.

6. The method according to claim 2, characterized in that, The method based on construction schedule change factors, using a graph attention mechanism to model the spatial dependencies between construction nodes, and generating a spatial topology weight matrix includes: Based on the node delay data in the construction progress change factor, the spatiotemporal correlation features between construction nodes are extracted through the spatiotemporal graph convolutional layer to generate an initial dependency graph with timestamps. Based on the initial dependency graph and combined with the geographic coordinate data in the BIM model, a multimodal attention mechanism is used to calculate the spatial correlation weights between nodes. The multimodal attention score matrix is ​​output by integrating three channel features: physical distance between nodes, construction technology dependence, and resource flow direction. The multimodal attention score matrix is ​​input into the dynamic residual connection network, and implicit dependencies in historical construction logs are fused through gated recurrent units to generate a spatiotemporally enhanced node dependency strength distribution map. Based on the dependency strength distribution map, a differential privacy-preserving graph pruning algorithm is used to remove weak dependency edges with confidence levels below a threshold. Then, a hierarchical spatial topology weight matrix is ​​generated through spectral clustering, and the spatial topology of construction nodes with security protection labels is output.

7. The method according to claim 4, characterized in that, The method, based on a encrypted dataset, employs multi-party secure computation to decompose collaborative tasks, and uses the Hungarian algorithm to solve for the optimal task allocation scheme under the balance of interests among multiple parties, outputting a task allocation matrix with weight constraints, including: Based on the homomorphically encrypted ciphertext dataset, the collaborative task is decomposed through a secure multi-party computation protocol to generate a privacy-preserving task requirement vector, in which the task parameters are divided into multiple shadow shares using secret sharing technology. Based on the task requirement vector, a multi-party interest game model is constructed, and the initial task allocation scheme is calculated through the Nash equilibrium solution algorithm. The model constraints include the owner's budget limit, the contractor's resource capacity, and the designer's tolerance for changes. The initial allocation scheme is input into the Hungarian algorithm, and a dynamic weight adjustment strategy is introduced to optimize the matching of conflicting tasks. The dynamic weights are adaptively updated according to the real-time supply chain risk index, and the intermediate task allocation matrix with elastic constraints is output. The fairness of the intermediate task allocation matrix is ​​verified by zero-knowledge proof. The zk-SNARK protocol is used to generate verifiable allocation proofs. Finally, a weighted and constrained task allocation matrix that is immutable and satisfies the balance of interests among multiple parties is output.

8. The method according to claim 5, characterized in that, The process, based on a simulation dataset, employs fuzzy comprehensive evaluation to calculate the fitness scores of each scheme, and uses a gradient boosting decision tree to select the optimal cost control strategy, including: Based on the multidimensional engineering parameters in the simulation dataset, an adaptive weight allocation is performed on the three objectives of cost deviation rate, quality defect index and risk outbreak probability through a dynamic weight generator based on Monte Carlo sampling, generating a dynamic evaluation weight vector with confidence intervals. Based on the dynamic evaluation weight vector, the fitness score of the scheme is calculated using the fuzzy comprehensive evaluation algorithm. The fuzzy comprehensive evaluation algorithm introduces a time decay factor to correct the contribution of historical simulation data and maps a high-dimensional nonlinear pattern through a radial basis function kernel to output a fitness score matrix with time-sensitive labels. Based on the fitness score matrix, a multimodal feature fusion channel is constructed to extract the hyperplane curvature features, decision variable correlation features, and simulation process stability features of the Pareto solution set. A unified-dimensional enhanced feature vector set is generated through a heterogeneous feature embedding network. The enhanced feature vector set is input into the gradient boosting decision tree model. The tree structure growth direction is optimized by adopting a multi-objective splitting criterion. The search capability of the high-dimensional non-convex solution space is enhanced by a cost-sensitive learning strategy. The global ranking result of the optimal cost control strategy is output. Based on the global ranking results and the physical constraint verification of the digital twin engine, the strategy parameters are iteratively optimized through a backpropagation correction mechanism to generate the optimal cost control strategy execution sequence that satisfies the construction feasibility boundary.

9. A multi-dimensional intelligent management system for full-process cost estimation, characterized in that, The system includes: The alignment module is used to perform cross-stage feature alignment based on heterogeneous cost data from each stage of engineering planning, design, construction, and completion, and generate a spatiotemporally correlated structured cost data cube through a federated learning-driven data fusion engine. The data fusion engine uses a differential privacy protection mechanism to eliminate data silos. The construction module is used to build a spatiotemporal graph neural network prediction model based on the structured cost data cube, combined with market price fluctuation trends and construction progress change factors, and output dynamic cost prediction curves and deviation-sensitive nodes. The spatiotemporal graph neural network prediction model captures the correlation effects of potential cost overrun risk factors through an attention mechanism. The optimization module is used to optimize the allocation of tasks among multiple parties using a blockchain-enabled BIM / CIM collaborative platform based on the dynamic cost prediction curve, and generate a collaborative instruction set encoded by smart contracts. The BIM / CIM collaborative platform achieves privacy-preserving data synchronization among owners, designers, and contractors through zero-knowledge proofs. The parsing module is used to analyze potential risk paths based on the collaborative instruction set and real-time engineering data stream through a dynamic knowledge graph engine, and output a risk probability matrix and early warning signals. The dynamic knowledge graph engine integrates industry knowledge base and reinforcement learning strategy to realize risk evolution simulation. The output module is used to generate an anti-interference decision scheme set based on the risk probability matrix using a multi-objective optimization algorithm, and output the optimal cost control strategy after verification by digital twin simulation. The optimal cost control strategy synchronously drives the three-dimensional risk heat map and decision path deduction animation in the visualization interface. The process of obtaining the structured cost data cube includes: Based on the BIM model parameters in the engineering planning stage, the bill of quantities in the design stage, the progress log in the construction stage, and the settlement documents in the completion stage, a heterogeneous data adapter is used to perform semantic mapping of fields to generate a four-dimensional tensor containing timestamps, spatial coordinates, and cost attributes, and output a standardized feature vector set. Based on the feature vector set, a cross-stage alignment algorithm under the federated learning framework is adopted. The local model parameters of each participant are optimized by projective gradient descent, and the distributed feature space alignment is completed under differential privacy constraints to generate a cross-stage joint feature matrix. The joint feature matrix is ​​input into the spatiotemporal correlation engine, which captures the coupling relationship between project progress, cost consumption and geographical distribution through three-dimensional convolution kernels. A six-dimensional cost data cube containing time series correlation, spatial topological correlation and cost transmission correlation is constructed, and a structured data cube with privacy protection label is output. Construction schedule change factors include the number of delay days for each node and the process dependencies.

Images