AI-assisted construction error real-time identification and automatic correction method and system

Abstract

The invention relates to an AI-assisted construction error real-time identification and automatic correction method and system. The method comprises the following steps: based on a real-time construction image, adopting a convolutional neural network to identify a construction error and generate an identification report; if the report has no error, constructing a three-dimensional construction component model by using a point cloud registration algorithm, and calculating the size and position deviation values of the three-dimensional construction component model and the design model; according to the construction error type or the deviation value, adopting a weighted decision tree model to generate a risk level; and based on the risk level matching correction strategy database, generating a correction prompt containing the positioning coordinate and the operation instruction, and pushing the correction prompt. By adopting the method, real-time three-dimensional quality monitoring, dynamic risk assessment and closed-loop correction guidance of the construction process can be realized, the error recognition precision and correction efficiency are improved, and the manual intervention cost is reduced.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for construction quality, and in particular relates to an AI-assisted method and system for real-time identification and automatic correction of construction errors. Background Technology

[0002] With the development of intelligent construction technology, computer vision-based construction quality inspection methods have emerged. This technology automatically identifies surface defects (such as cracks and misalignments) by analyzing construction images. This leads to the current construction quality monitoring methods or traditional manual inspection methods. In traditional technologies, construction quality control mainly relies on manual on-site inspections and offline measurements: in manual inspections, supervisors visually inspect construction errors based on experience and use tools such as measuring tapes and total stations to randomly check the dimensions of key components; existing automated technologies, some solutions use two-dimensional image recognition algorithms (such as target detection) to locate obvious construction errors, or use BIM models to compare static point cloud data to assess completion deviations. However, current monitoring methods have the following problems: insufficient real-time performance, long manual inspection cycles and slow response times, inability to cover dynamic construction processes, and difficulty in correcting errors by the time they are discovered; lack of three-dimensional quantization, existing image recognition technologies can only detect surface errors and lack the ability to model three-dimensional dimensional deviations of hidden structures (such as rebar spacing and component spatial pose) in real time; broken decision-making loop, traditional methods stop at error reports or deviation data, do not generate risk level-driven correction instructions, rely on secondary manual decision-making, resulting in low correction efficiency; weak anti-interference ability: factors such as construction obstruction and changes in lighting can easily lead to false detection or missed detection by image recognition algorithms. Summary of the Invention

[0003] Therefore, it is necessary to provide an AI-assisted method and system for real-time identification and automatic correction of construction errors that can solve the above problems.

[0004] Firstly, this application provides an AI-assisted method for real-time identification and automatic correction of construction errors, including:

[0005] Based on real-time construction images, a convolutional neural network is used to identify construction errors and generate an identification report.

[0006] If no construction errors are found in the identification report, a three-dimensional construction component model is constructed based on the real-time construction images using a point cloud registration algorithm, and the dimensional and positional deviations between the construction component model and the design model are calculated.

[0007] Based on the construction error types or deviation values ​​in the identification report, a weighted decision tree model is used to generate risk levels;

[0008] Based on the risk level, the system matches the construction plan from the correction strategy database, generates a correction prompt containing positioning coordinates and operation instructions, and pushes it to the user's terminal.

[0009] In one embodiment, based on real-time construction images, a convolutional neural network is used to identify construction errors and generate an identification report, including:

[0010] Extract feature points, edge contours, and texture features from real-time construction images to form an image feature set;

[0011] A convolutional neural network feature extraction layer is used to abstract and extract image features layer by layer, and a classification layer is used to output the probability distribution of each type of construction error.

[0012] Based on a preset probability screening threshold, if the probability of any construction error type is greater than or equal to the probability screening threshold, the construction error type with the highest probability is taken as the identification result; if the probability of all construction error types is less than the probability screening threshold, it is identified as no construction error.

[0013] A corresponding recognition report is generated based on the recognition results.

[0014] In one embodiment, a three-dimensional construction component model is constructed using a point cloud registration algorithm based on real-time construction images, and the dimensional and positional deviations between the construction component model and the design model are calculated, including:

[0015] Sparse point cloud data of construction components are generated based on real-time construction images;

[0016] Based on sparse point cloud data, segment and identify the point cloud of construction components;

[0017] The point cloud registration algorithm is used to register the point cloud of the construction component with the standard point cloud corresponding to the design model to construct a three-dimensional construction component model.

[0018] Based on the point cloud registration results, the deviation values ​​of the contour dimensions and spatial pose between the 3D construction component model and the design model are calculated.

[0019] In one embodiment, a risk level is generated using a weighted decision tree model based on the construction error type or deviation value in the identification report, including:

[0020] Based on the construction error type identified in the report, a preset weighting coefficient W is matched. 错误 and range influence coefficient C 错误 ;

[0021] Based on the deviation value D 偏差 Matching preset weight coefficient W 偏差 and range influence coefficient C 偏差 ;

[0022] Calculate the overall risk value using the following formula:

[0023] P=α×(W 错误 ×C错误 )+β×(D 偏差 ×W 偏差 ×C 偏差 )

[0024] Where P is the overall risk value, W 错误 C represents the preset weighting coefficient corresponding to the construction error type. 错误 D is the scope impact coefficient corresponding to the construction error type. 偏差 W represents the deviation values ​​of the outline dimensions and spatial pose between the 3D construction component model and the design model. 偏差 The deviation value D 偏差 The corresponding preset weighting coefficient, C 偏差 The deviation value D 偏差 The corresponding range influence coefficients, α and β are path weight factors and α+β=1;

[0025] Risk levels are generated based on the comprehensive risk value using a weighted decision tree model.

[0026] In one embodiment, after calculating the dimensional and positional deviations between the construction component model and the design model, the method further includes:

[0027] Extract the deviation value D from continuous time-series acquisition 偏差 Construct a time series dataset;

[0028] The time series data dataset is fitted by a time series data analysis algorithm to predict the deviation trend of spatial pose or contour size within a preset period in the future; the time series data analysis algorithm adopts the ARIMA model.

[0029] In one embodiment, the method further includes:

[0030] Based on the trend of prediction deviation, calculate the probability P that the prediction deviation value exceeds the preset risk threshold. 预警 ;

[0031] When probability P 预警 When the value is greater than or equal to the preset threshold, an early warning report is generated that includes the deviation location coordinates, development trend, and risk level.

[0032] Based on the early warning report, update and correct the construction scheme matching rules in the strategy database.

[0033] In one embodiment, before identifying construction errors using a convolutional neural network based on real-time construction images, the method further includes:

[0034] The semantic segmentation network is used to identify occluded areas in real-time construction images and generate occluded area masks.

[0035] Based on the occlusion area mask, a generative adversarial network is used to complete the content of the occluded area and generate a complete construction image.

[0036] The complete construction images are input as real-time construction images into the subsequent construction error identification process.

[0037] Secondly, this application also provides an AI-assisted real-time construction error identification and automatic correction system, including:

[0038] The real-time error identification module is used to identify construction errors based on real-time construction images using a convolutional neural network and generate an identification report.

[0039] The 3D reconstruction and deviation calculation module is used to construct a 3D construction component model based on real-time construction images using a point cloud registration algorithm if there are no construction errors in the identification report, and to calculate the dimensional and positional deviation values ​​between the construction component model and the design model.

[0040] The risk level assessment module is used to generate risk levels based on the types of construction errors or deviation values ​​identified in the reports, using a weighted decision tree model.

[0041] The correction strategy execution module is used to match construction plans from the correction strategy database based on risk level, generate correction prompts containing positioning coordinates and operation instructions, and push them to the user terminal.

[0042] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described AI-assisted real-time identification and automatic correction method for construction errors.

[0043] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described AI-assisted method for real-time identification and automatic correction of construction errors.

[0044] The aforementioned AI-assisted real-time construction error identification and automatic correction method and system, computer equipment, and storage medium utilize convolutional neural networks to identify errors in real-time construction images, generating immediate response reports and eliminating the lag in manual inspections. When no construction errors are found, a three-dimensional construction component model is constructed using a point cloud registration algorithm, and dimensional and positional deviation values ​​are quantitatively calculated, overcoming the blind spots of two-dimensional detection and achieving spatial diagnosis of hidden structures. A weighted decision tree model dynamically generates risk levels based on the type of construction error or deviation value. Based on the risk level matching correction strategy database, correction prompts containing positioning coordinates and operation instructions are generated and pushed in real time, forming a closed-loop control mechanism of detection-evaluation-correction. This directly outputs executable instructions, eliminating the need for secondary manual decision-making. It enhances the anti-interference capability and system robustness in complex construction environments, achieving a dual leap in the efficiency and accuracy of construction quality control. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of an AI-assisted method for real-time identification and automatic correction of construction errors according to the present invention.

[0047] Figure 2 This is a structural diagram of an AI-assisted real-time construction error identification and automatic correction system according to the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] In one embodiment, such as Figure 1As shown, an AI-assisted method for real-time identification and automatic correction of construction errors is provided. This embodiment illustrates the application of this method to an intelligent construction quality monitoring system, which consists of a front-end industrial camera, an edge computing device, a cloud platform, and an execution terminal. In a concrete pouring scenario, the front-end camera captures construction images of beam-column nodes, and the edge device calls a convolutional neural network model to identify errors. If no errors are detected, the edge device generates a sparse point cloud of the component and uploads it to the cloud. The cloud constructs a 3D model through point cloud registration and calculates the pose deviation from the BIM design model. Based on the deviation value, a weighted decision tree is activated to generate a risk level, and a correction strategy library is matched to generate instructions, which are then pushed to AR glasses via a 5G network to guide manual calibration of template offset.

[0050] In this embodiment, the method includes the following steps:

[0051] S01 uses a convolutional neural network to identify construction errors based on real-time construction images and generates an identification report.

[0052] This process can involve extracting scale- and rotation-invariant keypoints, identifying the edge contours of components in an image to form structured shape features, and extracting texture features such as cracks and rebar distribution on concrete surfaces to form an image feature set, integrating the geometric and texture information of the image. Using a convolutional neural network architecture such as ResNet, YOLO, or Faster R-CNN, the feature set is abstracted layer by layer through multiple convolutional kernels. A fully connected layer outputs the current image's classification as belonging to various construction error types (such as cracks, missing components, dimensional deviations, etc.), generating an identification report. This report can include the error type (such as honeycomb-like surface on concrete), confidence level, error location coordinates (based on image pixel coordinates or 3D world coordinates), image screenshots, and feature annotations (such as marking error areas with bounding boxes). This overcomes the efficiency bottleneck of traditional manual inspections, enabling real-time monitoring and digital recording of construction quality.

[0053] S02. If there are no construction errors in the identification report, construct a three-dimensional construction component model based on the real-time construction images using a point cloud registration algorithm, and calculate the size and position deviation values ​​between the construction component model and the design model.

[0054] This process involves using real-time construction images of the same component from multiple angles (e.g., four views: front, back, left, and right) to calculate camera pose parameters using the SFM (Structure of Motion) algorithm. An initial sparse point cloud is then generated using epipolar constraints and feature matching (e.g., SIFT feature point matching). A random sampling consensus algorithm is employed to segment the point cloud into geometric primitives such as planes (e.g., template surfaces) and columns (e.g., reinforcing bars), filtering out background point clouds (e.g., scaffolding, ground). Fast Point Feature Hash (FPFH) or Visual Bag-of-Words Hash (VFH) is extracted from the construction component point cloud and the standard point cloud of the design model. Feature point pairs are matched using the RANSAC algorithm, and an initial transformation matrix (rotation matrix R and translation vector T) is calculated to achieve preliminary alignment of the two point clouds. The registered point cloud is converted into a surface mesh model through triangular mesh reconstruction (such as the Ball Pivoting algorithm). Combined with the BIM parameters of the design model (such as material and dimension annotations), a 3D construction component model is generated. Dimensional and positional deviations are calculated. For straight components, the absolute deviations in length, width, and height are calculated. For curved components (such as arch beams), the shape deviation is calculated by finding the minimum distance between the sampled point cloud and the design curve. Translational and rotational deviations are calculated using the coordinate axes in the model. This forms a dual quality control mechanism of surface inspection and 3D verification.

[0055] S03. Based on the construction error type or deviation value in the identification report, a weighted decision tree model is used to generate the risk level.

[0056] The process involves assigning different weights to construction error types and deviation values ​​(e.g., for construction error types like rebar misalignment, and the threshold range of the deviation value) and calculating a comprehensive risk value. This comprehensive risk value is then input into a weighted decision tree model to generate a risk level. The splitting conditions for the weighted decision tree model can include: whether the error type belongs to the structural safety category (e.g., missing rebar, deviation of load-bearing components); whether the deviation value exceeds 10% of the design threshold; and whether the impact scope involves critical load-bearing components. The decision tree parameters can be iteratively optimized based on historical construction data. Through a multi-dimensional risk quantification system encompassing error type, deviation degree, and impact scope, data support is provided for the precise matching of corrective strategies.

[0057] S04: Based on the risk level, match the construction plan from the correction strategy database, generate a correction prompt containing positioning coordinates and operation instructions, and push it to the user terminal.

[0058] The database can be a relational database (such as MySQL) to store structured strategy data, including: an error type index, with primary keys established according to construction error types (such as rebar misalignment, formwork offset); risk level partitioning, with each error type corresponding to a strategy subset of different risk levels; a 3D positioning template, storing standard positioning coordinate parameters for various components (such as beam axis coordinates, column center point coordinates); and an operation instruction library: standardized instruction templates (such as tightening [bolt number], torque to [X] N·m). The matching algorithm can use a rule-based expert system (such as Drools) or vector similarity matching (such as the TF-IDF algorithm) to convert risk level, error type, and deviation value into feature vectors, calculate cosine similarity with the feature labels of the database strategies, and take strategies with similarity ≥ 0.8 as the matching result. The instruction template is called according to the strategy type, and corresponding correction prompts are generated according to the coordinates of the detected target, such as adjusting the horizontal displacement of beam KL-12 in area D1 by 15mm to the east using a hydraulic jack. Establish an automated mapping mechanism for risk level, remedial measures, and execution guidelines to achieve full-process digitalization from quality diagnosis to on-site operation.

[0059] The aforementioned AI-assisted real-time construction error identification and automatic correction method employs convolutional neural network recognition technology based on real-time construction images. It can generate construction error identification reports in real time, solving the problems of long cycles and slow response times associated with traditional manual inspections, and enabling real-time monitoring of the dynamic construction process. By constructing a 3D construction component model through a point cloud registration algorithm and calculating the deviation value from the design model, it overcomes the blind spot of 2D detection in monitoring the 3D dimensional deviations of hidden structures. Using a weighted decision tree model, it integrates error types and deviation values ​​to generate risk levels, and generates correction prompts containing positioning coordinates and operation instructions based on a risk level matching correction strategy database, forming a closed-loop control mechanism of detection-evaluation-correction. This eliminates the need for secondary manual decision-making and improves correction efficiency. Through the integration of multiple technologies, it achieves real-time 3D monitoring of construction quality, dynamic risk assessment, and closed-loop correction guidance, improving error identification accuracy and correction efficiency, reducing manual intervention costs, and solving the technical problems of insufficient real-time performance, lack of 3D quantification, and broken decision-making loops in existing technologies.

[0060] In one embodiment, based on real-time construction images, a convolutional neural network is used to identify construction errors and generate an identification report, including:

[0061] S11, extract feature points, edge contours and texture features of real-time construction images to form an image feature set;

[0062] S12 uses a convolutional neural network feature extraction layer to abstract and extract image features layer by layer, and combines the classification layer to output the probability distribution of each type of construction error.

[0063] S13. Based on a preset probability screening threshold, if the probability of any construction error type is greater than or equal to the probability screening threshold, the construction error type with the highest probability is taken as the identification result; if the probability of all construction error types is less than the probability screening threshold, it is identified as no construction error.

[0064] S14, Generate a corresponding recognition report based on the recognition results.

[0065] Specifically, an image feature set is formed by extracting feature points, edge contours, and texture features from real-time construction images. Feature point extraction can use SIFT or ORB algorithms to obtain key points with scale and rotation invariance. Edge contour extraction uses the Canny operator or HOG algorithm to identify component boundaries. Texture feature extraction uses Gabor filters or LBP algorithms to capture surface details (such as cracks and rebar distribution), constructing a multi-dimensional feature set containing geometric and texture information. The feature set is then abstracted layer by layer using feature extraction layers of convolutional neural networks (such as multi-layer convolutional modules in ResNet and YOLO architectures). The bottom convolutional layer extracts edges and corners. Based on basic features such as points, high-level convolutional layers generate semantic features such as rebar misalignment and formwork deformation. These features can be combined with classification layers and the Softmax function can output the probability distribution of each construction error type. Judgment is based on a preset probability filtering threshold (determined according to historical experience and actual working conditions): if the probability of any error type is greater than or equal to the threshold, the type with the highest probability is taken as the recognition result; if the probabilities of all types are less than the threshold, it is determined that there is no construction error. Low-confidence results can be filtered to reduce the false detection rate. A structured recognition report is generated based on the recognition results, which can include error type, confidence level, location coordinates, and image annotation information, realizing automated recognition and digital recording of construction errors.

[0066] In one embodiment, a three-dimensional construction component model is constructed using a point cloud registration algorithm based on real-time construction images, and the dimensional and positional deviations between the construction component model and the design model are calculated, including:

[0067] S21, generating sparse point cloud data of construction components based on real-time construction images;

[0068] S22, based on sparse point cloud data, segment and identify the point cloud of construction components;

[0069] S23, the point cloud of the construction component is registered with the standard point cloud corresponding to the design model through the point cloud registration algorithm to construct a three-dimensional construction component model;

[0070] S24, based on the point cloud registration results, calculate the deviation values ​​of the contour dimensions and spatial pose between the 3D construction component model and the design model.

[0071] For example, sparse point cloud data of construction components can be generated based on real-time construction images using multi-view stereo matching (MVS) technology. Camera pose parameters from multiple angle images are calculated using the SFM algorithm, and a sparse point cloud containing the 3D coordinates of key feature points on the component surface is constructed by combining SIFT feature point matching and epipolar constraints. Based on the sparse point cloud data, the RANSAC plane segmentation algorithm can be used to segment the point cloud into geometric primitives such as planes and cylinders, filtering background point clouds. Semantic annotation of the segmented point cloud is then performed using semantic segmentation networks such as PointNet, enabling automatic identification and extraction of point clouds from construction components. The point cloud registration algorithm registers the point cloud of the construction component with the standard point cloud corresponding to the design model. It uses FPFH features combined with RANSAC to achieve coarse registration, and uses Go-ICP or CPD algorithms to iteratively optimize transformation parameters to achieve fine registration. It then reconstructs a 3D construction component model through triangular mesh. Based on the point cloud registration results, it extracts the minimum bounding box or geometric fitting parameters of the 3D construction component model, compares it with the design model to calculate the deviation of the contour dimensions (such as the absolute deviation in the length and width directions), and calculates the spatial pose deviation (such as translation and rotation angle) through the centroid coordinate difference and Euler angles to achieve deviation quantification.

[0072] In one embodiment, a risk level is generated using a weighted decision tree model based on the construction error type or deviation value in the identification report, including:

[0073] S31, Match the preset weighting coefficient W according to the construction error type in the identification report. 错误 and range influence coefficient C 错误 ;

[0074] S32, based on the deviation value D 偏差 Matching preset weight coefficient W 偏差 and range influence coefficient C 偏差 ;

[0075] S33, Calculate the overall risk value according to the following formula:

[0076] P=α×(W 错误 ×C 错误 )+β×(D 偏差 ×W 偏差 ×C 偏差 )

[0077] Where P is the overall risk value, W 错误 C represents the preset weighting coefficient corresponding to the construction error type. 错误 D is the scope impact coefficient corresponding to the construction error type. 偏差 W represents the deviation values ​​of the outline dimensions and spatial pose between the 3D construction component model and the design model. 偏差 The deviation value D 偏差 The corresponding preset weighting coefficient, C偏差 The deviation value D 偏差 The corresponding range influence coefficients, α and β are path weight factors and α+β=1;

[0078] S34 generates risk levels based on comprehensive risk values ​​using a weighted decision tree model.

[0079] Specifically, the preset weighting coefficient W 错误 The impact of the error on structural safety can be assigned a value (e.g., 0.8 for rebar misalignment, 0.3 for surface cracks), and the range influence coefficient C. 错误 The spatial range representing the impact of errors (e.g., assigning a value of 0.6 to the impact of local components and 1.0 to the impact of the overall structure); preset weighting coefficient W. 偏差 and range influence coefficient C 偏差 Matching can be achieved using piecewise functions, such as D. 偏差 When ≤5mm, W 偏差 =0.2, C 偏差 =0.3, when 5mm≤D 偏差 When ≤15mm, W 偏差 =0.5, C 偏差 =0.6, where W 偏差 and C 偏差 The initial value assignment method is the same as W. 错误 and C 错误 And it can be adjusted based on piecewise functions; through the formula P=α×(W 错误 ×C 错误 )+β×(D 偏差 ×W 偏差 ×C 偏差 After calculating the comprehensive risk value P, a risk level can be generated using a weighted decision tree model constructed using the CART algorithm. The decision tree has P as its root node, and the splitting conditions include whether the error type involves structural safety and whether the deviation value exceeds the threshold of 10%. Leaf nodes can correspond to risk levels of low P = 0.3, medium 0.3 ≤ P < 0.7, and high P ≥ 0.7. Furthermore, the decision tree parameters can be iteratively optimized based on historical data (e.g., increasing W when a certain type of error repeatedly triggers high risk). 错误 By using a quantitative fusion and dynamic weight adjustment mechanism of multi-source risk factors, the automatic generation and adaptive optimization of risk levels can be achieved.

[0080] In one embodiment, after calculating the dimensional and positional deviations between the construction component model and the design model, the method further includes:

[0081] S41, Extract the deviation value D from the continuous time-series acquisition. 偏差 Construct a time series dataset;

[0082] S42 uses a time series data analysis algorithm to fit the time series dataset and predict the deviation trend of spatial pose or contour size within a preset period; the time series data analysis algorithm adopts the ARIMA model.

[0083] For example, the collection frequency of the time series dataset can be set according to the characteristics of the construction process (e.g., once per minute during the concrete pouring stage, and once every 10 minutes during the steel structure installation stage). The time series dataset is fitted using time series data analysis algorithms, and an ARIMA (Autoregressive Integral Moving Average) model is used to predict deviation trends: the stationarity of the deviation sequence is tested (e.g., ADF test); if the sequence is non-stationary, it is transformed into a stationary sequence by d-order differencing; then, the autoregressive order p and the moving average order q are determined using the autocorrelation function (ACF) and partial autocorrelation function (PACF), and an ARIMA(p,d,q) model is constructed; the model parameters are fitted using the maximum likelihood estimation method, and the randomness of the residual sequence is verified using the Ljung-Box test to ensure model effectiveness; based on the fitted model, the deviation trend of spatial pose or contour dimensions within a preset period (e.g., 8 hours) is predicted, and the deviation trend, including the deviation value prediction curve and confidence interval, is output. Through dynamic analysis of time series data, single-point deviation detection is upgraded to trend prediction, identifying deviation evolution risks in advance and realizing a shift from passive correction to proactive prevention.

[0084] In one embodiment, the method further includes:

[0085] S51, Based on the trend of the prediction deviation, calculate the probability P that the prediction deviation value exceeds the preset risk threshold. 预警 ;

[0086] S52, when probability P 预警 When the value is greater than or equal to the preset threshold, an early warning report is generated that includes the deviation location coordinates, development trend, and risk level.

[0087] S53, based on the early warning report, update and correct the construction scheme matching rules in the strategy database.

[0088] Specifically, based on the trend of the predicted deviation, Monte Carlo simulation or statistical inference methods are used to calculate the probability P that the predicted deviation value exceeds a preset risk threshold (such as the maximum permissible deviation defined by structural safety specifications). 预警 By performing more than 1000 random samplings on the prediction curve, the fluctuation range of the deviation value within a preset future period can be simulated, and the proportion of samples exceeding the risk threshold can be counted to obtain P. 预警 When probability P 预警When the deviation exceeds a preset threshold (determined based on experience and actual working conditions), an early warning report is generated, containing three-dimensional deviation positioning coordinates (e.g., X = 15.2m, Y = 3.8m, Z = 2.5m), a deviation development trend curve (e.g., displacement growth rate of 0.5mm / h in the next 24 hours), and a risk level (e.g., high risk). Based on the deviation characteristics and risk patterns in the early warning report, the construction scheme matching rules in the correction strategy database can be updated using machine learning algorithms: if the early warning frequency of a certain type of deviation (e.g., support settlement) exceeds a preset threshold (e.g., 3 times / week), the matching priority of the corresponding correction scheme (e.g., adding diagonal bracing) is increased by 20%, and the scheme parameters are adjusted (e.g., optimizing the bracing spacing from 1.5m to 1.2m). Simultaneously, if a new combination of deviations (e.g., a concurrent combination of displacement and cracks) appears in the early warning event, a new strategy matching entry is created, achieving dynamic evolution of the database. This solves the problems of missing risk prediction and rigid strategies in existing technologies.

[0089] In one embodiment, before identifying construction errors using a convolutional neural network based on real-time construction images, the method further includes:

[0090] S61, using a semantic segmentation network to identify occluded areas in real-time construction images and generate occluded area masks;

[0091] S62, based on the occlusion area mask, uses a generative adversarial network to complete the content of the occlusion area and generate a complete construction image;

[0092] S63 inputs the complete construction image as a real-time construction image into the subsequent construction error identification process.

[0093] For example, before identifying construction errors, a semantic segmentation network (such as U-Net or DeepLab) can be used to perform pixel-level semantic segmentation on real-time construction images to identify areas obscured by materials, personnel, or equipment, generating a binary mask of the obscured areas. White pixels can be used to mark the obscured areas, and black pixels can be used to mark the visible areas. Based on the mask of the obscured areas, a generative adversarial network (GAN) architecture (such as pix2pix or CycleGAN) can be used to complete the content of the obscured areas: the generator network takes the mask image and the features of the visible areas as input, and generates the predicted content of the obscured areas through multi-layer transposed convolutions. The discriminator network distinguishes the difference between the completed image and the real unobscured image. Through adversarial training, the completed content output by the generator is made consistent with the surrounding texture and structure (such as the distribution pattern of steel bars and the direction of formwork joints). The complete construction image is used as the input of the real-time construction image into the subsequent convolutional neural network recognition process, which can solve the problem of false detection and false detection caused by occlusion in the construction environment and improve the recognition robustness under complex working conditions.

[0094] The aforementioned AI-assisted real-time construction error identification and automatic correction method is based on convolutional neural network (CNN) recognition technology for real-time construction images. By extracting feature points, edge contours, and texture features and abstracting them layer by layer through a CNN, it achieves second-level error identification and report generation, eliminating the problems of long cycles and slow response times associated with traditional manual inspections, and covering the entire dynamic construction process monitoring. Utilizing a point cloud registration algorithm, it generates sparse point clouds from multi-view images, segments and identifies them, and then registers them with the standard point cloud of the design model to construct a 3D component model. It calculates dimensional and pose deviations, overcoming the blind spot of 3D quantization of hidden structures (such as rebar spacing and component spatial pose) in 2D detection. Finally, it integrates error type weight coefficients, range influence coefficients, and deviation value weight coefficients through a weighted decision tree model to calculate... The system calculates a comprehensive risk value and generates a risk level, driving the correction strategy database to match construction plans containing positioning coordinates and operation instructions, forming a closed-loop control of detection-evaluation-correction. This eliminates the need for secondary manual decision-making and improves correction efficiency. Semantic segmentation networks and generative adversarial networks are used to complete content in occluded areas. Combined with an ARIMA model, the system predicts deviation trends and calculates the probability of exceeding risk thresholds, dynamically updating correction strategies. This addresses interference factors such as construction occlusion and lighting changes, enabling real-time 3D monitoring of construction quality, dynamic risk assessment, and closed-loop correction guidance. It improves error identification accuracy, shortens correction time, and reduces manual intervention costs, solving technical problems in existing technologies such as insufficient real-time performance, lack of 3D quantization, broken decision-making loops, and weak anti-interference capabilities.

[0095] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0096] Based on the same inventive concept, this application also provides an AI-assisted real-time construction error identification and automatic correction system for implementing the aforementioned AI-assisted real-time construction error identification and automatic correction method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the AI-assisted real-time construction error identification and automatic correction system provided below can be found in the limitations of the AI-assisted real-time construction error identification and automatic correction method described above, and will not be repeated here.

[0097] In one exemplary embodiment, such as Figure 2 As shown, an AI-assisted real-time construction error identification and automatic correction system is provided, including:

[0098] The real-time error recognition module 101 is used to identify construction errors based on real-time construction images using a convolutional neural network and generate an identification report.

[0099] The 3D reconstruction and deviation calculation module 102 is used to construct a 3D construction component model based on real-time construction images using a point cloud registration algorithm if there are no construction errors in the identification report, and to calculate the size and position deviation values ​​between the construction component model and the design model.

[0100] The risk level assessment module 103 is used to generate risk levels based on the construction error types or deviation values ​​in the identification report using a weighted decision tree model.

[0101] The correction strategy execution module 104 is used to match construction plans from the correction strategy database based on risk level, generate correction prompts containing positioning coordinates and operation instructions, and push them to the user terminal.

[0102] In one embodiment, the real-time error detection module 101 is further configured to:

[0103] Extract feature points, edge contours, and texture features from real-time construction images to form an image feature set;

[0104] A convolutional neural network feature extraction layer is used to abstract and extract image features layer by layer, and a classification layer is used to output the probability distribution of each type of construction error.

[0105] Based on a preset probability screening threshold, if the probability of any construction error type is greater than or equal to the probability screening threshold, the construction error type with the highest probability is taken as the identification result; if the probability of all construction error types is less than the probability screening threshold, it is identified as no construction error.

[0106] A corresponding recognition report is generated based on the recognition results.

[0107] In one embodiment, the 3D reconstruction and deviation calculation module 102 is further configured to:

[0108] Sparse point cloud data of construction components are generated based on real-time construction images;

[0109] Based on sparse point cloud data, segment and identify the point cloud of construction components;

[0110] The point cloud registration algorithm is used to register the point cloud of the construction component with the standard point cloud corresponding to the design model to construct a three-dimensional construction component model.

[0111] Based on the point cloud registration results, the deviation values ​​of the contour dimensions and spatial pose between the 3D construction component model and the design model are calculated.

[0112] In one embodiment, the risk level assessment module 103 is further configured to:

[0113] Based on the construction error type identified in the report, a preset weighting coefficient W is matched. 错误 and range influence coefficient C 错误 ;

[0114] Based on the deviation value D 偏差 Matching preset weight coefficient W 偏差 and range influence coefficient C 偏差 ;

[0115] Calculate the overall risk value using the following formula:

[0116] P=α×(W 错误 ×C 错误 )+β×(D 偏差 ×W 偏差 ×C 偏差 )

[0117] Where P is the overall risk value, W 错误 C represents the preset weighting coefficient corresponding to the construction error type. 错误 D is the scope impact coefficient corresponding to the construction error type. 偏差 W represents the deviation values ​​of the outline dimensions and spatial pose between the 3D construction component model and the design model. 偏差 The deviation value D 偏差 The corresponding preset weighting coefficient, C 偏差 The deviation value D 偏差 The corresponding range influence coefficients, α and β are path weight factors and α+β=1;

[0118] Risk levels are generated based on the comprehensive risk value using a weighted decision tree model.

[0119] In one embodiment, the system further includes a deviation trend prediction module, used for:

[0120] Extract the deviation value D from continuous time-series acquisition 偏差 Construct a time series dataset;

[0121] The time series data dataset is fitted by a time series data analysis algorithm to predict the deviation trend of spatial pose or contour size within a preset period in the future; the time series data analysis algorithm adopts the ARIMA model.

[0122] In one embodiment, the deviation trend prediction module is further used for:

[0123] Based on the trend of prediction deviation, calculate the probability P that the prediction deviation value exceeds the preset risk threshold. 预警 ;

[0124] When probability P 预警 When the value is greater than or equal to the preset threshold, an early warning report is generated that includes the deviation location coordinates, development trend, and risk level.

[0125] Based on the early warning report, update and correct the construction scheme matching rules in the strategy database.

[0126] In one embodiment, the real-time error detection module 101 is further configured to:

[0127] The semantic segmentation network is used to identify occluded areas in real-time construction images and generate occluded area masks.

[0128] Based on the occlusion area mask, a generative adversarial network is used to complete the content of the occluded area and generate a complete construction image.

[0129] The complete construction images are input as real-time construction images into the subsequent construction error identification process.

[0130] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the AI-assisted real-time identification and automatic correction method for construction errors as described above.

[0131] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0132] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0133] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for real-time identification and automatic correction of construction errors assisted by AI, characterized in that, The method includes: Based on real-time construction images, a convolutional neural network is used to identify construction errors and generate an identification report. If there are no construction errors in the identification report, a three-dimensional construction component model is constructed using a point cloud registration algorithm based on the real-time construction image, and the size and position deviation values ​​between the construction component model and the design model are calculated. Based on the construction error type or deviation value in the identification report, a weighted decision tree model is used to generate a risk level; Based on the risk level, the system matches the construction plan from the correction strategy database, generates a correction prompt containing positioning coordinates and operation instructions, and pushes it to the user terminal.

2. The method according to claim 1, characterized in that, The method of identifying construction errors based on real-time construction images using a convolutional neural network and generating an identification report includes: Extract feature points, edge contours, and texture features from the real-time construction images to form an image feature set; The feature extraction layer of the convolutional neural network is used to abstract and extract the image feature set layer by layer, and the probability distribution of each construction error type is output by the classification layer. Based on a preset probability screening threshold, if the probability of any construction error type is greater than or equal to the probability screening threshold, the construction error type with the highest probability is taken as the identification result; if the probability of all construction error types is less than the probability screening threshold, it is identified as no construction error. A corresponding identification report is generated based on the identification results.

3. The method according to claim 1, characterized in that, The step of constructing a three-dimensional construction component model based on the real-time construction images using a point cloud registration algorithm, and calculating the dimensional and positional deviations between the construction component model and the design model, includes: Sparse point cloud data of construction components are generated based on the real-time construction images; Based on the sparse point cloud data, the point cloud of construction components is segmented and identified; The point cloud of the construction component is registered with the standard point cloud corresponding to the design model using a point cloud registration algorithm to construct a three-dimensional construction component model. Based on the point cloud registration results, the deviation values ​​of the contour dimensions and spatial pose between the three-dimensional construction component model and the design model are calculated.

4. The method according to claim 3, characterized in that, The step of generating a risk level using a weighted decision tree model based on the construction error type or deviation value in the identification report includes: A preset weighting coefficient W is matched based on the type of construction error in the identification report. 错误 and range influence coefficient C 错误 ; According to the deviation value D 偏差 Matching preset weight coefficient W 偏差 and range influence coefficient C 偏差 ; Calculate the overall risk value using the following formula: P=α×(W 错误 ×C 错误 )+β×(D 偏差 ×W 偏差 ×C 偏差 ) Where P is the overall risk value, W 错误 C represents the preset weighting coefficient corresponding to the construction error type. 错误 D is the scope impact coefficient corresponding to the construction error type. 偏差 W represents the deviation values ​​of the outline dimensions and spatial pose between the 3D construction component model and the design model. 偏差 The deviation value D 偏差 The corresponding preset weighting coefficient, C 偏差 The deviation value D 偏差 The corresponding range influence coefficients, α and β are path weight factors and α+β=1; Based on the comprehensive risk value, a risk level is generated using a weighted decision tree model.

5. The method according to claim 3, characterized in that, After calculating the dimensional and positional deviations between the construction component model and the design model, the method further includes: Extract the deviation value D from continuous time-series acquisition 偏差 Construct a time series dataset; The time series dataset is fitted using a time series data analysis algorithm to predict the deviation trend of spatial pose or contour size within a preset period in the future; wherein, the time series data analysis algorithm adopts the ARIMA model.

6. The method according to claim 5, characterized in that, The method further includes: Based on the trend of prediction deviation, calculate the probability P that the prediction deviation value exceeds the preset risk threshold. 预警 ; When the probability P 预警 When the value is greater than or equal to the preset threshold, an early warning report is generated that includes the deviation location coordinates, development trend, and risk level. Based on the warning report, update the construction scheme matching rules in the correction strategy database.

7. The method according to claim 1, characterized in that, Before using a convolutional neural network to identify construction errors based on real-time construction images, the method also includes: The semantic segmentation network is used to identify occluded areas in real-time construction images and generate occluded area masks. Based on the mask of the occluded area, a generative adversarial network is used to complete the content of the occluded area and generate a complete construction image. The complete construction image is used as the real-time construction image and input into the subsequent construction error identification process.

8. An AI-assisted real-time construction error identification and automatic correction system, characterized in that, The system includes: The real-time error identification module is used to identify construction errors based on real-time construction images using a convolutional neural network and generate an identification report. The 3D reconstruction and deviation calculation module is used to construct a 3D construction component model based on the real-time construction image using a point cloud registration algorithm if there are no construction errors in the identification report, and to calculate the size and position deviation values ​​between the construction component model and the design model. The risk level assessment module is used to generate risk levels based on the types of construction errors or deviation values ​​identified in the report, using a weighted decision tree model. The correction strategy execution module is used to match construction plans from the correction strategy database based on the risk level, generate correction prompts containing positioning coordinates and operation instructions, and push them to the user terminal.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images