Construction progress dynamic optimization method and system based on BIM and computer vision

Abstract

The invention discloses a construction progress dynamic optimization method and system based on BIM and computer vision, and particularly relates to the technical field of building construction management, and the method comprises the steps: carrying out the automatic registration of a BIM model and a construction site image; processing the construction site image by adopting a visual identification algorithm to generate a visual identification result; constructing a four-dimensional dynamic BIM model, and mapping a visual identification result to a corresponding component in real time through multi-feature similarity calculation; the progress deviation is monitored by using key path dynamic identification and a deviation propagation matrix, and the risk is predicted by combining a Bayesian network and Monte Carlo simulation. The BIM and computer vision technologies are fused, a construction progress optimization system integrating automatic registration, dynamic monitoring, risk prediction and intelligent decision making is constructed, and the problems that traditional manual inspection data collection is low in efficiency, progress monitoring is lagged, risk prejudgment is fuzzy and resource allocation is extensive are solved; accurate monitoring, risk early warning and resource optimization configuration of the construction progress are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of building construction management, and more specifically, to a method and system for dynamically optimizing construction progress based on BIM and computer vision. Background Art

[0002] In the process of construction management, dynamic monitoring and optimization of construction progress has always been a difficult point in project management.

[0003] Traditional construction progress management methods mainly rely on manual inspections and experience-based judgments, and have problems such as low data collection efficiency, poor real-time performance, and large errors. It is difficult to accurately reflect the actual progress of the construction site, and it is impossible to respond to various risks and deviations in the construction process in a timely manner.

[0004] With the development of BIM (Building Information Modeling) technology and computer vision technology, how to effectively combine the two to achieve dynamic optimization of construction progress has become a technical problem that needs to be solved urgently in this field. Summary of the Invention

[0005] The present invention aims to provide a construction progress dynamic optimization method and system based on BIM and computer vision to solve the problems of low construction progress management efficiency, poor real-time performance, inaccurate risk prediction, etc. in the existing technology, and realize accurate monitoring, risk prediction and dynamic optimization of construction progress.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a construction progress dynamic optimization method based on BIM and computer vision, comprising the following steps: Step 1: Obtain BIM model point cloud data, collect construction site images, establish a unified spatial coordinate system, and automatically align the BIM model with the construction site images; Step 2: Use visual recognition algorithms to process construction site images, identify component status, equipment operating parameters, quality defects, and safety violations, and generate visual recognition results; Step 3: Build a four-dimensional dynamic BIM model and map the visual recognition results to the corresponding components in real time through multi-feature similarity calculation; Step 4: Use dynamic critical path identification and deviation propagation matrix to monitor progress deviations, and combine Bayesian networks and Monte Carlo simulation to predict risks; Step 5: Generate a dynamic optimization plan for construction sequence and resource allocation through a hierarchical reinforcement learning architecture.

[0007] Specifically, based on step 1, the BIM model and the construction site image are automatically registered. The automatic registration method is as follows: S11, Feature Acquisition: Use a laser scanner to sample the BIM model’s point cloud to obtain point cloud data containing 3D coordinates; use a binocular camera to capture RGB-D images of the construction site, extract depth information, and generate on-site point cloud data; S12, feature extraction: The FPFH algorithm is used to extract the fast point feature histogram descriptors of the point cloud data and the on-site point cloud data respectively, and the initial correspondence is established through bidirectional nearest neighbor search; the iterative nearest point algorithm is used to optimize the spatial transformation parameters, and the rotation matrix and translation vector to be optimized are calculated through the objective function. The BIM model point cloud and the on-site point cloud are aligned by minimizing the function.

[0008] Specifically, based on step 2, a visual recognition algorithm is used to process the construction site image. The visual recognition algorithm processing specifically includes: S21, multimodal feature acquisition: Use industrial cameras to capture RGB images of the construction site and extract channel features; simultaneously obtain the depth map output by the ZED camera and generate depth features through bilateral filtering preprocessing; S22, feature fusion processing: ResNet50 is used as the backbone network to extract the semantic features of the RGB image, the spatial distance information of the deep features is extracted through the 3D convolution layer, and the features are nonlinearly transformed through the fusion function; S23, spatiotemporal attention calculation: input the fused features of consecutive set frames into the gated recurrent unit and calculate the temporal attention weight of the t-th frame.

[0009] Specifically, the visual recognition algorithm further includes: Quality defect feature acquisition and processing: Use the U-Net network to perform semantic segmentation on RGB images. The input channels include three RGB channels and a single depth channel, and output a pixel-level defect probability map. The model parameters are trained using a defect type database, and the defect severity is calculated. Safety hazard feature acquisition and processing: Preset construction safety inspection standards, identify real-time construction safety features from construction site images based on the YOLOv8 object detection network, and generate a set number of anchor boxes through K-means clustering; Then, through spatiotemporal correlation analysis, determine whether there are any safety hazards.

[0010] Specifically, based on step three, the construction of the four-dimensional dynamic BIM model includes: S31, state feature acquisition: read the component's geometric completion data through the RFID tag, then collect construction quality indicators to generate a component quality state vector; extract the construction process type and start / end time from the construction log and record it as a construction operation vector; collect manpower data, equipment data, and material data from the construction site and record it as a resource input vector; S32, model construction processing: establish a state transfer function and use LSTM network training, the input is the component quality state vector, construction operation vector, and resource input vector within the set time zone, and the output is the state prediction at the next moment; establish a resource-progress association model through the construction condition vector, the preset resource efficiency function, the construction progress completion amount and the maximum construction progress completion amount.

[0011] Specifically, the visual recognition results are mapped to the corresponding components in real time through multi-feature similarity calculation. The real-time mapping process includes: Mapping feature acquisition: Extract the geometric features, texture features, and material features of components from the visual recognition results, and extract the corresponding attribute parameters from the corresponding components in the BIM model; Mapping processing: Set corresponding weight coefficients for geometric features, texture features, and similarity material features respectively; construct a similarity function for multi-feature fusion, and the fused feature dimensions include geometric features, texture features, and material features; calculate geometric feature similarity through Euclidean distance, calculate texture feature similarity through LBP texture histogram algorithm, and calculate material feature similarity through HSV color space analysis; set corresponding weight coefficients for geometric features, texture features, and similarity material features respectively; perform weighted fusion of geometric features, texture features, and similarity material features with their set weight coefficients to obtain comprehensive similarity; comprehensive similarity is used to measure the degree of match between visual recognition results and BIM component features; Based on the similarity of multi-feature fusion, the mapping confidence is calculated. Specifically, the confidence calculation covers the dimensions corresponding to geometry, texture, and material. The final confidence is obtained by averaging the feature similarities after processing the feature extraction function of each dimension. A confidence threshold is set. When the calculated confidence is lower than the confidence threshold, it triggers the BIM engineer to perform manual verification through the WebGL interactive interface.

[0012] Specifically, the progress deviation is monitored using the dynamic identification of the critical path and the deviation propagation matrix. The monitoring method of the progress deviation is as follows: Acquiring progress characteristics: The duration data of each construction activity is obtained through the task list of the BIM model, and the actual start and end times of the construction activities are identified through on-site images. In the dynamic identification of critical paths, the earliest possible start time for any construction activity is determined by adding the maximum value of the earliest possible start time of all the preceding construction activities and its own duration. For any construction activity, its latest required start time is determined by subtracting the minimum of its own duration from the latest required start time of all subsequent construction activities; Deviation processing: Construct a deviation propagation matrix: The values ​​of the elements in the matrix are determined according to the logical relationship of the construction process: if a process is the immediate predecessor of another process, the corresponding matrix element value is set to 1; if there is a resource sharing relationship between the two processes, the corresponding matrix element value is set to 0.5; if there is no such relationship, the corresponding matrix element value is set to 0; By comparing the actual construction progress with the planned progress, the delay time of each process is counted and the delay vector of the current process is calculated; The deviation propagation matrix is ​​operated on the delay vector of the current process to calculate the execution deviation propagation, and the delay time distribution results of the affected subsequent processes are output to achieve dynamic monitoring and propagation simulation of construction progress deviations.

[0013] Specifically, based on step 4, we combine Bayesian networks with Monte Carlo simulation to predict risk, as follows: Risk feature acquisition: Extract historical construction data from the project management system; use a set number of historical risk event data sets, train and learn using the K2 algorithm, and generate a conditional probability table for the Bayesian network; construct a Bayesian network based on the conditional probability table; Risk handling: Set the number of simulations. During each simulation, generate risk scenarios based on the Bayesian network and calculate the corresponding risk values. Set the risk threshold. Use the indicator function to determine whether the risk scenario in each simulation triggers a risk. Count the proportion of risk events in all simulation results as the probability of construction risk occurrence. Generate and output a set number of high-probability risk scenarios.

[0014] Specifically, based on step 5, a dynamic optimization plan for construction sequence and resource allocation is generated through a hierarchical reinforcement learning architecture, which includes: Decision feature acquisition: Divide decision features into upper-level features and lower-level features. Upper-level features include the current construction stage and critical path activity status, while lower-level features include resource requirements for each process and material inventory levels. A reward function is also pre-set. Algorithm processing: In the hierarchical decision-making architecture, the upper-level strategy uses a convolutional neural network, which takes as input the construction stage feature map and outputs the construction stage transition probability; the lower-level strategy uses a fully connected network, which takes as input the resource feature vector and outputs the resource allocation plan; Actor network: Designed as a three-layer fully connected layer structure, with 256 neurons in each layer to output decision action probabilities; Critic network: uses value function approximation method to evaluate the value of the current decision; Training process: Set the learning rate and discount factor, and train multiple rounds through the PPO algorithm until the fluctuation of the cumulative reward value meets the preset fluctuation threshold. During training, the parameters of the Actor and Critic networks are dynamically updated based on the current construction status, decision actions, and reward feedback to optimize the construction sequence and resource allocation strategy.

[0015] The present invention also proposes a system using the above method, comprising the following modules: Data acquisition module, used to obtain BIM model point cloud data and construction site RGB-D images to establish a unified spatial coordinate system; A visual recognition module is used to process construction site images using visual recognition algorithms to identify component status, equipment operating parameters, quality defects, and safety violations, and generate visual recognition results; The four-dimensional BIM modeling module is used to build a four-dimensional dynamic BIM model and map the visual recognition results to the corresponding components in real time through multi-feature similarity calculation; The progress deviation monitoring module is used to monitor construction progress deviations in real time and generate delay time distribution through the dynamic identification algorithm of critical paths and the deviation propagation matrix; Risk prediction module, which integrates Bayesian networks and Monte Carlo simulation algorithms to predict construction risks based on historical risk data and real-time environmental parameters; A dynamic optimization decision module is used to generate dynamic optimization plans for construction sequence and resource allocation through a hierarchical reinforcement learning architecture.

[0016] Technical effects and advantages of the present invention: 1. This invention integrates BIM and computer vision technology, realizes automatic registration of models and on-site images through laser scanning and visual recognition, and obtains data such as component status and quality defects in real time, solving the problems of low efficiency and large errors in traditional manual inspection data collection.

[0017] 2. This invention builds a critical path dynamic monitoring and Bayesian network risk prediction system, combines the deviation propagation matrix with Monte Carlo simulation, and analyzes progress deviations in real time and quantifies risk probabilities, thus overcoming the limitations of traditional methods such as poor real-time performance and inaccurate risk prediction.

[0018] 3. The present invention relies on a hierarchical reinforcement learning architecture and a four-dimensional BIM model to dynamically generate construction sequence and resource allocation optimization plans, achieve accurate quantification of the relationship between resource input and progress, and solve the problems of extensive resource allocation and lagging progress optimization in traditional management.

[0019] In summary, the present invention integrates BIM and computer vision technology to build a construction progress optimization system that integrates automatic alignment, dynamic monitoring, risk prediction and intelligent decision-making, solves the problems of inefficient data collection of traditional manual inspections, delayed progress monitoring, vague risk prediction and extensive resource allocation, and realizes accurate monitoring of construction progress, risk warning and optimal resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flowchart of the construction progress dynamic optimization method based on BIM and computer vision proposed in this invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] like Figure 1 As shown in the figure, the construction progress dynamic optimization method based on BIM and computer vision includes the following steps: Step 1: Use a laser scanner to obtain BIM model point cloud data, use a binocular camera to capture construction site images, establish a unified spatial coordinate system, and automatically align the BIM model with the construction site images; Step 2: Use visual recognition algorithms to process construction site images, identify component status, equipment operating parameters, quality defects, and safety violations, and generate visual recognition results; Step 3: Build a four-dimensional dynamic BIM model and map the visual recognition results to the corresponding components in real time through multi-feature similarity calculation; Step 4: Use dynamic critical path identification and deviation propagation matrix to monitor progress deviations, and combine Bayesian networks and Monte Carlo simulation to predict risks; Step 5: Generate a dynamic optimization plan for construction sequence and resource allocation through a hierarchical reinforcement learning architecture.

[0023] In this application, based on step 1, the BIM model and the construction site image are automatically aligned. The automatic alignment method is as follows: S11, feature acquisition: Use FAROS70 laser scanner to perform point cloud sampling on the BIM model and obtain point cloud data containing three-dimensional coordinates, which is recorded as , n represents the number of points in the BIM model point cloud data P, and the three-dimensional coordinate information of the i-th point in the BIM model point cloud data is recorded as ; Use the ZED2i binocular camera to collect RGB-D images of the construction site, extract depth information and generate on-site point cloud data, recorded as , m represents the number of points in the on-site point cloud data, and the three-dimensional coordinate information of the j-th point in the on-site point cloud data is recorded as ; S12, feature extraction: The FPFH algorithm is used to extract the fast point feature histogram descriptors FP and FQ of the point cloud data P and the field point cloud data Q respectively, and the initial correspondence is established through bidirectional nearest neighbor search. ,in yes The nearest neighbor in FQ, and yes Nearest neighbor in FP; The spatial transformation parameters are optimized using the iterative closest point algorithm, and the objective function Calculate the rotation matrix to be optimized and the translation vector t, the BIM model point cloud is aligned with the on-site point cloud by minimizing the function; where the weight matrix , 、 Calculate the point cloud curvature by local surface fitting, Indicates the influence of the i-th point pair in calculating the optimal rotation matrix and translation vector, Obtained by estimating the normal vector of the point neighborhood, 、 、 It is an adjustable parameter used to adjust the weight distribution of point cloud registration.

[0024] It should be noted that this automatic registration method first uses the FAROS70 laser scanner and the ZED2i binocular camera to obtain the BIM model point cloud data and the on-site point cloud data respectively; then uses the FPFH algorithm to extract features and the bidirectional nearest neighbor search to establish the initial correspondence; finally, using the iterative nearest point algorithm, based on the weighted objective function including curvature and normal vector, the rotation matrix and translation vector are optimized to achieve point cloud registration, laying a data foundation for the subsequent construction progress correlation analysis.

[0025] In this application, based on step 2, a visual recognition algorithm is used to process the construction site image. The visual recognition algorithm processing specifically includes: S21, Multimodal Feature Acquisition: Using the Basler acA2500-14gc industrial camera to capture RGB images of the construction site and extract channel features ; Synchronously obtain the depth map output by the ZED camera and generate depth features through bilateral filtering preprocessing ; Where H is the image height, W is the image width, Represents the channel feature matrix of the RGB image, Represents the deep feature matrix; S22, feature fusion processing: ResNet50 is used as the backbone network to extract the semantic features of RGB images, and the spatial distance information of the depth features is extracted through the 3D convolution layer. ,in 、 are the weight matrices of feature fusion, b1 represents the bias vector used to adjust the overall offset when calculating the semantic features and depth features of the RGB image, is an activation function used to perform nonlinear transformation on the fused features; S23, spatiotemporal attention calculation: set the fusion features of T frames continuously Input gated recurrent unit and pass Calculate the temporal attention weight of the t-th frame ;in is the hidden state at the previous moment, W represents the hidden state at the previous moment The weight matrix for linear transformation, U represents the weight matrix used to fuse the features of the current frame The weight matrix for linear transformation, b2 is the bias vector for offset adjustment of the linear transformation result when calculating the temporal attention weight by fusion features of the previous hidden state and the current frame.

[0026] It should be noted that, through the visual recognition algorithm processing flow in this application, the Basler industrial camera and ZED camera are used to collect RGB images and depth maps of the construction site respectively, and multimodal features are obtained through preprocessing; then the ResNet50 backbone network and 3D convolution layer are used to extract semantic and spatial distance information, and feature fusion is achieved through a fusion function containing a weight matrix, a bias vector and an activation function; finally, the continuous T-frame fusion features are input into the gated recurrent unit, combined with the hidden state of the previous moment and the current frame features, and the temporal attention weight is calculated through a specific formula to complete the full-process visual recognition processing of the construction site image from multimodal feature acquisition, fusion to spatiotemporal attention calculation, providing accurate image feature support for subsequent construction progress analysis.

[0027] In this application, the visual recognition algorithm also includes: Quality defect feature acquisition and processing: Use the U-Net network to perform semantic segmentation on RGB images. The input channels include RGB three channels and a single depth channel, and the output is a pixel-level defect probability map. ; The model parameters are trained by the defect type database (including 12 types of defect samples such as cracks, exposed steel bars, honeycombs, etc.), and the calculation formula of defect severity D is , where N represents the total number of defective pixels, It represents the defect probability of the i1th defective pixel, that is, the possibility that the pixel is a quality defect. Its value is in the range of [0,1]. The closer the value is to 1, the higher the probability of defect. Denotes the defect weight of the i-th defective pixel. This defect weight is preset based on the defect location (for example, load-bearing locations such as beam bottoms and column corners have higher weight coefficients, while non-load-bearing locations have lower weight coefficients) and defect type (for example, different crack widths have different weights). It is used to reflect the different impacts of defects of different locations and types on the overall quality. (The defect severity is a comprehensive indicator calculated from the information related to the defect pixel, used to quantitatively assess the severity of the quality defect). Acquisition and processing of safety hazard characteristics: preset construction safety inspection standards, and record the construction rules in the construction safety monitoring standards as (This is a pre-set basis for judging safety, such as "Construction workers must wear safety helmets" and "Scaffolding must meet spacing requirements.") Using the YOLOv8 object detection network, it identifies real-time construction safety features from construction site images and generates a set number of anchor boxes through K-means clustering. Then pass Realize spatiotemporal correlation analysis, where Indicates the result of the safety hazard judgment, which is a Boolean value. If the calculation result is true (which can be understood as 1), it is determined that there is a security violation; if it is false (which can be understood as 0), it is determined that there is no security violation; Represents the matching function, which is used to judge the construction safety features identified in real time , whether it is consistent with the construction rules in the construction safety inspection standards If it violates, it returns true, otherwise it returns false; Represents a set of real-time construction safety features obtained from construction site image recognition, including target categories (such as "personnel without safety helmets" and "illegally erected scaffolding"), target locations (location information in the image or construction site space), and duration (the number of image frames in which the safety feature appears continuously). Represents a method for calculating real-time construction safety features A function of the number of frames in which the security feature persists, i.e., the duration for which the security feature exists in consecutive image frames; Indicates setting the number of frames as a time threshold for determining security violations. The calculated duration frame is greater than When , it means that the security feature may constitute a security violation (needs to be determined in conjunction with the Match result) to avoid misjudging security violations due to temporary misidentification and other situations; ∧ represents the logical AND operator, which is only valid when Returns true (there is a rule violation) and (Continuous frame rate exceeds the set frame rate threshold) is established at the same time, then It is determined that there is a security violation (true). If any of the conditions are not met, then It is determined that there is no security violation (false).

[0028] It should be noted that by further expanding the visual recognition algorithm, multimodal features are first collected and preprocessed through Basler industrial cameras and ZED cameras, and feature fusion is achieved by using ResNet50, 3D convolutional layers and fusion functions. The gated recurrent unit is then input to combine the hidden state and the current frame features to calculate the temporal attention weight to complete the basic visual recognition of the image. At the same time, the identification of quality defects and safety hazards is also expanded. The U-Net network semantic segmentation RGB image is used to output the defect probability map and calculate the defect severity according to the preset weights. The safety features are identified based on the YOLOv8 target detection network. Combined with the K-means clustering anchor frame and spatiotemporal correlation analysis logic (matching safety rules and determining the number of continuous frames), quality defects and safety hazards are accurately identified, providing comprehensive and accurate image features and risk information support for construction progress optimization and construction management.

[0029] In this application, based on step three, the construction of the four-dimensional dynamic BIM model specifically includes: S31, state feature acquisition: read the geometric completion data of the component through the RFID tag, and then collect the construction quality indicators (such as concrete strength through the rebound hammer), and generate the component quality state vector recorded as ; Extract the construction process type and start / end time from the construction log and record it as a construction operation vector By collecting manpower data (such as the number of team members through the facial recognition attendance system), equipment data (such as tower crane operating parameters through GPS positioning), and material data (such as cement / steel usage through weighing on the construction site), it is recorded as a resource input vector. ; S32, model building process: establishing state transfer function , and adopts LSTM network training, with the input being the component quality state vector, construction operation vector, and resource input vector within the set time zone, and the output being the state prediction at the next moment; Establish resource-progress association model, model expression: , is the construction condition vector, and environmental data such as precipitation and temperature are obtained through the weather station. g is the resource efficiency function, which can be obtained by fitting the historical resource-progress data of a set number of groups through the XGBoost algorithm. Indicates the construction progress completion amount, Indicates the maximum completion amount of construction progress.

[0030] It should be noted that by constructing a four-dimensional dynamic BIM model, three types of vectors are first obtained through RFID tags, construction logs, facial recognition, GPS positioning, weighing scales, etc.: component quality status (geometric completion, concrete strength, etc.), construction operations (process type and time), and resource input (manpower, equipment, and material data); then the LSTM network is used to train the state transfer function, and the state at the next moment is predicted based on the three types of vector inputs within the set time range; at the same time, a resource-progress association model is constructed, and construction condition vectors (environmental data such as precipitation and temperature) are introduced. The resource efficiency function is fitted based on the XGBoost algorithm, and the relationship between resource input and construction progress is quantified by combining the construction progress completion amount and the maximum completion amount, providing model support for the dynamic optimization of the construction progress; The four-dimensional dynamic BIM model of the present invention incorporates dynamic construction data in the time dimension. Through an automatic registration process, it maps progress and on-site parameters in real time, dynamically simulates the construction process, and realizes the four-dimensional association of space, time, and dynamic data. It is expanded from a design model to a dynamic construction management model, which is different from the traditional static model.

[0031] In this application, the visual recognition results are mapped to the corresponding components in real time through multi-feature similarity calculation. The real-time mapping process includes: Mapping feature acquisition: Extract the component's geometric features (such as length and angle), texture features (such as surface roughness), and material features (such as the color threshold of concrete / steel) from the visual recognition result V, and extract the corresponding attribute parameters from the corresponding component in the BIM model; Mapping processing: Set the corresponding weight coefficients for geometric features, texture features, and similarity material features as q1, q2, and q3 respectively; construct a similarity function for multi-feature fusion, and perform weighted fusion on the geometric features, texture features, and similarity material features with the weight coefficients set to obtain the comprehensive similarity ,in is the geometric feature similarity, is the texture feature similarity, is the material feature similarity; Calculate mapping confidence based on the similarity of multi-feature fusion , its calculation formula In the example, n1=3 (geometry / texture / material dimensions), represents the m1th feature extraction function, represents the i2th component feature in the visual recognition result, Represents the j2th component feature in the BIM component; set a confidence threshold δ. If the confidence level is lower than the confidence threshold δ, it triggers manual verification by the BIM engineer through the WebGL interactive interface.

[0032] It should be noted that, through the real-time mapping process proposed in this application, three types of features, namely geometry, texture, and material, are first extracted from the visual recognition results and the corresponding components of the BIM model; then a weight coefficient is set for each feature, a similarity function of multi-feature fusion is constructed, and the weighted comprehensive similarity is calculated to measure the degree of matching between the visual recognition components and the BIM components; then, the mapping confidence is calculated based on the comprehensive similarity, and a confidence threshold is set. When the confidence is lower than the threshold, it triggers manual verification by the BIM engineer through the WebGL interactive interface, so as to achieve accurate and reliable mapping of the visual recognition results to the BIM components, and provide accurate component association data support for the dynamic optimization of the construction progress.

[0033] In this application, the progress deviation is monitored by using the dynamic identification of critical paths and the deviation propagation matrix. The progress deviation monitoring method is as follows: Acquisition of progress characteristics: obtain activity duration D(i3) through the task list of the BIM model, and obtain actual start / end time through on-site image recognition; in the dynamic identification of critical paths, the earliest start time , latest start time in is the set of predecessor processes of activity number j3, The set of subsequent processes of activity number i3; Deviation processing: propagate the elements of the deviation matrix A Determined by the logical relationship of the processes: If process i3 is the immediate predecessor of process j3, then ; If there is a resource sharing relationship, then ;otherwise The current process delay vector Δt is calculated by the time difference between the actual progress and the planned progress, and the deviation propagation calculation formula is Output the delay time distribution of subsequent processes , realizing progress deviation propagation analysis.

[0034] It should be noted that the construction progress deviation monitoring proposed in this application first obtains the activity duration from the BIM model task list, combines the actual start / end time of on-site image recognition, and uses the critical path method to calculate the earliest start time and the latest start time of the activity through the set of predecessor processes and the set of successor processes, so as to clarify the construction critical path; then constructs the deviation propagation matrix, determines the matrix elements according to the logical relationship of the processes (predecessor processes, resource sharing, etc.), combines the current process delay vector (the difference between actual and planned progress time), and outputs the delay time distribution of subsequent processes through matrix operations, so as to realize dynamic identification and propagation analysis of progress deviations and provide data support for optimization and adjustment of construction progress.

[0035] In this application, based on step 4, we use dynamic identification of critical paths and the deviation propagation matrix to monitor progress deviations, and combine Bayesian networks and Monte Carlo simulation to predict risks. Specifically: Risk feature acquisition: Extract historical progress deviation data from the project management system. Specifically, this includes records of material shortages such as steel and concrete from the material management system and the probability of precipitation during the construction period from the meteorological database. This includes variables such as progress deviation (discretized into high, medium, and low states), material shortages (binary states of yes and no), and precipitation (binary states of yes and no). Using a set number of groups of historical risk event data, the K² algorithm is used to learn the conditional probability table P of the Bayesian network, thereby forming a Bayesian network BN=(G,P) to characterize the correlation between risk variables. Risk management: Execute Monte Carlo simulation and set the number of simulations; generate risk scenarios (i.e., combinations of risk variable states) based on the Bayesian network (BN) and calculate the corresponding risk value (quantified in days of schedule delay); set the risk threshold (based on the proportion of the contract duration or the absolute number of days), and use the indicator function I to determine whether the risk scenario triggers a risk (I = 1 if and only if the risk value of the i4th simulation is , otherwise I=0), calculate the probability of risk occurrence based on all simulation results , and its calculation formula is ;in The preset risk threshold is a critical value set in advance to determine whether a risk has occurred. For example, it can be a standard of schedule delay days determined based on the contract duration. When the risk value obtained by simulation reaches or exceeds the risk threshold, it is determined that a risk has occurred. Refers to the risk value obtained from the i4th Monte Carlo simulation; It also generates and outputs a set number of high-probability risk scenarios (such as the probability that the combination of "precipitation + steel shortage" will cause a schedule delay of ≥10 days) to achieve construction risk prediction.

[0036] It should be noted that based on the progress deviation monitoring in step four, risk-related data such as historical progress deviation, material shortage, and precipitation are first extracted from multiple channels of the project management system (material management, meteorological database, etc.). After discretization processing, the K2 algorithm is used to learn and construct a Bayesian network (BN) to characterize the correlation between risk variables; then a Monte Carlo simulation is performed to generate risk scenarios based on BN and quantify the risk value (number of days of progress delay), set the risk threshold, determine the risk trigger through the indicator function, calculate the probability of risk occurrence through statistical simulation results, and output high-probability risk scenarios (such as the case of progress delay caused by "precipitation + steel shortage") to achieve accurate prediction of construction risks and provide risk warning support for progress optimization decisions.

[0037] In this application, based on step 5, a dynamic optimization scheme for construction sequence and resource allocation is generated through a hierarchical reinforcement learning architecture, wherein the hierarchical reinforcement learning architecture includes: Decision feature acquisition: This is divided into upper-level and lower-level decision features. Upper-level decision features include the current construction stage (foundation / main structure / decoration) and the status of critical path activities, while lower-level decision features include resource requirements for each process (manpower shortage, equipment idle rate), and material inventory levels. A reward function is also pre-set, with specific examples including a +10 reward for progressing one day ahead of schedule, a -20 penalty for a one-day delay, and a +5 reward for every 1% reduction in resource waste rate. Algorithm processing: In the hierarchical decision-making architecture, the upper-layer strategy uses a CNN network (with the input being the stage feature map) to output the stage transition probability, and the lower-layer strategy uses a fully connected network (with the input being the resource feature vector) to output the resource allocation plan; the Actor network structure is a 3-layer fully connected layer (with 256 neurons per layer), and the Critic network uses the value function approximation method; the parameter update formula is Learning rate , discount factor γ = 0.95, and set the number of rounds through PPO algorithm training until the cumulative reward fluctuates, where is the Actor network parameter, is the Critic network parameter, is the decision action, s is the current state, is the next state, r is the reward value, π is the policy function, and V() is the value function.

[0038] It should be noted that, based on step five, a hierarchical reinforcement learning architecture is constructed to generate construction sequences and resource allocation plans. Decision features are first extracted hierarchically, with the upper layer focusing on construction stages and critical path status, and the lower layer focusing on process resource requirements and material inventory. A preset reward function is used to reward and punish based on schedule advance / delay and resource waste rate optimization. In terms of algorithmic processing, the upper layer uses a CNN network to output stage transition probabilities, while the lower layer uses a fully connected network to output resource allocation. The Actor-Critic network collaborates and is trained using the PPO algorithm. It dynamically updates parameters based on status, action, and reward, iteratively optimizing strategies to achieve dynamic adaptation and optimization of construction decisions, facilitating efficient coordination of construction progress and resource management.

[0039] In this application, a system using the above method is also proposed, comprising: Data acquisition module, used to obtain BIM model point cloud data and construction site RGB-D images to establish a unified spatial coordinate system; A visual recognition module is used to process construction site images using visual recognition algorithms to identify component status, equipment operating parameters, quality defects, and safety violations, and generate visual recognition results; The four-dimensional BIM modeling module is used to build a four-dimensional dynamic BIM model and map the visual recognition results to the corresponding components in real time through multi-feature similarity calculation; The progress deviation monitoring module is used to monitor construction progress deviations in real time and generate delay time distribution through the dynamic identification algorithm of critical paths and the deviation propagation matrix; Risk prediction module, which integrates Bayesian networks and Monte Carlo simulation algorithms to predict construction risks based on historical risk data and real-time environmental parameters; A dynamic optimization decision module is used to generate dynamic optimization plans for construction sequence and resource allocation through a hierarchical reinforcement learning architecture.

[0040] The above formulas all use dimensionless numerical calculations, which can be achieved through standardization and other means, and will not be detailed here. The formulas are derived from software simulations based on a large amount of measured data, and the preset parameters are configured by those skilled in the art based on actual working conditions.

[0041] This embodiment can be implemented through software, hardware, firmware, or a combination thereof. When implemented in software, it can be integrated into a computer program product containing computer instructions or programs that, when loaded and executed, implement the processes and functions of this application. The computer can be a general-purpose or dedicated device or a network device, and the instructions can be stored in a computer-readable medium (such as a magnetic, optical, or semiconductor medium) and transmitted via wired or wireless means.

[0042] It should be noted that the sequence numbers of the steps do not represent the order of execution, which is determined by the functional logic. Those skilled in the art will appreciate that each unit and algorithm step can be implemented through electronic hardware or a combination of hardware and software, depending on the application scenario and design constraints. The systems and devices disclosed in this application can also be implemented through other partitioning methods, and the functional units can be integrated or independent.

[0043] If the function is implemented and sold in the form of a software unit, it can be stored in a computer-readable storage medium such as a USB flash drive, hard disk, ROM, RAM, etc. The medium contains instructions to drive the device to execute the steps of the method of this application.

[0044] It should be understood by those skilled in the art that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware or a combination of computer software and electronic hardware. Whether a specific function is executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can choose different methods to implement the described functions for each specific application, and such implementation should not be considered to exceed the scope of protection of this application.

Claims

1. A construction progress dynamic optimization method based on BIM and computer vision, characterized by: The steps include: Step 1: Obtain BIM model point cloud data, collect construction site images, establish a unified spatial coordinate system, and automatically align the BIM model with the construction site images; Step 2: Use visual recognition algorithms to process construction site images, identify component status, equipment operating parameters, quality defects, and safety violations, and generate visual recognition results; Step 3: Build a four-dimensional dynamic BIM model and map the visual recognition results to the corresponding components in real time through multi-feature similarity calculation; Step 4: Use dynamic critical path identification and deviation propagation matrix to monitor progress deviations, and combine Bayesian networks and Monte Carlo simulation to predict risks; Step 5: Generate a dynamic optimization plan for construction sequence and resource allocation through a hierarchical reinforcement learning architecture.

2. The construction progress dynamic optimization method based on BIM and computer vision according to claim 1 is characterized in that: Based on step 1, the BIM model and the construction site image are automatically registered. The automatic registration method is as follows: S11, Feature Acquisition: Use a laser scanner to sample the BIM model’s point cloud to obtain point cloud data containing 3D coordinates; use a binocular camera to capture RGB-D images of the construction site, extract depth information, and generate on-site point cloud data; S12, feature extraction: The FPFH algorithm is used to extract the fast point feature histogram descriptors of the point cloud data and the on-site point cloud data respectively, and the initial correspondence is established through the bidirectional nearest neighbor search; The iterative closest point algorithm is used to optimize the spatial transformation parameters, and the rotation matrix and translation vector to be optimized are calculated through the objective function. The BIM model point cloud and the on-site point cloud are aligned by minimizing the function.

3. The construction progress dynamic optimization method based on BIM and computer vision according to claim 2 is characterized in that: Based on step 2, a visual recognition algorithm is used to process the construction site image. The visual recognition algorithm processing specifically includes: S21, multimodal feature acquisition: Use industrial cameras to capture RGB images of the construction site and extract channel features; simultaneously obtain the depth map output by the ZED camera and generate depth features through bilateral filtering preprocessing; S22, feature fusion processing: ResNet50 is used as the backbone network to extract the semantic features of the RGB image, the spatial distance information of the deep features is extracted through the 3D convolution layer, and the features are nonlinearly transformed through the fusion function; S23, spatiotemporal attention calculation: input the fused features of consecutive set frames into the gated recurrent unit and calculate the temporal attention weight of the t-th frame.

4. The construction progress dynamic optimization method based on BIM and computer vision according to claim 3 is characterized in that: The visual recognition algorithm also includes: Quality defect feature acquisition and processing: Use the U-Net network to perform semantic segmentation on RGB images. The input channels include three RGB channels and a single depth channel, and output a pixel-level defect probability map. The model parameters are trained using a defect type database, and the defect severity is calculated. Safety hazard feature acquisition and processing: Preset construction safety inspection standards, identify real-time construction safety features from construction site images based on the YOLOv8 object detection network, and generate a set number of anchor boxes through K-means clustering; Then, through spatiotemporal correlation analysis, determine whether there are any safety hazards.

5. The construction progress dynamic optimization method based on BIM and computer vision according to claim 1 is characterized in that: Based on step three, the construction of the four-dimensional dynamic BIM model specifically includes: S31, state feature acquisition: read the component's geometric completion data through the RFID tag, then collect construction quality indicators to generate a component quality state vector; extract the construction process type and start / end time from the construction log and record it as a construction operation vector; collect manpower data, equipment data, and material data from the construction site and record it as a resource input vector; S32, model construction processing: establish a state transfer function and use LSTM network training, the input is the component quality state vector, construction operation vector, and resource input vector within the set time zone, and the output is the state prediction at the next moment; establish a resource-progress association model through the construction condition vector, the preset resource efficiency function, the construction progress completion amount and the maximum construction progress completion amount.

6. The construction progress dynamic optimization method based on BIM and computer vision according to claim 1 is characterized in that: The visual recognition results are mapped to the corresponding components in real time through multi-feature similarity calculation. include: Mapping feature acquisition: extracting the geometric features, texture features, and material features of components from the visual recognition results, and extracting the corresponding attribute parameters from the corresponding components in the BIM model; Mapping processing: Set corresponding weight coefficients for geometric features, texture features, and similarity material features respectively; construct a similarity function for multi-feature fusion, and the fused feature dimensions include geometric features, texture features, and material features; calculate geometric feature similarity through Euclidean distance, calculate texture feature similarity through LBP texture histogram algorithm, and calculate material feature similarity through HSV color space analysis; set corresponding weight coefficients for geometric features, texture features, and similarity material features respectively; perform weighted fusion of geometric features, texture features, and similarity material features with their set weight coefficients to obtain comprehensive similarity; comprehensive similarity is used to measure the degree of match between visual recognition results and BIM component features; Based on the similarity of multi-feature fusion, the mapping confidence is calculated. Specifically, the confidence calculation covers the dimensions corresponding to geometry, texture, and material. The final confidence is obtained by averaging the feature similarities after processing the feature extraction function of each dimension. A confidence threshold is set. When the calculated confidence is lower than the confidence threshold, it triggers the BIM engineer to perform manual verification through the WebGL interactive interface.

7. The construction progress dynamic optimization method based on BIM and computer vision according to claim 1 is characterized in that: The progress deviation is monitored using the dynamic identification of critical paths and the deviation propagation matrix. The monitoring method for progress deviation is as follows: Acquiring progress characteristics: The duration data of each construction activity is obtained through the task list of the BIM model, and the actual start and end times of the construction activities are identified through on-site images. In the dynamic identification of critical paths, the earliest possible start time for any construction activity is determined by adding the maximum value of the earliest possible start time of all the preceding construction activities and its own duration. For any construction activity, its latest required start time is determined by subtracting the minimum of its own duration from the latest required start time of all subsequent construction activities; Deviation processing: Construct a deviation propagation matrix: The values ​​of the elements in the matrix are determined according to the logical relationship of the construction process: if a process is the immediate predecessor of another process, the corresponding matrix element value is set to 1; if there is a resource sharing relationship between the two processes, the corresponding matrix element value is set to 0.5; if there is no such relationship, the corresponding matrix element value is set to 0; By comparing the actual construction progress with the planned progress, the delay time of each process is counted and the delay vector of the current process is calculated; The deviation propagation matrix is ​​operated on the delay vector of the current process to calculate the execution deviation propagation, and the delay time distribution results of the affected subsequent processes are output to achieve dynamic monitoring and propagation simulation of construction progress deviations.

8. The construction progress dynamic optimization method based on BIM and computer vision according to claim 1 is characterized in that: Based on step 4, we combine Bayesian networks and Monte Carlo simulation to predict risks, specifically: Risk feature acquisition: extract historical construction data from the project management system; use a set number of historical risk event data, train and learn using the K2 algorithm, and generate a conditional probability table for the Bayesian network; Construct a Bayesian network based on the conditional probability table; Risk processing: Set the number of simulations. During each simulation, generate risk scenarios based on the Bayesian network and calculate the corresponding risk value. Setting risk thresholds; Determine whether each simulated risk scenario triggers risk through the indicator function; Count the proportion of risk events in all simulation results as the probability of construction risk occurrence; Generate and output a set number of high-probability risk scenarios.

9. The construction progress dynamic optimization method based on BIM and computer vision according to claim 1, characterized in that: Based on step 5, a dynamic optimization plan for construction sequence and resource allocation is generated through a hierarchical reinforcement learning architecture, which includes: Decision feature acquisition: Divide decision features into upper-level features and lower-level features. Upper-level features include the current construction stage and critical path activity status, while lower-level features include resource requirements for each process and material inventory levels. A reward function is also pre-set. Algorithm processing: In the hierarchical decision-making architecture, the upper-level strategy uses a convolutional neural network, which takes as input the construction stage feature map and outputs the construction stage transition probability; the lower-level strategy uses a fully connected network, which takes as input the resource feature vector and outputs the resource allocation plan; Actor network: It is designed as a 3-layer fully connected layer structure with 256 neurons in each layer to output the decision action probability Critic network: uses value function approximation method to evaluate the value of the current decision; Training process: Set the learning rate and discount factor, and train multiple rounds through the PPO algorithm until the fluctuation of the cumulative reward value meets the preset fluctuation threshold. During training, the parameters of the Actor and Critic networks are dynamically updated based on the current construction status, decision actions, and reward feedback to optimize the construction sequence and resource allocation strategy.

10. A construction progress dynamic optimization system based on BIM and computer vision using the method according to any one of claims 1 to 9, characterized in that: Includes the following modules: Data acquisition module, used to obtain BIM model point cloud data and construction site RGB-D images to establish a unified spatial coordinate system; A visual recognition module is used to process construction site images using visual recognition algorithms to identify component status, equipment operating parameters, quality defects, and safety violations, and generate visual recognition results; The four-dimensional BIM modeling module is used to build a four-dimensional dynamic BIM model and map the visual recognition results to the corresponding components in real time through multi-feature similarity calculation; The progress deviation monitoring module is used to monitor construction progress deviations in real time and generate delay time distribution through the dynamic identification algorithm of critical paths and the deviation propagation matrix; Risk prediction module, which integrates Bayesian networks and Monte Carlo simulation algorithms to predict construction risks based on historical risk data and real-time environmental parameters; A dynamic optimization decision module is used to generate dynamic optimization plans for construction sequence and resource allocation through a hierarchical reinforcement learning architecture.

Images