A construction project management method based on big data

Abstract

This application proposes a construction project management method based on big data, which belongs to the field of construction project management technology. Real-time images and data are obtained through drones, cameras, sensors and other equipment, and machine learning, deep learning and other technologies are combined to optimize the scheduling of construction progress, quality monitoring, safety management, material management and other aspects to achieve refined construction management. This method can improve construction quality and efficiency, reduce the occurrence of safety accidents, and reduce construction costs. The study uses long short-term memory networks (LSTM) to predict construction progress trends, introduces reinforcement learning algorithms, combines historical project data, real-time resource status, weather forecasts and other information, and dynamically adjusts project plans to shorten construction periods and reduce delay costs. Combined with multivariate data sets, the nodes are embedded and learned through the GCN model, and reinforcement learning is used to optimize material management. Ensure construction site safety through face recognition technology and deep learning algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of construction project management, and in particular relates to a construction project management method based on big data. Background Art

[0002] Using the concept of big data, we build a construction project quality supervision platform to comprehensively monitor the construction of the project and improve the overall construction quality of the project. Using the information monitoring platform, we can understand the project construction process and dynamically monitor the construction project, which is conducive to the efficient operation of project management.

[0003] A document with prior art publication number CN116823171A proposes a construction project quality supervision and management method based on big data. A three-dimensional model database of construction projects is pre-established in a construction project detection platform using a big data method, and combined with real-time on-site engineering data, a three-dimensional visualization model with the highest matching degree is selected from the three-dimensional model database of the construction project, and the selected three-dimensional visualization model is then modified and annotated to indicate information such as the dimensions of major construction parts, the types of materials used, the corresponding quality control points, and key maintenance parts.

[0004] Although this method can improve the efficiency and quality of construction project management to a certain extent, there are still some problems and challenges. The direction of construction project management is relatively single, and the intelligence level of material management, progress management and safety management still needs to be improved. In terms of quality management, although quality control can be carried out based on three-dimensional models, there is a lack of dynamic data monitoring and refined quality management cannot be achieved.

[0005] In view of this, we propose a construction project management method based on big data. Summary of the invention

[0006] The present invention provides a construction project management method based on big data, and the technical solution includes the following aspects.

[0007] A construction project management method based on big data, the method comprising:

[0008] S1. Set up data collection devices and inspections at the construction site to collect real-time images and data and transmit them back to the central database;

[0009] S2. Divide the construction site into different sub-areas, conduct quality monitoring for each area, determine the quality level of each sub-area, and mark the corresponding area in the 3D model;

[0010] S3, using drones to obtain construction images, using semantic segmentation, calculating construction progress, and predicting remaining construction time. If a delay is predicted, using reinforcement learning algorithms to optimize project planning and task allocation;

[0011] Reinforcement learning algorithms optimize project planning and task allocation:

[0012] Calculate and predict construction progress P^ t Design and construction schedule plan，t Deviation: ΔP t =P^ t -P plan，t ;

[0013] If |ΔP t ∣≥δ, δ is the set threshold, and the reinforcement learning algorithm is enabled;

[0014] The goal of the reinforcement learning algorithm is to maximize the expected cumulative reward Gt: G t =∑ k＝0 ∞ γ k R t+k+1 ;

[0015] Among them, R t is the immediate reward obtained at time step t, R t+k+1 is the next immediate reward, k is the number of steps of immediate reward, γ is the discount factor, select a policy π, and update the policy to maximize G by interacting with the environment. t ;

[0016] Status t Including historical project data, real-time resource status, weather forecast, action a t Represents project plan adjustment decisions.

[0017] S4. Implement facial recognition and protective equipment recognition to enter the construction site, install radio frequency devices on protective equipment to ensure information consistency, and use cameras in the construction site to detect and identify dangerous behaviors from time to time;

[0018] S5. Embed RFID tags on material packaging to track the entire chain flow, treat materials and construction locations as directed graphs, use graph neural networks for embedding learning, and optimize material scheduling plans through deep reinforcement learning;

[0019] Consider materials and construction locations as a directed graph:

[0020] The construction site is divided into multiple work areas, materials are regarded as nodes, and the flow of materials between different areas is regarded as directed edges;

[0021] Construct a directed graph model including material nodes, location nodes and the transfer relationship between them, and define two types of nodes, namely material nodes M and location nodes L;

[0022] Each material node m∈M represents a specific material, including attributes such as type, specification, and demand;

[0023] Each location node l∈L represents a work area on the construction site, and contains attributes such as spatial distance and construction progress relevance;

[0024] Define directed edges E to represent the flow of materials between different locations;

[0025] For a material node m and a location node l, if material m can flow from location l to another location l′, then there is a directed edge ∈E starting from m and pointing to l′;

[0026] The weight of an edge can be used to represent the strength or probability of material flowing between two locations;

[0027] Optimizing material scheduling through deep reinforcement learning:

[0028] Embed the node into vector h v As the state representation of the reinforcement learning environment, for each node v where the agent needs to make a decision, its state is composed of its own embedding vector and the embedding vectors of its neighboring nodes;

[0029] Define the set of available node actions for the agent at each node. Each node action a usually corresponds to a real-valued or discrete decision variable. Build a deep neural network with the input as the node state representation and the output as the probability distribution of each node action or the deterministic node action value.

[0030] For the deterministic policy gradient method, the network directly outputs the deterministic node action value, a v =π(s v ; θ) = f policy (h v (K) ;θ), where π is the policy function, s v is the state of node v, θ is the parameter of the policy network, and f policy is the forward propagation function of the policy network;

[0031] The reward function can be expressed as R = w cost r cost +w transport r transport -w inventory r inventory -w outage r outage +w timelin ess r timeliness +w usage r usage , where w n is the corresponding reward component rn The weight of ∑w n =1, ensuring the same scale of reward value; procurement cost r cost : Give positive (savings) or negative (overspending) rewards based on the difference between the actual purchase price and the budget or market benchmark price, cost = (budget cost - actual purchase cost) / budget cost, transportation cost r transport , r transport = (estimated transportation cost - actual transportation cost) / estimated transportation cost, inventory holding cost r inventory , r inventory =-[(actual inventory-target inventory) / target inventory] 2 , out-of-stock penalty r outage , r outage ={-λ outage ; 0}, where λ outage is a preset positive value representing the severity of the shortage, the timeliness of material arrival timeliness , r timeliness ={1-(actual arrival time-planned arrival time) / maximum allowed delay time;-1}, material utilization efficiency r usage , r usage = (actual consumption / ideal consumption)·(completed engineering quantity / planned engineering quantity),

[0032] Use the deterministic policy gradient method REINFORCE algorithm to update the policy network parameters θ. The parameter update rule is: Among them, J is the expected cumulative reward, T is the number of sampling steps, and R t is the reward received at time step t;

[0033] The agent executes node actions in the environment according to the deterministic node action values ​​output by the current policy network, observes environmental feedback, and repeats this process, continuously accumulating experience and updating policy network parameters until the preset number of training rounds or convergence conditions are reached.

[0034] Furthermore, the quality level of each sub-area is determined:

[0035] Calculate the qualified rate of each test in each sub-area and build a quality rating model based on neural network;

[0036] Use existing labeled data to train the model, adjust model parameters through cross-validation and grid search methods, and optimize model performance;

[0037] The features of the sub-region to be evaluated are input into the trained model, and the model outputs the quality level of the sub-region.

[0038] Further, calculate the construction progress:

[0039] The drone performs flight missions according to the plan, captures construction images from different angles, and generates 3D model views of corresponding angles from the building information model;

[0040] Process the construction images and 3D model views collected by drones, and apply semantic segmentation models to segment the construction images and 3D model views;

[0041] For the construction image and 3D model view at the same angle, the intersection-and-union ratio corresponding to each segmentation result is calculated as a quantitative indicator of the construction progress at that angle;

[0042] The construction image acquired by the drone is I real , the corresponding 3D model view is I model After semantic segmentation, the pixels of the same category in the two images are marked as S real and S model , the intersection over union ratio is defined as: IoU = |S real ∩S model ∣ / ∣S real ∪S model ∣;

[0043] Among them, |S real ∩S model ∣ represents the intersection area of ​​two types of pixel sets, ∣S real ∪S model ∣ represents the area of ​​the union;

[0044] For a certain shooting angle, the construction progress P angle It can be estimated by combining the intersection ratio with the engineering quantity W; the ratio of the actual completed engineering quantity to the ideal completed engineering quantity under this angle is r, then: r = / W;

[0045] The construction progress percentage of this angle is: P angle =100×r=100×IoU;

[0046] For multiple shooting angles, the weight of each angle is w j , the corresponding construction progress is P j ; Overall construction progress P total Calculated by weighted average: P total =∑ j＝1 N w j ×P j , N represents the total number of shooting angles.

[0047] Furthermore, the remaining construction time is predicted:

[0048] Collect and integrate future weather forecasts, resource status, and attendance to form a multivariate data set that affects construction progress, and normalize or standardize the influencing factor data;

[0049] The influencing factor data D t =Convert to time series format, set the number of LSTM units, input dimension, return sequence, use a fully connected layer with a linear activation function f=wx+b, (where x is input, w is weight, and b is bias term), delete

[0050] The output of the LSTM layer is converted into a construction progress prediction value, and the loss function is the mean square error MSE;

[0051] Train the model, optimize the model parameters through back propagation and gradient update, and use the trained LSTM model to predict the construction progress at the future time step t;

[0052] The input is the normalized influencing factor data sequence D of time step t and several days before t-n:t-1 =(D t-n , D t-n+1 ,,,,,D t-1 ), where n is the observation window length required by the model;

[0053] Generate predicted value P^ t :P^ t =LSTM(D t-n:t-1 ; φ), where φ represents all parameters of the LSTM model;

[0054] Denormalized predicted value P^ t , and obtain the construction progress forecast value on the original scale.

[0055] Further, face recognition and protective equipment recognition are implemented:

[0056] The identity of the entrants is confirmed through facial recognition technology, and computer vision is used to identify whether they are wearing protective equipment correctly;

[0057] The protective equipment is equipped with RFID radio frequency devices to ensure information consistency, and the attendance status is determined by integrating facial recognition and RFID radio frequency devices;

[0058] Conduct posture recognition on construction workers in high-risk areas from time to time to determine whether they are engaging in dangerous behaviors.

[0059] Furthermore, RFID tags are embedded in the material packaging:

[0060] In the logistics process, RFID tags are scanned at warehouses, transportation nodes, and construction site entrances through fixed or handheld reading and writing devices, and uploaded to the central database in real time;

[0061] An RFID access control system is set up at the entrance of the construction site to automatically record the time when materials enter the site;

[0062] Each work area on site is equipped with handheld devices or fixed read-write points to scan materials regularly or on demand to ensure the whereabouts of materials and prevent loss or misuse.

[0063] Furthermore, we use graph neural networks for embedding learning:

[0064] Assign an initial node feature vector x to each node v v , corresponding to their attributes in the graph, the feature vectors of all nodes are arranged into a matrix H = [x1, x2, ..., x N ] T , where N = |M| + |L| is the total number of nodes; x N is the feature vector of the Nth node;

[0065] Message passing and node update are performed through multi-layer GCN to calculate the node feature matrix H of the lth layer (l) :H (l+1) =σ(D ～-1 / 2 A ～ D ～-1 / 2 H (l) W (l) ), where σ is the activation function, A ～ =A+I N is the adjacency matrix with self-loops, A is the original adjacency matrix, I N is the identity matrix, D ～ ii =∑ j A ～ ij is the degree matrix, W (l) is the weight matrix of the lth layer;

[0066] After training is completed, the final hidden state h of each node v , where K is the index of the last layer, which is used as the embedding vector representation of the node. These low-dimensional, dense vectors contain material properties, position characteristics, and transfer rules between nodes.

[0067] The technical solution provided in the embodiments of the present application can bring the following beneficial effects:

[0068] By subdividing the construction site into multiple sub-areas and conducting independent quality monitoring and risk assessment for each sub-area, the construction status can be more accurately grasped, problems can be discovered and solved in a timely manner, and construction quality and efficiency can be improved. Using drones to obtain multi-angle construction images, calculating the construction progress through semantic segmentation of three-dimensional models and images, and predicting the remaining construction time in combination with weather, resource status, and attendance, the construction progress and potential delay risks can be more accurately grasped, providing strong support for project planning and task allocation. Through face recognition and protective equipment recognition technology, the compliance and safety of entrants can be ensured, and the occurrence of safety accidents can be effectively prevented. At the same time, the automatic identification and early warning system for dangerous behaviors can timely discover potential safety hazards and take corresponding measures to ensure construction safety. In terms of material management, by embedding RFID tags, the entire chain of materials can be tracked in real time, ensuring that the source of materials can be traced and preventing the mixing of counterfeit and shoddy products. At the same time, the use of graph neural networks and deep reinforcement learning algorithms for intelligent management and optimized scheduling of materials can improve the efficiency of material use, reduce waste, and reduce construction costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A flowchart of a construction project management method based on big data provided by one embodiment of the present application;

[0070] Figure 2 A structural diagram of a construction project management system based on big data provided by an embodiment of the present application DETAILED DESCRIPTION

[0071] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0072] Please refer to Figure 1 , which shows a flow chart of a construction project management method based on big data provided by an embodiment of the present application;

[0073] S1. Install cameras and deploy various sensors in the construction site, use drones for aerial inspections, and arrange inspection personnel to conduct on-site inspections, collect real-time images and data, and transmit them back to the central database;

[0074] S2. Divide the construction site into different sub-areas, conduct quality monitoring for each area, use machine learning to determine the quality level of each sub-area, and mark the corresponding area in the 3D model;

[0075] S3. Use drones to obtain construction images from different angles, use 3D models and images for semantic segmentation, calculate construction progress, and predict the remaining construction time based on weather, resource status, and attendance. If a delay is predicted, use reinforcement learning algorithms to optimize project planning and task allocation;

[0076] S4. Implement facial recognition and protective equipment recognition to enter the construction site, install radio frequency devices on protective equipment to ensure information consistency, and use cameras in the construction site to detect and identify dangerous behaviors from time to time;

[0077] S5. Embed RFID tags on material packaging to track the entire chain flow, treat materials and construction locations as directed graphs, use graph neural networks for embedded learning, and optimize material scheduling plans through deep reinforcement learning.

[0078] As an optional solution of the present application, in step S2, according to the physical structure, construction stage and management requirements of the construction site, the site is divided into multiple sub-areas, key data monitoring points are set in each sub-area, and the detection data of the monitoring points of each sub-area are continuously collected by various sensors, cameras, drones and manual inspections deployed at the construction site, including but not limited to structural stress, material properties, construction process parameters, etc., and the original detection data is cleaned, missing value processing, abnormal value detection and correction, data standardization and other pre-processing operations are performed to ensure data quality and prepare for subsequent analysis, and the qualified rate of each detection in each sub-area is calculated, and a quality rating model is constructed based on a neural network, and the model is trained with existing annotated data (including sub-area features and corresponding quality grades), and the model parameters are adjusted by cross-validation, grid search and other methods to optimize the model performance, and the features of the sub-area to be evaluated are input into the trained model, and the model outputs the quality grade of the sub-area (such as excellent, good, qualified, unqualified, etc.), and the quality grade of each sub-area is mapped to the corresponding three-dimensional building model. Color coding, icon marking, text annotation and other forms can be used to intuitively display the quality status of each sub-area.

[0079] In step S3, the construction progress is calculated by combining semantic segmentation technology with intersection over union ratio. Specifically, the drone performs flight missions according to the plan and captures high-definition construction images from different angles in real time. According to the current construction progress, a 3D model view of the corresponding angle is generated from the building information model (BIM). The construction images and 3D model views collected by the drone are adjusted in brightness and contrast and geometrically corrected to ensure comparability. A deep learning model (such as U-Net, DeepLabv3+, etc.) is applied to perform semantic segmentation on the construction images and 3D model views, and the images are divided into different categories (such as completed structures, parts under construction, and unconstructed areas). For the construction images and 3D model views at the same angle, the intersection area and union area of ​​the corresponding categories of the respective segmentation results are calculated to obtain the intersection over union ratio (IoU) as a quantitative indicator of the construction progress at that angle. The construction images acquired by the drone are IoU. real , the corresponding 3D model view is Imodel After semantic segmentation, the pixels of the same category (such as completed structure) in the two images are marked as S real and S model , the intersection over union (IoU) is defined as: IoU = |S real ∩S mod el ∣ / ∣S real ∪S model ∣, where ∣S real ∩S model ∣ represents the intersection area of ​​two types of pixel sets, ∣S re al ∪S model ∣ represents the union area. The IoU value range is between 0 and 1. The larger the value, the more consistent the segmentation results are, that is, the construction progress is closer to the expected. For a certain shooting angle, the construction progress P angle It can be estimated by combining the IoU value with the project volume W. The ratio of the actual completed project volume to the ideal completed project volume at this angle is r, then: r = (IoU × W) / W, so the construction progress percentage at this angle is: P angle =100×r=100×IoU, for multiple shooting angles, the weight of each angle is w j , the corresponding construction progress is P j ; Overall construction progress P total Calculated by weighted average: P total =∑ j＝1 N w j ×P j .

[0080] Use the Long Short-Term Memory (LSTM) network to predict the construction progress trend, collect and integrate future weather forecasts, resource status, attendance and other data to form a multivariate data set that affects the construction progress, and normalize or standardize the influencing factor data (weather conditions T, resource status R, attendance A) so that they have similar scales and distributions, which is conducive to model training. t =(T t ,R t ,A t) is converted into a time series format. Each moment t corresponds to a sample. Each sample contains the current moment and possible historical data (weather, resources, attendance in the past few days). The number of LSTM units is set, the input dimension is equal to the number of influencing factors, and the return sequence is set to True to capture long-term dependencies. A fully connected layer (Dense layer) is used with a linear activation function f(x) = wx + b to convert the output of the LSTM layer into a construction progress prediction value. The loss function is the mean square error MSE. The training set, validation set, and test set are divided. The model is trained using the training set data. The model parameters are optimized through back propagation and gradient update. The model performance is evaluated on the validation set. Hyperparameters (such as learning rate, number of LSTM units, etc.) are adjusted to prevent overfitting or underfitting. When the performance of the model on the validation set reaches a satisfactory level, training is stopped and the trained LSTM model is used to predict the construction progress on the future day t. The input is the normalized influencing factor data sequence D of the tth day and several days before. t-n:t-1 =(D t-n ,D t-n+1 ,,,,,D t-1 ), where n is the observation window length required by the model. Inside the model, the LSTM layer updates the hidden state according to the input sequence and generates a predicted value P^ through the output layer t :P^ t =LSTM(D t-n:t-1 ; θ), where θ represents all parameters of the LSTM model, and the denormalized prediction value P^ t , and obtain the construction progress forecast value on the original scale.

[0081] Compare the forecast results with the design construction schedule, assess whether there is a risk of failure to complete on schedule, and calculate the forecast construction progress P^ t Design and construction schedule plan,t Deviation: ΔP t =P^ t -P plan,t If |ΔP t ∣≥δ (where δ is the set threshold), it is considered that there is a progress deviation. When the prediction results show that there is a risk of progress delay, the reinforcement learning algorithm is enabled. Input historical project data, real-time resource status, weather forecast and other information, and learn and update the optimal task allocation strategy and construction sequence through interaction with the environment (i.e., the construction process), and dynamically adjust the project plan to minimize the construction period and reduce the delay cost. Specifically, the goal of the reinforcement learning algorithm is to maximize the expected cumulative reward Gt: G t =∑ k＝0 ∞ γ k R t+k+1 , where R tis the immediate reward obtained at time step t, γ is the discount factor (usually 0≤γ<1), and a policy π(a t ∣s t )(Given state s t Take action a t The probability of G is updated by interacting with the environment (i.e., the construction process) to maximize G t The update process is based on the Q-learning algorithm. In this scenario, the state s t Including historical project data, real-time resource status, weather forecast and other information, action a t Represents project plan adjustment decisions (such as task allocation, construction sequence changes, etc.).

[0082] In step S4, the identity of the entrants is confirmed through face recognition technology, and computer vision is used to identify whether they are wearing protective equipment correctly. The protective equipment (work clothes, safety helmets, etc.) is equipped with RFID radio frequency devices to ensure information consistency. Attendance is determined by combining face recognition and RFID radio frequency. The openpose model is used to periodically identify the posture of construction workers in high-risk areas to determine whether there are any dangerous behaviors.

[0083] In step S5, RFID tags are embedded on the material packaging. During the logistics process, RFID tags are scanned at key locations such as warehouses, transportation nodes, and construction site entrances through fixed or handheld reading and writing devices to record the location, timestamp and other information of the materials, which are uploaded to the central database in real time. An RFID access control system is set up at the entrance of the construction site to automatically record the time when the materials enter the site. Each work area on site is equipped with a handheld device or a fixed reading and writing point to scan the materials regularly or on demand to ensure that the whereabouts of the materials are clear and prevent loss or misuse.

[0084] The construction site is divided into multiple work areas, materials are regarded as nodes, and the flow of materials between different areas is regarded as directed edges. A directed graph model containing material nodes, location nodes and the transfer relationship between the two is constructed, and two types of nodes are defined, namely material nodes M and location nodes L. Each material node m∈M represents a specific material (such as concrete, steel, etc.), including attributes such as type, specification, demand, etc. Each location node l∈L represents a work area on the construction site, including attributes such as spatial distance, construction progress correlation, etc. Directed edges E are defined to represent the flow of materials between different locations. For material node m and location node l, if material m may flow from location l to another location l′, there is a directed edge (m, l′)∈E starting from m and pointing to l′. The weight of the edge (such as flow rate, frequency) can be used to represent the intensity or probability of the flow of materials between two locations.

[0085] The graph neural network GCN is used to embed nodes and extract high-dimensional features such as material properties (such as type, specification, and demand), location characteristics (such as spatial distance and construction progress correlation), and transfer rules between nodes to form a low-dimensional, dense vector representation. An initial node feature vector x is assigned to each node v. v , corresponding to its properties in the graph (such as material properties, position characteristics, etc.), the eigenvectors of all nodes are arranged into a matrix H (0) =[x1,x2,…,x N ] T , where N = |M| + |L| is the total number of nodes. Message passing and node updates are performed through multi-layer GCN to calculate the node feature matrix H of the lth layer (l) :H (l+1) =σ(D ～-1 / 2 A ～ D ～-1 / 2 H (l) W (l) ), where σ is the activation function (such as ReLU, tanh, etc.), A ～ =A+I N is an adjacency matrix with self-loops (A is the original adjacency matrix, I N is the unit matrix), D ～ ii =∑ j A ～ ij is the degree matrix, W (l) is the weight matrix of the lth layer. After training, the final hidden state of each node h v (K) (where K is the index of the last layer) is used as the embedding vector representation of the node. These low-dimensional, dense vectors contain the encoding of high-dimensional features such as material properties, position characteristics, and transfer rules between nodes.

[0086] Use deep reinforcement learning to explore and optimize directed graphs, embedding nodes into vectors h v (K) As the state representation of the reinforcement learning environment. For each node v where the agent needs to make a decision, its state is composed of its own embedding vector and the embedding vectors of its neighboring nodes (or their aggregate representations), defining the set of available node actions for the agent at each node, such as material management decisions such as procurement, storage, distribution, and use. Each node action a usually corresponds to a real-valued or discrete decision variable. A deep neural network (such as a fully connected network) is constructed with the input as the node state representation and the output as the probability distribution of each node action or the deterministic node action value. For the deterministic policy gradient method, the network directly outputs the deterministic node action value, a v =π(s v ; θ) = fpolicy (h v (K) ;θ), where π is the policy function, s v is the state of node v, θ is the parameter of the policy network, and f policy is the forward propagation function of the policy network, and the reward function can be expressed as R = w cost r cost +w transport r transport -w inventory r inventory -w outage r outage +w timeliness r timeliness +w u sage r usage , where w n is the corresponding reward component r n The weight of ∑w n =1, ensuring the same scale of reward value, procurement cost r cost : Give positive (savings) or negative (overspending) rewards based on the difference between the actual purchase price and the budget or market benchmark price, cost = (budget cost - actual purchase cost) / budget cost, transportation cost r transport : rewards are given based on the difference between actual delivery costs and estimated costs, encouraging optimized delivery routes and bulk shipments to reduce costs, transport = (estimated transportation cost - actual transportation cost) / estimated transportation cost, inventory holding cost r inventory :Punishment is given based on the gap between inventory level and target inventory level (excess or shortage), encouraging the maintenance of reasonable inventory level, reducing capital occupation and storage costs, inventory =-[(actual inventory-target inventory) / target inventory] 2 , out-of-stock penalty r outage :When material demand cannot be met in time and construction is stopped, serious negative rewards are given to emphasize the importance of avoiding construction progress affected by material shortages. outage ={-λ outage ; 0}, where λ outage is a preset positive value representing the severity of the shortage, the timeliness of material arrival tim eliness :Reward materials for being delivered to the construction site on time as planned, avoiding idle labor caused by waiting time, tim eliness ={1-(actual arrival time-planned arrival time) / maximum allowed delay time;-1}, material utilization efficiency r usage:Reward material consumption and project progress are well matched to avoid waste or delays caused by excessive stockpiling or frequent replenishment,r usage =(actual consumption / ideal consumption)·(completed engineering volume / planned engineering volume), the deterministic policy gradient method REINFORCE algorithm is used to update the policy network parameters θ, and the parameter update rule is: Among them, J(θ) is the expected cumulative reward, T is the number of sampling steps, and R t is the reward received at time step t. The agent performs node actions in the environment according to the deterministic node action value output by the current policy network, and observes the environmental feedback (new state and reward). This process is repeated, and experience is accumulated and policy network parameters are updated continuously until the preset number of training rounds or convergence conditions are reached. The agent can explore and optimize in the graph environment and learn the optimal material management strategy.

[0087] Please refer to Figure 2 , which shows a structural diagram of a construction project management system based on big data provided by an embodiment of the present application;

[0088] Data collection module: By installing cameras and deploying various sensors in the construction site, using drones for aerial inspections, and arranging inspection personnel for on-site inspections, the collected images and data are transmitted back to the control center in real time, providing rich original data for subsequent analysis;

[0089] Quality management module: The construction site is subdivided into multiple sub-areas, and data monitoring points are set for each sub-area. The detection data in the sub-area is collected and input into the machine learning model through pre-processing and feature extraction. The model outputs the quality level and risk degree of each sub-area to reflect the construction status. The evaluation results are accurately mapped to the corresponding 3D building model to form a visual quality map.

[0090] Progress management module: drones regularly take multi-angle construction images, compare them with the three-dimensional models constructed from the design drawings, perform semantic segmentation on the compared images, and determine the construction progress by calculating the intersection ratio. Integrate weather forecasts, resource status, attendance data and other information to predict the remaining construction time and potential delay risks. When predicting delays, use reinforcement learning algorithms combined with historical project data, real-time resource status, weather forecasts, etc. to dynamically adjust project plans and construction task allocations, and strive for the optimal construction schedule;

[0091] The safety management module uses facial recognition technology to confirm the identity of the entrants and computer vision to identify whether they are wearing protective equipment correctly to ensure compliance. It deploys cameras to monitor the construction site around the clock and at irregular intervals. It uses deep learning algorithms to analyze surveillance videos in real time, automatically identify potentially dangerous behaviors, trigger early warning systems, and prevent safety accidents.

[0092] Material management module: RFID tags are embedded in material packaging to track the entire chain of material flow from suppliers to construction sites in real time through wireless networks, ensuring that the source of materials can be traced and preventing the mixing of counterfeit and shoddy products. At the same time, the efficiency of material use is monitored to reduce waste. Materials and construction locations are regarded as directed graphs, and graph neural networks and deep reinforcement learning algorithms are used to perform intelligent management and optimized scheduling of materials.

Claims

1. A construction project management method based on big data, characterized in that: The method comprises: S1. Set up data collection devices and inspections at the construction site to collect real-time images and data and transmit them back to the central database; S2. Divide the construction site into different sub-areas, conduct quality monitoring for each area, determine the quality level of each sub-area, and mark the corresponding area in the 3D model; S3, using drones to obtain construction images, using semantic segmentation, calculating construction progress, and predicting remaining construction time. If a delay is predicted, using reinforcement learning algorithms to optimize project planning and task allocation; Reinforcement learning algorithms optimize project planning and task allocation: Calculate and predict construction progress P^ t Design and construction schedule plan，t Deviation: ΔP t =P^ t -P plan，t ; If |ΔP t ∣≥δ, δ is the set threshold, and the reinforcement learning algorithm is enabled; The goal of the reinforcement learning algorithm is to maximize the expected cumulative reward Gt: G t =∑ k=0 ∞ γ k R t+k+1 ; Among them, R t is the immediate reward obtained at time step t, R t+k+1 is the next immediate reward, k is the number of steps of immediate reward, γ is the discount factor, select a policy π, and update the policy to maximize G by interacting with the environment. t ; Status t Including historical project data, real-time resource status, weather forecast, action a t Represent project plan adjustment decisions; S4. Implement facial recognition and protective equipment recognition to enter the construction site, install radio frequency devices on protective equipment to ensure information consistency, and use cameras in the construction site to detect and identify dangerous behaviors from time to time; S5. Embed RFID tags on material packaging to track the entire chain flow, treat materials and construction locations as directed graphs, use graph neural networks for embedding learning, and optimize material scheduling plans through deep reinforcement learning; Consider materials and construction locations as a directed graph: The construction site is divided into multiple work areas, materials are regarded as nodes, and the flow of materials between different areas is regarded as directed edges; Construct a directed graph model including material nodes, location nodes and the transfer relationship between them, and define two types of nodes, namely material nodes M and location nodes L; Each material node m∈M represents a specific material, including attributes such as type, specification, and demand; Each location node l∈L represents a work area on the construction site, and contains attributes such as spatial distance and construction progress relevance; Define directed edges E to represent the flow of materials between different locations; For each material node m and location node l, if material m can flow from location l to another location l′, then there is a directed edge ∈E from m to l′; The weight of an edge can be used to represent the strength or probability of material flowing between two locations; Optimizing material scheduling through deep reinforcement learning: Embed the node into the final hidden state h v （K） As the state representation of the reinforcement learning environment, for each node v where the agent needs to make a decision, its state is composed of its own final hidden state and the final hidden state of its adjacent nodes; Define the set of available node actions for the agent at each node. Each node action a usually corresponds to a real-valued or discrete decision variable. Build a deep neural network with the input as the node state representation and the output as the probability distribution of each node action or the deterministic node action value. For the deterministic policy gradient method, the network directly outputs the deterministic node action value, a v =π(s v ;θ) = f policy (h v （K） ;θ), where π is the policy function, s v is the state of node v, θ is the parameter of the policy network, and f policy is the forward propagation function of the policy network; The reward function can be expressed as: R=w cost r cost +w transport r transport -w inventory r inventory -w outage r outage +w timeliness r timeliness +w usage r usage , where w n is the corresponding reward component r n The weight of the purchase cost r cost , transportation cost r transport , inventory holding cost r inventory , out-of-stock penalty r outage , Timely arrival of materials timeliness , material utilization efficiency r usage , Use the deterministic policy gradient method REINFORCE algorithm to update the policy network parameters θ. The parameter update rule is: ▽ θ J≈∑ t=1 T ▽ θ logπR t , where J is the expected cumulative reward, T is the number of sampling steps, and R t is the reward received at time step t; The agent executes node actions in the environment according to the deterministic node action values ​​output by the current policy network, observes environmental feedback, and repeats this process, continuously accumulating experience and updating policy network parameters until the preset number of training rounds or convergence conditions are reached.

2. A construction project management method based on big data according to claim 1, characterized in that: Determine the quality level of each sub-area: Calculate the qualified rate of each test in each sub-area and build a quality rating model based on neural network; Use existing labeled data to train the model, adjust model parameters through cross-validation and grid search methods, and optimize model performance; The features of the sub-region to be evaluated are input into the trained model, and the model outputs the quality level of the sub-region.

3. A construction project management method based on big data according to claim 1, characterized in that: Calculate construction progress: The drone performs flight missions according to the plan, captures construction images from different angles, and generates 3D model views of corresponding angles from the building information model; Process the construction images and 3D model views collected by drones, and apply semantic segmentation models to segment the construction images and 3D model views; For the construction image and 3D model view at the same angle, the intersection-and-union ratio corresponding to each segmentation result is calculated as a quantitative indicator of the construction progress at that angle; The construction image acquired by the drone is I real , the corresponding 3D model view is I model After semantic segmentation, the pixels of the same category in the two images are marked as S real and S model , the intersection over union ratio is defined as: IoU=|S real ∩S model ∣ / ∣S real ∪S model ∣; Among them, |S real ∩S model ∣ represents the intersection area of ​​two types of pixel sets, ∣S real ∪S model ∣ represents the area of ​​the union; For a certain shooting angle, the construction progress P angle It can be estimated by combining the intersection ratio with the engineering quantity W; the ratio of the actual completed engineering quantity to the ideal completed engineering quantity under this angle is r, then: r= (IoU×W) / W; The construction progress percentage of this angle is: P angle =100×r=100×IoU; For multiple shooting angles, the weight of each angle is w j , the corresponding construction progress is P j ; Overall construction progress P total Calculated by weighted average: P total =∑ j=1 Z w j ×P j .

4. A construction project management method based on big data according to claim 1, characterized in that: Estimated remaining construction time: Collect and integrate future weather forecasts, resource status, and attendance to form a multivariate data set that affects construction progress, and normalize or standardize the influencing factor data; The influencing factor data D t =(T t ,R t ,A t ) is converted to a time series format, the number of LSTM units, input dimension, and return sequence are set, and a fully connected layer is used with a linear activation function f=wx+b. The output of the LSTM layer is converted into a construction progress prediction value, and the loss function is the mean square error MSE; Train the model, optimize the model parameters through back propagation and gradient update, and use the trained LSTM model to predict the construction progress at the future time step t; The input is the normalized influencing factor data sequence D of time step t and several days before t-n:t-1 =(D t-n , D t-n+1 , ..., D t-1 ), where n is the observation window length required by the model; Generate predicted value P^ t :P^ t =LSTM(D t-n:t-1 ; φ), where φ represents all parameters of the LSTM model; Denormalized predicted value P^ t , and obtain the construction progress forecast value on the original scale.

5. The construction project management method based on big data according to claim 1 is characterized in that: Implement facial recognition and protective equipment recognition: The identity of the entrants is confirmed through facial recognition technology, and computer vision is used to identify whether they are wearing protective equipment correctly; The protective equipment is equipped with RFID radio frequency devices to ensure information consistency, and the attendance status is determined by integrating facial recognition and RFID radio frequency devices; Conduct posture recognition on construction workers in high-risk areas from time to time to determine whether they are engaging in dangerous behaviors.

6. A construction project management method based on big data according to claim 1, characterized in that: Embed RFID tags on material packaging: In the logistics process, RFID tags are scanned at warehouses, transportation nodes, and construction site entrances through fixed or handheld reading and writing devices, and uploaded to the central database in real time; An RFID access control system is set up at the entrance of the construction site to automatically record the time when materials enter the site; Each work area on site is equipped with handheld devices or fixed read-write points to scan materials regularly or on demand to ensure the whereabouts of materials and prevent loss or misuse.

7. The construction project management method based on big data according to claim 1 is characterized in that: Embedding Learning with Graph Neural Networks: Assign an initial node feature vector x to each node v v , corresponding to its attributes in the graph, the feature vectors of all nodes are arranged into a matrix H = [x1, x2, ..., x N ] T , where N = |M| + |L| is the total number of nodes; x N is the feature vector of the Nth node; Message passing and node update are performed through multi-layer GCN to calculate the node feature matrix H of the lth layer （l） :H （l+1） =σ(D ~-1 / 2 A ~ D ~-1 / 2 H （l） W （l） ), where σ is the activation function, A ~ =A+I N is the adjacency matrix with self-loops, A is the original adjacency matrix, I N is the identity matrix, D ~ ii =∑ j A ~ ij is the degree matrix, W （l） is the weight matrix of the lth layer; After training is completed, the final hidden state h of each node v （K） , where K is the index of the last layer, which is used as the embedding vector representation of the node. These low-dimensional, dense vectors contain material properties, position characteristics, and transfer rules between nodes.

Images