Construction cost analysis system and method based on big data

Abstract

The invention relates to the field of engineering construction, in particular to a construction cost analysis system and method based on big data. The method comprises the following steps: acquiring multi-modal construction data in a system, sequentially performing data preprocessing and standardization on the multi-modal construction data, aligning texts and numerical values in the initial multi-modal construction data, constructing a knowledge graph by using aligned text numerical value data and image spatio-temporal data, and connecting an LSTM-GRU hybrid network with an HMM hidden Markov model. The method comprises the following steps: establishing a construction cost prediction model, adding a space-time attention mechanism in the model, inputting feature multi-modal construction data into the model for prediction, iterating model training through a PPO algorithm, and carrying out unified scheduling management on materials and manpower in a construction project based on a minimum cost path and a risk plan library. Key space-time regions can be dynamically focused, so that the spatial resolution of a prediction result is improved to a section level, and a basis is provided for accurate scheduling.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network security, in particular to a construction cost analysis system and method based on big data. BACKGROUND

[0002] When the traditional construction cost analysis relies on manual experience and static historical data, the following defects exist: single data dimension: only based on the cost records within the enterprise, lacking real-time integration of market dynamics and environmental factors. In addition, related external data such as project peripheral supporting facility construction cost and temporary site rental price cannot be effectively collected, resulting in that the cost analysis is difficult to fully cover the complex economic environment where the project is actually located. Significant lag: long cost model update cycle, unable to adapt to sudden risks. After the risk occurs, it often takes a long time from information collection, arrangement to inclusion in the cost analysis system, making the cost adjustment measures seriously lagging behind, which may cause adverse consequences such as project fund chain tension or even rupture. SUMMARY

[0003] The present application is designed to solve the above problems, and a construction cost analysis system and method based on big data are designed.

[0004] To achieve the above purpose, the technical scheme of the present application is as follows: further, in the above construction cost analysis system based on big data, the construction cost analysis system comprises the following modules:

[0005] A construction data acquisition module is used to acquire multi-modal construction data in the system, and the multi-modal construction data is sequentially subjected to data preprocessing and standardization to obtain initial multi-modal construction data;

[0006] A feature data alignment module is used to align the text and numerical value in the initial multi-modal construction data, and align the image and space-time, and construct a knowledge graph for the aligned text and numerical value data and image and space-time data to obtain feature multi-modal construction data;

[0007] A prediction model construction module is used to connect an LSTM-GRU hybrid network and an HMM hidden Markov model to establish an LSTM-GRU-HMM construction cost prediction model, and add a space-time attention mechanism in the model to obtain a target LSTM-GRU-HMM construction cost prediction model;

[0008] A cost risk prediction module is used to input the feature multi-modal construction data into the target LSTM-GRU-HMM construction cost prediction model for prediction, and through the training of the PPO algorithm iteration model, a Pareto optimal solution set is generated to obtain a minimum cost path and a risk plan library;

[0009] The human resource and material scheduling module is configured to schedule and manage materials and human resources in the construction project based on the minimum cost path and the risk plan library.

[0010] Further, in the construction cost analysis system based on big data, the construction data acquisition module comprises the following sub-modules:

[0011] The data acquisition sub-module is configured to access enterprise ERP systems, BIM models and public databases through API interfaces to acquire material inventory data, bill of quantities data and building material price indexes.

[0012] The image scanning sub-module is configured to acquire construction log image in an OCR scanner and extract text data of the construction log image.

[0013] The sensor sub-module is configured to collect site environment data and human resource data in a construction site by using temperature and humidity sensors and vibration sensors.

[0014] The standardization sub-module is configured to perform outlier deletion, missing value supplementation and standardization processing on the material inventory data, bill of quantities data, building material price indexes, construction log image, text data, site environment data and human resource data to obtain initial multi-modal construction data.

[0015] Further, in the construction cost analysis system based on big data, the feature data alignment module comprises the following sub-modules:

[0016] The extraction sub-module is configured to extract text name data in the initial multi-modal construction data by using an NLP natural language processing model, at least including material name, construction period node and risk event, associate the text name data with numerical value labels to establish a mapping relationship to obtain first associated data.

[0017] The positioning sub-module is configured to position an image region in the initial multi-modal construction data based on a YOLO target detection model, label quality defect labels, acquire text data in the quality defect labels to obtain image data.

[0018] The matching sub-module is configured to match date and location information in the image data with 3D coordinates of a BIM model, fuse sensor data and environment states in the image to generate spatio-temporal joint features to obtain second associated data.

[0019] The establishment sub-module is configured to establish static rule library and dynamic rule library of the first associated data and the second associated data, store heterogeneous data association by using a Neo4j graph database, encode multi-hop relationships in a knowledge graph into feature vectors to obtain feature multi-modal construction data.

[0020] Further, in the construction cost analysis system based on big data, the prediction model construction module comprises the following sub-modules:

[0021] A connection sub-module is configured to connect the LSTM-GRU hybrid network and the HMM hidden Markov model to obtain a feature vector of historical construction data in the system ;

[0022] An extraction sub-module is configured to extract long-term dependence features of the feature vector using an LSTM layer to output a first hidden state and learn short-term fluctuation patterns of the feature vector using a GRU layer to output a second hidden state;

[0023] A splicing sub-module is configured to splice the first hidden state and the second hidden state into a hybrid feature to obtain initial construction feature data;

[0024] The market policy stage in the historical construction data is defined as a hidden state of the HMM hidden Markov model , and the state transition matrix of the model is as follows:

[0025]

[0026] wherein, represents a probability of transition from a state to a state , and represents a hidden state at a time point ;

[0027] A generation sub-module is configured to generate an observation probability distribution using the initial construction feature data to obtain an observation probability of the HMM hidden Markov model:

[0028]

[0029]

[0030] wherein, represents an observation vector, represents a state generates an observation vector , and represents initial construction feature data, represents an activation function, represents a weight matrix, represents a bias;

[0031] A calculation sub-module is configured to solve an optimal state sequence in the model to obtain macro construction state label data. ​​​

[0032] Further, in the above construction cost analysis system based on big data, the prediction model construction module comprises the following units:

[0033] An attention mechanism adding unit is configured to add a spatio-temporal attention mechanism in the LSTM-GRU-HMM construction cost prediction model, and dynamically allocate cost weights of different regions and construction stages in the macro construction state label data by using the spatio-temporal attention mechanism.

[0034] A prediction weight unit is configured to calculate importance weights of historical time steps in the macro construction state label data for current prediction by using a time attention mechanism .

[0035]

[0036]

[0037] wherein, represents a scoring function output in the attention mechanism, represents a weight matrix in the time attention mechanism, which performs linear transformation on the spliced features , represents macro construction state label data, represents features at the current time, represents a bias term in the time attention mechanism;

[0038] A weight calculation unit is configured to calculate spatial weights according to region features in the macro construction state label data by using a spatial attention mechanism .

[0039] A model obtaining unit is configured to combine the attention weights and the mixed features to generate a final prediction input, and obtain a target LSTM-GRU-HMM construction cost prediction model after training.

[0040] Further, in the above construction cost analysis system based on big data, the cost risk prediction module comprises the following units:

[0041] The feature multi-modal construction data is input into the target LSTM-GRU-HMM construction cost prediction model for prediction.

[0042] A PPO algorithm is used to iteratively train the model to generate a Pareto optimal solution set, and the optimization of the PPO algorithm at least includes a policy network configured to output an action probability distribution ; and a value network configured to evaluate state values .

[0043] The importance sampling ratio is:

[0044]

[0045] The clipping objective function is:

[0046]

[0047] wherein, represents the advantage function, represents the immediate reward, represents the discount factor, represents the clipping threshold, represents the action probability distribution under the old policy, represents the action, represents the state, represents the expectation, represents the current policy, represents the old policy.

[0048] The Pareto dominance relationship is defined, a diversity solution set is generated by adjusting weights, a non-dominated solution is screened to constitute a Pareto front, and a minimum cost path and a risk plan library are obtained.

[0049] Further, in the construction cost analysis system based on big data, the human and material resource scheduling module comprises the following units:

[0050] A task disassembly unit is configured to disassemble construction tasks into material procurement, human resource allocation and equipment scheduling sub-tasks according to key nodes in the minimum cost path, generate a space-time constraint graph based on a BIM model, and identify resource conflict areas;

[0051] A resource allocation unit is configured to mark tasks with a cost weight > 15% in the path as critical path tasks, and allocate resources preferentially, and reserve 20% resource redundancy for high-risk tasks in the plan library;

[0052] A risk judgment unit is configured to match a three-level response scheme from the plan library according to the type and level of the risk event, at least including a first-level response, a second-level response and a third-level response, the first-level response is used to directly start a preset action, the second-level response is used to generate a multi-scheme selection, and the third-level response is used to trigger an artificial intervention process.

[0053] Further, in the construction cost analysis method based on big data, the construction cost analysis method comprises the following steps:

[0054] Obtaining multi-modal construction data in the system, sequentially performing data preprocessing and standardization on the multi-modal construction data to obtain initial multi-modal construction data;

[0055] aligning texts and numerical values in the initial multi-modal construction data, and aligning images and space-time, constructing a knowledge graph from the aligned text and numerical value data and the image and space-time data, and obtaining feature multi-modal construction data;

[0056] The LSTM-GRU-HMM construction cost prediction model is connected with an LSTM-GRU hybrid network and an HMM hidden Markov model, a space-time attention mechanism is added to the model, and a target LSTM-GRU-HMM construction cost prediction model is obtained;

[0057] The feature multi-modal construction data is input into the target LSTM-GRU-HMM construction cost prediction model for prediction, a PPO algorithm is used to iteratively train the model, a Pareto optimal solution set is generated, a minimum cost path and a risk plan library are obtained;

[0058] Based on the minimum cost path and the risk plan library, the materials and manpower in the construction project are uniformly scheduled and managed.

[0059] Further, in the above construction cost analysis method based on big data, the multi-modal construction data in the acquisition system is sequentially subjected to data preprocessing and standardization to obtain initial multi-modal construction data, including:

[0060] The material inventory data, the bill of quantities data, and the building material price index are obtained by accessing the enterprise ERP system, the BIM model, and the public database through the API interface;

[0061] The construction log image in the OCR scanner is obtained, and the text data of the construction log image is extracted;

[0062] The site environment data and manpower data in the construction site are collected by using the temperature and humidity sensor and the vibration sensor;

[0063] The material inventory data, the bill of quantities data, the building material price index, the construction log image, the text data, the site environment data, and the manpower data are subjected to outlier deletion, missing value supplementation, and standardization processing to obtain initial multi-modal construction data.

[0064] Further, in the above construction cost analysis method based on big data, the multi-modal construction data in the acquisition system is sequentially subjected to data preprocessing and standardization to obtain initial multi-modal construction data, including:

[0065] extract text name data in the initial multi-modal construction data using an NLP natural language processing model, at least including material name, construction period node and risk event; correlate the text name data and the numerical label to establish a mapping relationship to obtain first correlation data;

[0066] Based on the YOLO target detection model, the image region in the initial multi-modal construction data is positioned, the quality defect label is labeled, the text data in the quality defect label is obtained, and the image data is obtained.

[0067] The date and location information in the image data are matched with the 3D coordinates of the BIM model, the sensor data and the environment state in the image are fused, the spatio-temporal joint features are generated, and the second correlation data is obtained.

[0068] The static rule base and the dynamic rule base of the first correlation data and the second correlation data are established, and the Neo4j graph database is used to store heterogeneous data correlation, the multi-hop relationship in the knowledge graph is encoded into a feature vector, and the feature multi-modal construction data is obtained.

[0069] The beneficial effects are that, by obtaining multi-modal construction data in the system, the multi-modal construction data is sequentially subjected to data preprocessing and standardization, the text and numerical values in the initial multi-modal construction data are aligned, the aligned text and numerical value data and image spatio-temporal data are used to construct a knowledge graph, an LSTM-GRU-HMM construction cost prediction model is established by connecting an LSTM-GRU hybrid network and an HMM hidden Markov model, a spatio-temporal attention mechanism is added to the model, the feature multi-modal construction data is input into the model for prediction, the model is iteratively trained by a PPO algorithm, a Pareto optimal solution set is generated, and materials and manpower in construction projects are uniformly scheduled and managed based on the minimum cost path and the risk contingency library.1. The traditional multi-modal data semantic fragmentation problem is solved, 2. The key spatio-temporal area (such as a high-cost risk section) can be dynamically focused, the spatial resolution of the prediction result is improved to the section level, and a basis is provided for accurate scheduling; 3. Based on historical data and real-time prediction, emergency plans for more than 20 kinds of risk scenarios (such as heavy rain and material shortage) are automatically generated, and the response speed is several times faster than that of manual preparation; 4. The minimum cost path and the risk contingency library are combined to realize dynamic allocation of materials and manpower, and the resource utilization rate is improved. BRIEF DESCRIPTION OF DRAWINGS

[0070] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered a limitation of the present application.

[0071] Figure 1A first embodiment schematic view of a construction cost analysis system based on big data in the embodiments of the present application;

[0072] Figure 2 A second embodiment schematic view of a construction cost analysis system based on big data in the embodiments of the present application;

[0073] Figure 3 A third embodiment schematic view of a construction cost analysis system based on big data in the embodiments of the present application;

[0074] Figure 4 A first embodiment schematic view of a construction cost analysis method based on big data in the embodiments of the present application. DETAILED DESCRIPTION

[0075] In order to make the objects, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0076] Those skilled in the art can understand that, unless specifically stated, the singular form "a", "an", "said" used herein can also include the plural form. It should be further understood that the phrase "comprising" used in the specification of the present application means that the features, integers, steps, operations, elements and / or components exist, but does not exclude the existence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0077] The present application will be specifically described below in combination with the accompanying drawings, such as Figure 1 As shown in the figure, the construction cost analysis system based on big data comprises the following modules:

[0078] 101, a construction data acquisition module, configured to acquire multi-modal construction data in the system, sequentially perform data preprocessing and standardization on the multi-modal construction data, and obtain initial multi-modal construction data;

[0079] Specifically, the embodiment further comprises a data acquisition sub-module configured to access an enterprise ERP system, a BIM model and a public database through an API interface, and acquire material inventory data, bill of quantities data and building material price index;

[0080] An image scanning sub-module configured to acquire construction log image in an OCR scanner, and extract text data of the construction log image;

[0081] A sensor sub-module configured to collect site environment data and manpower data in a construction site by using a temperature and humidity sensor and a vibration sensor;

[0082] The standardized submodule is configured to perform outlier deletion, missing value supplementation and standardization processing on material inventory data, bill of quantities data, building material price index, construction log image, text data, construction site environment data and human data, to obtain initial multi-modal construction data.

[0083] The feature data alignment module is configured to align text and numerical values in the initial multi-modal construction data, and align images and space-time, to construct a knowledge graph based on the aligned text and numerical value data and image and space-time data, to obtain feature multi-modal construction data.

[0084] Specifically, the embodiment further includes an extraction submodule configured to extract text name data in the initial multi-modal construction data by using an NLP natural language processing model, the text name data at least including material name, construction period node and risk event; the extraction submodule is further configured to associate the text name data and numerical value label, to establish a mapping relationship, to obtain first associated data.

[0085] The positioning submodule is configured to position an image region in the initial multi-modal construction data based on a YOLO target detection model, to label quality defect labels, to obtain text data in the quality defect labels, and to obtain image data.

[0086] The matching submodule is configured to match date and location information in the image data with 3D coordinates of a BIM model, to fuse sensor data and environment states in the image, to generate space-time joint features, and to obtain second associated data.

[0087] The establishment submodule is configured to establish a static rule base and a dynamic rule base of the first associated data and the second associated data, and to store heterogeneous data association by using a Neo4j graph database, to encode multi-hop relationships in the knowledge graph as feature vectors, to obtain feature multi-modal construction data.

[0088] The prediction model construction module is configured to connect an LSTM-GRU hybrid network and an HMM hidden Markov model, to establish an LSTM-GRU-HMM construction cost prediction model, and to add a space-time attention mechanism in the model, to obtain a target LSTM-GRU-HMM construction cost prediction model.

[0089] Specifically, the embodiment further includes a connection submodule configured to connect the LSTM-GRU hybrid network and the HMM hidden Markov model, to obtain feature vectors of historical construction data in the system .

[0090] The extraction submodule is configured to extract long-term dependence features of the feature vectors by using an LSTM layer, to output a first hidden state, and to learn short-term fluctuation patterns of the feature vectors by using a GRU layer, to output a second hidden state.

[0091] a splicing submodule configured to splice the first hidden state and the second hidden state into a mixed feature to obtain initial construction feature data;

[0092] define a market policy stage in historical construction data as a hidden state of an HMM (Hidden Markov Model) , and a state transition matrix of the model is as follows:

[0093]

[0094] wherein, denotes a probability of transition from a state to a state , and denotes a hidden state at a time point .

[0095] a generation submodule configured to generate an observation probability distribution by using the initial construction feature data to obtain an observation probability of the HMM (Hidden Markov Model):

[0096]

[0097]

[0098] wherein, denotes an observation vector, denotes a state generates a probability of the observation vector , and denotes the initial construction feature data, denotes an activation function, denotes a weight matrix, denotes a bias;

[0099] a calculation submodule configured to solve an optimal state sequence in the model to obtain macro construction state label data.

[0100] an attention mechanism adding unit configured to add a space-time attention mechanism in an LSTM-GRU-HMM construction cost prediction model, and dynamically allocate cost weights of different regions and construction stages in the macro construction state label data by using the space-time attention mechanism;

[0101] a prediction weight unit configured to calculate an importance weight of a historical time step in the macro construction state label data for a current prediction by using a time attention mechanism.

[0102]

[0103]

[0104] wherein, represents the output of the scoring function in the attention mechanism, represents the weight matrix in the time attention mechanism, which linearly transforms the spliced features , represents the macro construction state label data, represents the features at the current time, represents the bias term in the time attention mechanism;

[0105] importance weight so that the model can automatically identify and strengthen the historical fragments that are strongly related to the current prediction. For example, when predicting the concrete pouring cost of a certain section, if there is a material price fluctuation or an extreme weather event in the historical data at the same period, the importance weight of the corresponding time step will be significantly increased, so that the contribution of key historical information (such as a jump in steel price three months ago) to the total prediction is increased to more than 60%, and the influence of irrelevant time periods (such as stable period data one year ago) is weakened. This dynamic focusing mechanism solves the problem of long-term dependence loss caused by fixed time windows in traditional RNN.

[0106] importance weight reflects the confidence of historical data on the current prediction. When the model encounters data noise or outliers (such as cost record deviation caused by sensor failure), the importance weight of the relevant time step will automatically decrease to avoid the transmission of false information. For example, if the cost data of a certain week shows abnormally low values due to equipment failure, the time attention mechanism will compress its importance weight to less than 0.1, while increasing the weights of the normal weeks before and after it (total > 0.9), thereby maintaining time continuity while suppressing abnormal interference. This quantitative uncertainty capability reduces the prediction error of the model in a non-stationary construction environment (such as a policy adjustment period) by 15-20%.

[0107] By adjusting the calculation granularity of the importance weight , the model can capture both short-term fluctuations and long-term trends. In daily prediction, the importance weight will focus on the data of the last 7 days (accounting for > 70%) to reflect changes in construction progress; in monthly prediction, the importance weight is evenly distributed to the past 3 months, highlighting the influence of material procurement cycles. This multi-scale fusion enables the model to accurately capture sudden events (such as downtime costs caused by equipment failure) at the section level, as well as accurately predict monthly funding needs (such as engineering payments based on progress), with a spatial resolution of specific construction areas (such as the main tower pouring area), which is 30% higher than traditional methods.

[0108] a weight calculation unit configured to calculate spatial attention weights according to region features in the macro construction state label data calculate spatial weights .

[0109] a model obtaining unit configured to combine the attention weights with the mixed features to generate a final prediction input, and obtain a target LSTM-GRU-HMM construction cost prediction model after training is completed.

[0110] The LSTM-GRU-HMM construction cost prediction model combines the long-term memory capability of LSTM, the calculation efficiency of GRU, and the probability inference advantage of HMM, and after combining the spatio-temporal attention mechanism, can significantly improve the ability to capture the dynamic changes of construction cost. The following explains from three aspects of network structure, super parameter value selection, and training steps:

[0111] The network structure adopts a hierarchical hybrid architecture, the bottom layer is an LSTM-GRU parallel module, the middle layer is a spatio-temporal attention mechanism, and the top layer is an HMM state transition layer. The LSTM module controls the long-term information flow through the forget gate, input gate, and output gate, and the size of the hidden layer can be 128-256, which is suitable for processing long-term dependent factors in construction cost (such as material price cycle); the GRU module simplifies the calculation through the update gate and the reset gate, and the size of the hidden layer can be 64-128, which is suitable for capturing short-term fluctuations (such as construction progress changes). The outputs of the two are input into the spatio-temporal attention layer after being spliced or weighted, the spatio-temporal attention mechanism includes spatial attention (assigning weights to different cost items) and temporal attention (adjusting the contribution of different time steps), and the output feature vector is input into the HMM layer. The HMM layer defines the hidden state (such as cost rising, falling, and stable) and the observation probability, decodes the optimal state sequence through the Viterbi algorithm, and finally outputs the cost prediction value.

[0112] The range of hyperparameters should consider both performance and computational efficiency. The learning rate can be initially set to 1e-3 and adjusted dynamically (e.g., ReduceLROnPlateau) to 1e-4 when the validation loss does not decrease for 5 consecutive rounds. The dropout rate is set to 0.2-0.3 between the LSTM and GRU input layers and 0.1-0.2 between the hidden layers to prevent overfitting. The batch size is selected based on the data volume, with 32-64 for small datasets (e.g., single projects) and 128-256 for large datasets (e.g., regional projects). The sequence length should cover the cost variation period, with 10-30 for short-term prediction (e.g., monthly) and 50-100 for long-term prediction (e.g., annual). The number of iterations is controlled by early stopping, with patience set to 10 and a maximum of 200 iterations. For weight initialization, Xavier is used for LSTM to stabilize gradients, and He is used for GRU to accelerate convergence. The optimizer is recommended to be Adam with β1=0.9 and β2=0.999 for noisy data, or Nadam for stable data. The number of hidden states in HMM is set based on the cost fluctuation pattern, with 3-5 for simple scenarios (e.g., fixed processes) and 5-8 for complex scenarios (e.g., multiple process intersections).

[0113] The training step first performs data preprocessing, normalizing the construction cost sequence to [0, 1] using Min-Max normalization, and dividing it into training set (70%), validation set (15%), and test set (15%). The data is divided by sequence length to ensure that each sample contains a complete cost period.

[0114] Then, the model is built using PyTorch to implement the hybrid network. After the LSTM-GRU module outputs are concatenated, they are reduced to feature vectors through a fully connected layer. The spatio-temporal attention layer calculates spatial weights (Softmax assigns cost item weights) and temporal weights (adjusts time step contribution), and outputs the weighted features. The HMM layer initializes the state transition matrix (randomly or based on prior knowledge) and the observation probability matrix.

[0115] The training configuration is then performed, with the loss function combining MSE regression tasks and HMM log likelihood, the optimizer using Adam, and the learning rate scheduler using ReduceLROnPlateau. The device is selected to be GPU-accelerated, and the batch is grouped according to the Batch Size.

[0116] The training loop is then performed, with forward propagation through LSTM-GRU, attention layer, and HMM layer in sequence, calculating the total loss and then backpropagating to update the parameters. After each validation, the learning rate is reduced if the loss does not decrease. Early stopping monitors the validation loss, and training is terminated if it does not decrease for 10 consecutive rounds.

[0117] Finally, the model is fine-tuned and evaluated. Optuna is used to search for the optimal hyperparameters (such as hidden layer size, learning rate), and the evaluation metrics are MAE, RMSE (regression), and accuracy (HMM state prediction). The model's generalization ability is verified on the test set, and the error sources (such as data noise or model structure) are analyzed. Finally, the LSTM-GRU-HMM construction cost prediction model is obtained.

[0118] 104. A cost risk prediction module for inputting feature multi-modal construction data into a target LSTM-GRU-HMM construction cost prediction model for prediction, generating a Pareto optimal solution set through PPO algorithm iteration model training, and obtaining a minimum cost path and risk plan library;

[0119] Specifically, in the present embodiment, feature multi-modal construction data is input into a target LSTM-GRU-HMM construction cost prediction model for prediction.

[0120] Through PPO algorithm iteration model training, a Pareto optimal solution set is generated. The optimization of the PPO algorithm includes at least a policy network for outputting action probability distribution , and a value network for evaluating state value .

[0121] The importance sampling ratio is:

[0122]

[0123] The clipping target function is:

[0124]

[0125] wherein, represents the advantage function, represents the immediate reward, represents the discount factor, represents the clipping threshold, represents the action probability distribution under the old policy, represents the action, represents the state, represents the expectation, represents the current policy, represents the old policy.

[0126] A Pareto dominance relationship is defined, a diversity solution set is generated by adjusting weights, non-dominated solutions are screened to form a Pareto front, and a minimum cost path and risk plan library are obtained.

[0127] In optimizing the LSTM-GRU-HMM construction cost prediction model based on the PPO algorithm, the reward function design needs to quantify multiple objectives such as cost and risk into calculable, immediate rewards to guide the model in generating a Pareto optimal solution set. Specifically, the cost objective can be quantified as the reciprocal of the deviation between the predicted cost and the actual cost, i.e.:

[0128]

[0129] Where Cpred is the model-predicted cost and Ctrue is the actual cost. This function maps the deviation to the interval (0,1], with smaller deviations resulting in higher rewards, encouraging accurate model predictions. The risk objective is defined by the negative logarithm of risk indicators (such as the probability of cost overruns and the probability of project delays), i.e.:

[0130]

[0131] Where Prisk is the risk probability evaluated by the model, and ϵ is a local minimum (e.g., 1e-5) to avoid the undefined logarithm. This design ensures that high-risk scenarios receive significant negative rewards, forcing the model to avoid risky paths. The overall reward is calculated using a weighted summation:

[0132]

[0133] The weights wcost and wrisk are determined through Pareto front analysis. Balancing cost minimization and risk control, the model is iteratively optimized through the PPO algorithm. In multiple rounds of training, the model gradually approaches the Pareto front, ultimately generating a minimum cost path and corresponding risk contingency plan library covering different risk-cost trade-offs.

[0134] 105. The Human Resources and Materials Scheduling Module is used to uniformly schedule and manage materials and human resources in construction projects based on the minimum cost path and risk contingency plan library.

[0135] Specifically, this embodiment also includes a task decomposition unit, which is used to decompose the construction task into sub-tasks of material procurement, manpower allocation and equipment scheduling according to the key nodes in the minimum cost path, generate a spatiotemporal constraint diagram based on the BIM model, and identify resource conflict areas.

[0136] The resource allocation unit is used to mark tasks with a cost weight >15% in the path as critical path tasks, prioritize resource allocation, and reserve 20% resource redundancy for high-risk tasks in the contingency plan library.

[0137] The risk assessment unit is used to match three-level response plans from the contingency plan library based on the type and level of the risk event. It includes at least level one, level two, and level three responses. Level one response is used to directly initiate preset actions, level two response is used to generate multiple options, and level three response is used to trigger manual intervention procedures.

[0138] Sub-step 1: Resource Schedule Decomposition and Priority Setting;

[0139] Path Analysis and Task Decomposition:

[0140] According to the key nodes in the minimum cost path (prefer to purchase local sand and gravel), the construction tasks are decomposed into material procurement (material type, quantity), manpower allocation (work type, shift), and equipment scheduling (tower crane, excavator usage period) sub-tasks.

[0141] Based on the BIM model, a space-time constraint graph is generated to identify resource conflict areas (A area concrete pouring and B area steel bar binding need to share tower crane).

[0142] Priority Dynamic Sorting:

[0143] Cost-sensitive tasks: tasks with cost weight > 15% in the path (high-priced material procurement) are marked as "critical path tasks" and resources are allocated preferentially;

[0144] Risk-averse tasks: 20% resource redundancy is reserved in advance for high-risk tasks in the contingency library (construction in rainy season).

[0145] Sub-step 2: Multi-level Inventory and Flexible Labor Pool Management;

[0146] Material Classification Inventory Strategy:

[0147] Strategic Inventory: For long-cycle materials (special steel), stock up 3 months in advance according to forecast demand;

[0148] Tactical Inventory: For volatile materials (cement), adopt "Vendor Managed Inventory (VMI)" mode, replenish weekly;

[0149] Emergency Inventory: For short materials in the risk plan (waterproofing membrane), set up safety stock, trigger threshold automatic replenishment.

[0150] Flexible Labor Scheduling Mechanism:

[0151] Core Team Fixed: Technical workers (welders, surveyors) are stationed on site for the entire project period;

[0152] Temporary Labor Pool: General workers (laborers) are hired on a daily basis through a labor platform and their number is dynamically adjusted according to progress;

[0153] Cross-project Allocation: Share enterprise-level labor pool, allocate personnel from low-priority projects during peak periods.

[0154] Sub-step 3: Real-time Monitoring and Dynamic Scheduling Engine;

[0155] Resource Status Board:

[0156] Visualized display of real-time resource utilization (82% of tower crane usage), inventory warning (rebar inventory below safety line), and manpower gap (3 workers short for night shift).

[0157] Dynamic scheduling rule engine:

[0158] Rule base: "If a material is delayed for more than 3 days, activate the backup supplier and adjust the construction sequence";

[0159] Automatic triggering mechanism: When the sensor detects "continuous heavy rain for more than 6 hours", automatically push the "drainage plan" in the pre-plan library and suspend open-air operations.

[0160] Sub-step 4: Risk pre-plan integration and emergency response;

[0161] Pre-plan matching and execution:

[0162] According to the risk event type (supply chain interruption, extreme weather) and level (high / medium / low), match the three-level response plan from the pre-plan library:

[0163] First-level response: Directly activate the preset action (call strategy library);

[0164] Second-level response: Generate multiple scheme options (adjust process or delay construction) for decision-making;

[0165] Third-level response: Trigger manual intervention process (report to headquarters for resource coordination).

[0166] Post-mortem and pre-plan iteration:

[0167] Record the emergency event handling results, and optimize the pre-plan library through feedback loop (reduce the priority of a supplier who has violated multiple times).

[0168] Sub-step 5: Deployment of multi-dimensional collaborative management platform;

[0169] Platform function modules:

[0170] Resource cockpit: Integrates cost prediction, inventory, manpower, and equipment data, supports multi-dimensional drilling analysis;

[0171] Mobile collaboration: Site managers receive dispatching instructions and feedback execution status through APP;

[0172] Supplier portal: Open interface for suppliers to query order status and report supply exceptions.

[0173] Cross-departmental collaboration process:

[0174] Purchasing-construction linkage: The purchasing department dynamically adjusts the order delivery time according to the construction progress;

[0175] Financial-engine reconciliation: automatically generate resource consumption and cost deviation reports to support quick correction.

[0176] The beneficial effects are that 1) the traditional multi-modal data semantic split problem is solved, 2) the key space-time region (such as a high-cost risk section) can be dynamically focused, so that the spatial resolution of the prediction result is improved to the section level, providing a basis for accurate scheduling; 3) based on historical data and real-time prediction, an emergency plan for more than 20 kinds of risk scenarios (such as heavy rain and material shortage) is automatically generated, and the response speed is several times faster than that of manual preparation; 4) combined with the minimum cost path and the risk plan library, dynamic allocation of materials and manpower is realized, and resource utilization is improved.

[0177] In this embodiment, please refer to Figure 2 In the second embodiment of the construction cost analysis system based on big data in the embodiment of the application, the feature data alignment module includes the following submodules:

[0178] The extraction submodule is configured to extract text name data in the initial multi-modal construction data by using an NLP natural language processing model, the text name data at least including material names, construction period nodes, and risk events; the text name data is associated with numerical labels to establish a mapping relationship, and first associated data is obtained;

[0179] The positioning submodule is configured to position an image region in the initial multi-modal construction data based on a YOLO target detection model, label quality defect labels, obtain text data in the quality defect labels, and obtain image data;

[0180] The matching submodule is configured to match date and location information in the image data with 3D coordinates of a BIM model, fuse sensor data and environmental states in the image, generate spatio-temporal joint features, and obtain second associated data;

[0181] The establishment submodule is configured to establish a static rule library and a dynamic rule library of the first associated data and the second associated data, store heterogeneous data associations by using a Neo4j graph database, encode multi-hop relationships in a knowledge graph as feature vectors, and obtain feature multi-modal construction data.

[0182] The construction of the knowledge graph realizes the mapping from raw data to structured knowledge through multi-modal data fusion. The following takes a bridge project as an example to completely show a rule instance of the knowledge graph:

[0183] The raw data includes three types:

[0184] Text data (extracted by NLP): "2023-05-10, C50 concrete pouring of main tower is completed, 120 tons of HRB400 steel is used, and 3 honeycomb defects are found" is recorded in the construction log;

[0185] Image data (YOLO localization): A photo taken by the drone at the pouring site, labeled "Honeycomb defect (L5m x W0.3m x D0.1m) at the southeast corner of the main tower" and timestamped "2023-05-10 14:20";

[0186] BIM model data: 3D coordinates of the main tower (X=120.5, Y=85.2, Z=68.3) and environmental temperature and humidity (28°C, 65%RH) monitored by sensors.

[0187] Rule instance mapping process:

[0188] NLP extraction submodule: Identify entities "HRB400 steel" (material name), "2023-05-10" (project milestone), and "honeycomb surface" (risk event) in the text, associate the value label "120 tons", and generate the first associated data:

[0189] (HRB400 steel)-[consumption]->(120 tons);

[0190] (2023-05-10)-[risk event]->(honeycomb surface);

[0191] YOLO localization submodule: Locate the defect area coordinates (X=120.8, Y=85.0, Z=68.0) from the image, extract the text label "honeycomb defect", and generate the image data:

[0192] ( southeast corner of the main tower)-[defect type]->(honeycomb defect)-[size]->(L5m x W0.3m x D0.1m);

[0193] Matching submodule: Match the image date "2023-05-10" with the BIM model coordinates, and fuse the sensor temperature and humidity data to generate the second associated data:

[0194] (2023-05-10 14:20)-[environmental conditions]->(temperature 28°C / humidity 65%RH)-[spatial location]->(X120.8, Y85.0, Z68.0);

[0195] Establishment submodule: Through static rules (such as "honeycomb defect → need to be reworked") and dynamic rules (such as "temperature and humidity > 25°C / 60% → concrete strength reduction by 5%"), build multi-hop relationships in Neo4j:

[0196] (HRB400 steel)-[used for]->(main tower pouring)-[occurrence time]->(2023-05-10);

[0197] (main tower pouring)-[produces defect]->(honeycomb defect)-[associated risk]->(project delay 3 days);

[0198] (2023-05-10)-[environmental impact]->(concrete strength C45→C42.75);

[0199] The final encoding is a feature vector [material consumption, defect level, environmental factor, risk probability], forming a feature multi-modal data supporting a spatio-temporal attention mechanism.

[0200] Its beneficial effects are that it solves the traditional multi-modal data semantic fragmentation problem. In knowledge graph construction, through entity relationship extraction technology, the component information in engineering drawings is associated with construction progress data to form a structured knowledge network.

[0201] In this embodiment, please refer to Figure 3 The third embodiment of the construction cost analysis system based on big data in the embodiment of the application includes the following units:

[0202] The task decomposition unit is used to decompose the construction task into material procurement, human resource allocation and equipment scheduling sub-tasks according to the key nodes in the minimum cost path, and generate a space-time constraint graph based on the BIM model to identify resource conflict areas;

[0203] The resource allocation unit is used to mark the tasks with a cost weight > 15% in the path as key path tasks, and allocate resources preferentially, and reserve 20% resource redundancy for high-risk tasks in the contingency library;

[0204] The risk judgment unit is used to match a three-level response scheme from the contingency library according to the risk event type and level, including at least a first-level response, a second-level response and a third-level response, the first-level response is used to directly start a preset action, the second-level response is used to generate a multi-scheme selection, and the third-level response is used to trigger a manual intervention process.

[0205] Its beneficial effects are that it combines the minimum cost path and the risk contingency library to realize dynamic allocation of materials and human resources. When the demand for concrete in a certain section increases sharply, the system automatically calls the inventory of nearby warehouses and adjusts the transportation route to reduce costs.

[0206] The above describes the construction cost analysis system based on big data provided by the embodiment of the application, and the construction cost analysis method based on big data of the embodiment of the application is described below, please refer to Figure 4 The construction cost analysis method based on big data in the embodiment of the application includes:

[0207] Step 401, acquire multi-modal construction data in the system, and sequentially perform data preprocessing and standardization on the multi-modal construction data to obtain initial multi-modal construction data;

[0208] In step 402, the text and numerical values in the initial multi-modal construction data are aligned, and the images and space-time are aligned, the aligned text and numerical value data and the image and space-time data are used to construct a knowledge graph, and feature multi-modal construction data is obtained.

[0209] In step 403, an LSTM-GRU-HMM construction cost prediction model is established by connecting an LSTM-GRU hybrid network and an HMM hidden Markov model, and a space-time attention mechanism is added to the model to obtain a target LSTM-GRU-HMM construction cost prediction model.

[0210] In step 404, the feature multi-modal construction data is input into the target LSTM-GRU-HMM construction cost prediction model for prediction, the model is iteratively trained by a PPO algorithm, a Pareto optimal solution set is generated, a minimum cost path and a risk plan library are obtained.

[0211] In step 405, the materials and manpower in the construction project are uniformly scheduled and managed based on the minimum cost path and the risk plan library.

[0212] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various modifications and improvements can be made to the present application, and these modifications and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A construction cost analysis system based on big data, characterized in that, The construction cost analysis system includes the following modules: A construction data acquisition module is used to acquire multimodal construction data in the system, and to perform data preprocessing and standardization on the multimodal construction data in sequence to obtain initial multimodal construction data; The feature data alignment module is used to align the text and numerical values ​​in the initial multimodal construction data, and to align the images and spatiotemporal data. The aligned text and numerical data and image spatiotemporal data are used to construct a knowledge graph to obtain the feature multimodal construction data. The prediction model building module is used to connect the LSTM-GRU hybrid network and the HMM hidden Markov model to establish an LSTM-GRU-HMM construction cost prediction model. A spatiotemporal attention mechanism is added to the model to obtain the target LSTM-GRU-HMM construction cost prediction model. The cost risk prediction module is used to input the feature multimodal construction data into the target LSTM-GRU-HMM construction cost prediction model for prediction. Through training the PPO iterative model, a Pareto optimal solution set is generated to obtain the minimum cost path and risk contingency plan library. The human resources and materials scheduling module is used to uniformly schedule and manage materials and human resources in construction projects based on the minimum cost path and risk contingency plan library.

2. The construction cost analysis system based on big data as described in claim 1, characterized in that, The construction data acquisition module includes the following sub-modules: The data acquisition submodule is used to access the enterprise's ERP system, BIM model and public database through API interface to obtain material inventory data, bill of quantities data and building material price index; The image scanning submodule is used to acquire construction log images from the OCR scanner and extract text data from the construction log images; The sensor submodule is used to collect site environmental data and human data using temperature and humidity sensors and vibration sensors. The standardization submodule is used to perform outlier removal, missing value supplementation, and standardization processing on the material inventory data, bill of quantities data, building material price index, construction log images, text data, construction site environment data, and manpower data to obtain initial multimodal construction data.

3. The construction cost analysis system based on big data as described in claim 1, characterized in that, The feature data alignment module includes the following sub-modules: The extraction submodule is used to extract text name data from the initial multimodal construction data using an NLP natural language processing model, including at least material names, construction period nodes, and risk events; and to associate the text name data with numerical labels to establish a mapping relationship and obtain the first associated data. The localization submodule is used to locate the image region in the initial multimodal construction data based on the YOLO object detection model, label the quality defect, obtain the text data in the quality defect label, and obtain the image data. The matching submodule is used to match the date and location information in the image data with the 3D coordinates of the BIM model, fuse sensor data and environmental conditions in the image, generate spatiotemporal joint features, and obtain the second associated data. A submodule is established to create static and dynamic rule bases for the first and second associated data, and to store heterogeneous data associations using the Neo4j graph database. Multi-hop relationships in the knowledge graph are encoded into feature vectors to obtain feature multimodal construction data.

4. The construction cost analysis system based on big data as described in claim 1, characterized in that, The prediction model construction module includes the following sub-modules: The connection submodule is used to connect the LSTM-GRU hybrid network and the HMM hidden Markov model to obtain the feature vectors of historical construction data in the system. ; The extraction submodule is used to extract the feature vector using an LSTM layer. The long-term dependency features are used to output the first hidden state, and the feature vector is learned using a GRU layer. The short-term fluctuation pattern outputs the second hidden state; The splicing submodule is used to splice the first hidden state and the second hidden state into a hybrid feature to obtain the initial construction feature data; The market policy phase in the historical construction data is defined as the hidden state of the Hidden Markov Model (HMM). Then the state transition matrix of the model as follows: ； in, Indicates from state Transition to state The probability, Indicates time The hidden state; The generation submodule is used to generate an observation probability distribution using the initial construction feature data, thereby obtaining the observation probability of the Hidden Markov Model (HMM): ； ； in, Represents the observation vector. Representing state Generate observation vectors The probability, This represents the initial construction feature data. This represents the activation function. Represents the weight matrix. Indicates bias; The computation submodule is used to solve for the optimal state sequence in the model. This yields macro-level construction status label data.

5. The construction cost analysis system based on big data as described in claim 1, characterized in that, The prediction model construction module includes the following units: An attention mechanism addition unit is used to add a spatiotemporal attention mechanism to the LSTM-GRU-HMM construction cost prediction model, and to dynamically allocate cost weights for different regions and construction stages in the macro construction status label data using the spatiotemporal attention mechanism. The prediction weighting unit is used to calculate the importance weight of historical time steps in the macro-construction status label data to the current prediction using a time attention mechanism. ; ； ； in, This represents the output of the scoring function in the attention mechanism. This represents the weight matrix in the temporal attention mechanism, applied to the concatenated features. Perform a linear transformation. This represents macro-level construction status label data. Indicates the characteristics of the current moment. This represents the bias term in the time attention mechanism; The weight calculation unit is used to utilize spatial attention based on the regional characteristics in the macroscopic construction status label data. Calculate spatial weights ; The model obtains units, which are used to combine attention weights with hybrid features to generate the final prediction input. After training, the target LSTM-GRU-HMM construction cost prediction model is obtained.

6. The construction cost analysis system based on big data as described in claim 1, characterized in that, The cost risk prediction module includes the following units: The feature-based multimodal construction data is input into the target LSTM-GRU-HMM construction cost prediction model for prediction. By training the PPO iterative model, a Pareto optimal solution set is generated. The optimization of the PPO includes at least a policy network for outputting action probability distributions. Value networks are used to evaluate the value of a state. ; Importance sampling ratio is: ； Pruning objective function: ； in, Represents the dominance function. Indicates an immediate reward. Indicates the discount factor. Indicates the clipping threshold. This represents the probability distribution of actions under the old policy. Indicates an action, Indicates state, Expressing expectations, Indicates the current strategy. Indicates the old strategy; Define the Pareto dominance relationship, generate a diverse set of solutions by adjusting the weights, select non-dominated solutions to form the Pareto front, and obtain the minimum cost path and risk contingency plan library.

7. The construction cost analysis system based on big data as described in claim 1, characterized in that, The manpower and material scheduling module includes the following units: The task decomposition unit is used to decompose the construction task into sub-tasks such as material procurement, manpower allocation and equipment scheduling based on the key nodes in the minimum cost path, and generate a spatiotemporal constraint diagram based on the BIM model to identify resource conflict areas. The resource allocation unit is used to mark tasks with a cost weight >15% in the path as critical path tasks, prioritize resource allocation, and reserve 20% resource redundancy for high-risk tasks in the contingency plan library. The risk assessment unit is used to match a three-level response plan from the contingency plan library based on the type and level of the risk event. The plan includes at least a level 1 response, a level 2 response, and a level 3 response. The level 1 response is used to directly initiate a preset action, the level 2 response is used to generate multiple options, and the level 3 response is used to trigger a manual intervention process.

8. A construction cost analysis method based on big data, characterized in that, The construction cost analysis method includes the following modules: The system acquires multimodal construction data, and performs data preprocessing and standardization sequentially on the multimodal construction data to obtain initial multimodal construction data; The text and numerical values ​​in the initial multimodal construction data are aligned, and the images and spatiotemporal data are aligned. The aligned text and numerical data and image spatiotemporal data are used to construct a knowledge graph to obtain feature multimodal construction data. By connecting the LSTM-GRU hybrid network and the HMM hidden Markov model, an LSTM-GRU-HMM construction cost prediction model is established. A spatiotemporal attention mechanism is added to the model to obtain the target LSTM-GRU-HMM construction cost prediction model. The feature multimodal construction data is input into the target LSTM-GRU-HMM construction cost prediction model for prediction. Through training the PPO iterative model, a Pareto optimal solution set is generated, and the minimum cost path and risk contingency plan library are obtained. Based on the minimum cost path and risk contingency plan library, materials and manpower in the construction project are uniformly scheduled and managed.

9. The construction cost analysis method based on big data as described in claim 8, characterized in that, The acquisition system obtains multimodal construction data, which is then preprocessed and standardized sequentially to obtain initial multimodal construction data, including: By connecting to the enterprise's ERP system, BIM model and public database through API interface, material inventory data, bill of quantities data and building material price index can be obtained; Acquire construction log images from an OCR scanner and extract text data from the construction log images; Temperature and humidity sensors and vibration sensors are used to collect site environmental data and manpower data in the construction site. The material inventory data, bill of quantities data, building material price index, construction log images, text data, site environment data, and manpower data are processed by outlier removal, missing value supplementation, and standardization to obtain initial multimodal construction data.

10. The construction cost analysis method based on big data as described in claim 8, characterized in that, The process involves aligning the text and numerical values ​​in the initial multimodal construction data, and aligning the images and spatiotemporal data. The aligned text and numerical data, along with the image and spatiotemporal data, are then used to construct a knowledge graph, resulting in feature-based multimodal construction data, including: The text name data in the initial multimodal construction data is extracted using an NLP natural language processing model, including at least material names, construction period nodes, and risk events; the text name data and numerical labels are associated to establish a mapping relationship and obtain the first associated data; Based on the YOLO object detection model, image regions in the initial multimodal construction data are located, quality defect labels are marked, and text data in the quality defect labels are obtained to obtain image data; The date and location information in the image data are matched with the 3D coordinates of the BIM model, and the sensor data and the environmental state in the image are fused to generate spatiotemporal joint features, thus obtaining the second associated data. Establish static and dynamic rule bases for the first and second associated data, and use the Neo4j graph database to store heterogeneous data associations. Encode multi-hop relationships in the knowledge graph into feature vectors to obtain feature multimodal construction data.

Images