Pile plate type roadbed detection and prediction maintenance method and system based on AI analysis

Abstract

The application discloses a pile-slab type roadbed detection and prediction maintenance method and system based on AI analysis, and the method comprises the following steps: acquiring historical record data, multi-source real-time data and engineering profile data of the pile-slab type roadbed; analyzing and fusing the multi-source real-time data based on an AI intelligent model, extracting real-time structure characteristic data of the pile-slab type roadbed, and acquiring real-time state data of the pile-slab type roadbed; establishing a state intelligent prediction model, taking the real-time state data and the historical record data as inputs, and predicting the state of the pile-slab type roadbed in the next detection period; performing risk assessment on the state of the pile-slab type roadbed in the next detection period, and correspondingly selecting a maintenance decision to be executed based on the risk assessment result. The method efficiently processes and fuses real-time data collected by multi-source sensors by using an artificial intelligence model, and establishes an iterative optimization prediction model to predict the state of the pile-slab type roadbed in the next stage and perform risk prediction and estimation, so that the automatic detection, prediction and maintenance of the pile-slab type roadbed are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of road and bridge engineering, and in particular to a pile-slab type roadbed detection and predictive maintenance method and system based on AI analysis. BACKGROUND

[0002] The pile-slab type structure roadbed is a common high-stability roadbed structure, which is widely used in engineering scenarios such as soft soil areas and high-filled road sections. However, long-term load, geological changes and environmental factors can cause roadbed verticality deviation and settlement, affecting road safety. Traditional detection methods such as total station and level devices rely on manual measurement, which is low in efficiency and high in cost, and it is difficult to realize real-time monitoring and predictive maintenance.

[0003] In recent years, AI technology has been applied in the field of civil engineering monitoring, but existing technologies mainly focus on single detection or prediction functions, and lack an integrated solution for verticality detection-settlement prediction-intelligent maintenance of pile-slab type roadbeds. Therefore, there is an urgent need for an intelligent, automated and high-precision monitoring and maintenance system to improve the efficiency and safety of pile-slab type structure roadbed maintenance. SUMMARY

[0004] To solve the above problems of the prior art, the present application provides a pile-slab type roadbed detection and predictive maintenance method and system based on AI analysis, which aims to efficiently process and fuse real-time data collected by multiple sensors using an artificial intelligence model, and to establish an iterative optimization prediction model to predict the next stage of pile-slab type roadbed state, and to use a neural network to estimate risks to realize detection, prediction and automated maintenance of pile-slab type roadbeds.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides a pile-slab type roadbed detection and predictive maintenance method based on AI analysis, comprising:

[0007] In response to a preset detection period, historical record data, multi-source real-time data and engineering profile data of the pile-slab type roadbed are obtained; the multi-source real-time data includes inspection image data, laser point cloud data, contact measurement data and temperature and humidity data;

[0008] The multi-source real-time data is analyzed and fused based on an AI intelligent model to extract real-time structural feature data of the pile-slab type roadbed; real-time state data of the pile-slab type roadbed is obtained based on the real-time structural feature data and the engineering profile data;

[0009] A state intelligent prediction model is established, and the real-time state data and the historical record data are taken as inputs to predict the state of the pile-slab type roadbed in the next detection period;

[0010] Risk assessment is performed on the pile-slab subgrade state of the next detection cycle, and a maintenance decision is correspondingly selected and executed based on a result of the risk assessment.

[0011] Preferably, the historical record data includes historical detection data, historical maintenance data, and historical stress condition data; and the engineering profile data includes geological environment data and pile-slab structure size parameters.

[0012] Preferably, the analysis and fusion of the multi-source real-time data based on the AI intelligent model extracts real-time structure feature data of the pile-slab subgrade, including:

[0013] A pile column 3D model template created offline in advance is projected to a space where real-time laser point clouds are located, point cloud registration is performed based on an iterative closest point algorithm, and a 3D contour description of a pile column in the real-time laser point clouds is obtained;

[0014] Based on projection of the 3D contour description in an inspection image and gradient changes of the inspection image, a pile column in the laser point clouds and the inspection image is matched, and a real-time true pose of the pile column is obtained through iterative optimization;

[0015] The real-time true pose, contact type measurement data, and temperature and humidity data are input into the AI intelligent model for data fusion, and real-time structure feature data of the pile-slab subgrade is output.

[0016] Preferably, the matching of the pile column in the laser point clouds and the inspection image based on the projection of the 3D contour description in the inspection image and the gradient changes of the inspection image and the iterative optimization to obtain the real-time true pose of the pile column includes:

[0017] The 3D contour description is projected into the inspection image to obtain a predicted pile column pose, and a binary mask of the inspection image with respect to the predicted pile column pose is generated;

[0018] The minimum distance of each pixel point in the binary mask to the predicted pile column pose in the inspection image is calculated, and the sign of the minimum distance in different binary value regions is negated;

[0019] The gray scale gradient amplitudes of each pixel point in the inspection image are extracted based on a gradient operator;

[0020] A joint observation probability of the pixel point is constructed based on the gray scale gradient amplitudes of the pixel point and the minimum distance, and a feature matching loss function between the predicted pile column pose and the inspection image is calculated based on the joint observation probability;

[0021] The minimum value of the feature matching loss function is iteratively calculated, and an offset update amount between the predicted pile column pose and the inspection image when the feature matching loss function reaches the minimum value is output.

[0022] Adjust the projection position of the 3D contour description in the inspection image based on the offset update amount, obtain a real pose image of the pile, and output the real-time real pose of the pile.

[0023] Preferably, the joint observation probability of the pixel point is constructed based on the gray gradient amplitude and the minimum distance, and the joint observation probability of the pixel point is specifically:

[0024] ;

[0025] Wherein, represents the pixel point in the binary mask to the minimum distance of the predicted pile pose in the inspection image; represents a smoothing function, which is used to describe the probability distribution function between the minimum distance and the pixel point belongs to the foreground region; is a foreground gradient model of the inspection image, which is used to describe the probability distribution function of the gradient of the foreground region in the inspection image, represents the gradient of the pixel point in the foreground region, represents the image foreground region; is a background gradient model of the inspection image, which is used to describe the probability distribution function of the gradient of the background region in the inspection image, represents the gradient of the pixel point in the background region, represents the image background region.

[0026] Preferably, the foreground gradient model is specifically:

[0027] ;

[0028] The background gradient model is specifically:

[0029] ;

[0030] Wherein, represents the normalized gray gradient amplitude; is a smoothing factor.

[0031] Preferably, the state intelligent prediction model is:

[0032] ;

[0033] In the formula, represents the optimal estimation based on the first detection cycle to the pile plate roadbed state prediction of the first detection cycle; represents the real-time state data of the first detection cycle; a real-time state data control input matrix is represented; a first a historical record data of a detection period is detected an iterative optimization data of a subgrade state of a detection period; a historical record data state transition matrix is represented; a first a smart model estimation data after maintenance and repair of a detection period; an estimation matrix is represented; a first a deviation control matrix of a detection period.

[0034] Preferably, the first a real-time state data of a detection period wherein, is a vertical angle deviation, is a vertical displacement deviation, is a settlement amount, is a settlement rate;

[0035] the smart model estimation data is:

[0036] ;

[0037] wherein, a real-time state data of a detection period is represented; a first a first a first a first a deviation control matrix of a detection period.

[0038] Preferably, the risk assessment of the pile-slab subgrade state of the next detection period is that the real-time state data of the pile-slab subgrade of different detection periods is taken as input, a neural network model is trained and optimized for short-term and long-term risk trend prediction, and maintenance decisions are made according to the risk level: high-risk level takes emergency measures, medium-risk level takes planned maintenance, and low-risk level takes routine maintenance.

[0039] In a second aspect, the present application provides a pile-slab subgrade detection and prediction maintenance system based on AI analysis, comprising:

[0040] a data acquisition module for acquiring historical record data, multi-source real-time data and engineering profile data of a pile-slab subgrade in response to a preset detection period; the multi-source real-time data includes inspection image data, laser point cloud data, contact measurement data, temperature and humidity data;

[0041] an AI analysis module configured to analyze the fused multi-source real-time data based on an AI intelligent model to extract real-time structural feature data of the pile-slab subgrade, and obtain real-time state data of the pile-slab subgrade based on the real-time structural feature data and the engineering profile data;

[0042] an intelligent prediction module configured to establish a state intelligent prediction model, and predict the state of the pile-slab subgrade in the next detection cycle by taking the real-time state data and the historical record data as inputs;

[0043] a maintenance decision module configured to perform risk assessment on the state of the pile-slab subgrade in the next detection cycle, and select a corresponding maintenance decision to be executed based on the risk assessment result.

[0044] The pile-slab subgrade detection and predictive maintenance method and system based on AI analysis has the following beneficial effects:

[0045] The present application can improve the detection efficiency by more than 80% compared with the traditional manual method (such as total station and level), greatly reducing the manual intervention and time cost. The deep learning algorithm is used to realize the integration of detection, prediction and maintenance decision, solving the limitation of single function of traditional technology and forming a closed-loop management. The neural network is used to predict the short-term and long-term risk trends, and the risks are divided into three levels of high, medium and low according to the quantitative indicators (such as settlement speed and inclination), and the differentiated maintenance strategies are matched. Predictive maintenance can reduce the cost of emergency repair and reduce long-term maintenance costs. Precise intervention can delay the deterioration of the structure and prolong the service life of the pile-slab subgrade. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a flowchart of the pile-slab subgrade detection and predictive maintenance method based on AI analysis of the present application;

[0047] Figure 2 is a framework diagram of the pile-slab subgrade detection and predictive maintenance method based on AI analysis of the present application;

[0048] Figure 3 is a flowchart of the real-time true pose of the pile column obtained by iterative optimization in an embodiment of the present application;

[0049] Figure 4 is a verticality thermal map of the pile-slab subgrade structure provided in an embodiment of the present application;

[0050] Figure 5 is a flowchart of the risk assessment on the state of the pile-slab subgrade in the next detection cycle in an embodiment of the present application;

[0051] Figure 6is a structural block diagram of a pile plate type roadbed detection and prediction maintenance system based on AI analysis. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0053] To facilitate the understanding of the present embodiment, first, a kind of pile plate type roadbed detection and prediction maintenance method based on AI analysis disclosed in the embodiments of the present application will be introduced in detail.

[0054] Reference Figure 1 And Figure 2 As shown in the figure, the method comprises:

[0055] In response to a preset detection period, historical record data, multi-source real-time data and engineering profile data of the pile plate type roadbed are acquired;The multi-source real-time data includes inspection image data, laser point cloud data, contact measurement data, temperature and humidity data;

[0056] Based on AI intelligent model analysis fusion of the multi-source real-time data, real-time structure feature data of the pile plate type roadbed is extracted;Based on the real-time structure feature data and the engineering profile data, real-time state data of the pile plate type roadbed is acquired;

[0057] A state intelligent prediction model is established, and the real-time state data and the historical record data are taken as inputs to predict the state of the pile plate type roadbed in the next detection period;

[0058] The state of the pile plate type roadbed in the next detection period is risk evaluated, and the maintenance decision is correspondingly selected and executed based on the risk evaluation result.

[0059] The present application utilizes AI intelligent model to analyze and fuse multi-source real-time data, efficiently and accurately extracts real-time structure features reflecting the health condition of the roadbed, and provides high-quality data basis for state evaluation. By constructing a state intelligent prediction model, real-time state and historical data are analyzed in time and space, and based on the deep learning intelligent model, the calculation data after executing the maintenance decision in the last period is obtained, and the data of multiple iterations and optimization is summarized, which can accurately predict the evolution trend of the roadbed state in the next detection period. On this basis, the system automatically performs risk evaluation, significantly improves the scientificity, accuracy and timeliness of the maintenance work, and provides intelligent protection for the long-term safety and durability of the roadbed structure.

[0060] Preferably, the historical record data includes historical detection data, historical maintenance data, and historical stress condition data; and the engineering profile data includes geological environment data and pile plate structure size parameters.

[0061] It should be noted that the inspection image data is a two-dimensional image, for example, an RGB image captured by a drone inspection. The laser point cloud data is a three-dimensional image, which is a set of discrete points generated by a three-dimensional scanning device, each point at least contains three-dimensional coordinate information (XYZ), and can also contain color (RGB) or reflection intensity (Intensity) attributes. These image data can macroscopically, intuitively and quickly cover a large range, directly describe the position, shape and direction changes of the pile plate structure in space, and easily find visible cracks and other surface diseases.

[0062] The contact measurement data can directly contact the structure surface or be embedded in the structure, directly measure extremely small structure changes, and is not affected by light and weather, and can be collected in real time 7x24 hours. In the scenario of the present embodiment, the measurement of physical quantities such as strain and vibration of the pile plate subgrade is included. In the present embodiment, a fiber grating sensor is used to realize direct measurement of the strain of the subgrade, to obtain the micro-strain inside the measured concrete or steel bar, and to directly reflect the stress level and load response of the structure; a vibration sensor is used to realize direct measurement of the vibration of the subgrade, to monitor the vibration caused by traffic load, construction, etc., for analyzing fatigue effect and dynamic response. An inclination sensor is used to detect the verticality change of the pile plate subgrade. Changes in environmental temperature and humidity will cause materials to expand and contract with heat and changes in water content, and temperature and humidity data can reflect the stress change and deformation of the structure.

[0063] The geological environment data can include soil properties, underground water, etc., which are the fundamental geological causes of settlement.

[0064] The image data has limited indirect measurement accuracy and is greatly affected by the environment; the contact measurement is usually point measurement, which can only obtain local data at the sensor installation position, and it is difficult to reflect the overall deformation field of the entire section. Therefore, the present application processes and fuses these multi-source and multi-type data through an AI intelligent model, complements and verifies the multi-level and multi-angle information, and after fusion, not only has continuous time series data, but also can obtain high-precision spatial details when needed, realizes full coverage in time and space dimensions, reduces the interference of external factors, and reduces the false alarm rate of the system.

[0065] Specifically, the step S2 of analyzing and fusing the multi-source real-time data based on the AI intelligent model to extract real-time structure feature data of the pile plate subgrade includes:

[0066] S21, project a pre-off-line created pile column 3D model template to a space where real-time laser point cloud is located, perform point cloud registration based on an iterative closest point algorithm, and obtain a 3D contour description of the pile column in the real-time laser point cloud;

[0067] Firstly, the position of the pile column needs to be determined in the 3D point cloud. In a specific embodiment, step S21 can be implemented by the following steps:

[0068] After the construction of the pile slab subgrade is completed, the pile column is scanned from multiple angles by a ground three-dimensional laser scanner or an airborne LiDAR to obtain a multi-station point cloud covering the entire surface thereof.

[0069] The point cloud scanned by multiple stations is unified in the same coordinate system by an iterative closest point algorithm, and noise points and outliers generated in the scanning process are removed.

[0070] A Poisson reconstruction algorithm is used to convert the point cloud into a continuous mesh model, and a high-precision 3D template model of the pile column is created and stored in the system memory. Specifically, a indicator function is estimated by the point cloud, and the isosurface of the function is the reconstructed pile column surface.

[0071] When monitoring is performed in each detection cycle, the system needs to process a large amount of real-time data of the current frame. In order to obtain the precise structure details of the pile column, an initial and approximately accurate position and attitude estimation of the pile column is needed, and then the real-time real position and attitude are further optimized.

[0072] The iterative closest point algorithm is used to register the current sparse laser point cloud with the fine 3D contour description, and the error between the two point clouds is minimized by iterative calculation to obtain the 3D contour description of the pile column in the real-time laser point cloud.

[0073] S22, based on the projection of the 3D contour description in the inspection image and the gradient change of the inspection image, the laser point cloud and the pile column in the inspection image are matched, and the real-time real position and attitude of the pile column are obtained by iterative optimization.

[0074] Due to the sparse nature of the real-time laser point cloud data (the laser point cloud obtained in the current frame is usually sparse, for example, only one side of the pile column may be scanned), the structure change details of the pile column cannot be fully reflected, and the surface of the concrete pile column is usually smooth, the texture features are few, and the geometric shape is single and repeated, which causes the matching algorithm based on traditional feature points (such as SIFT and ORB) to fail easily, and it is difficult to reconstruct the precise model structure through single-dimensional image data.

[0075] Therefore, the present application considers fusing the 3D laser point cloud data and the 2D inspection image data, providing depth information from the 3D point cloud, fusing the high-resolution information of the 2D image to improve the recognition accuracy of the weak texture pile column, distinguishing objects with the same geometric shape but different surfaces, and obtaining the accurate three-dimensional deformation and corresponding surface appearance state of the pile column.

[0076] In a specific embodiment, referring to Figure 3, step S22 can be implemented by the following steps:

[0077] S221, project the 3D contour description into the inspection image to obtain a predicted pile position, and generate a binary mask of the inspection image with respect to the predicted pile position.

[0078] It should be understood that the predicted pile position is the position of the pile generated by mapping the 3D contour description of the pile into the 2D inspection image. The GPS on the inspection UAV provides its own longitude and latitude, and the position of the inspection image capturing device and the position of the laser radar can be obtained. By coordinate conversion between the laser point cloud and the inspection image and the sensor external parameter, the approximate position and direction of the pile in the current two-dimensional image sensor coordinate system are calculated, and the 3D contour description is projected into the inspection image as the predicted pile position.

[0079] In image processing, a mask is a binary image used to specify a region of interest in an image, which is used to select, filter or operate on a specific region of the image.

[0080] In this embodiment, the white area (value 1) of the binary mask image is inside the predicted pile position (i.e. "foreground"), and the black area (value 0) is outside the predicted pile position (i.e. "background"), and the boundary between the two is the predicted pile position.

[0081] S222, calculate the minimum distance of each pixel point in the binary mask to the predicted pile position in the inspection image, wherein the sign of the minimum distance in different binary value regions is reversed.

[0082] Considering that the pile and the surrounding environment (soil, concrete pile cap) may have similar colors and weak textures, the present application extracts the geometric features of the foreground provided in the 3D laser point cloud image by step S222. When calculating the minimum distance, the minimum distance value outside the contour is set to positive, and the minimum distance value inside the contour is set to negative.

[0083] The geometric shape of the pile is regular, usually cylindrical or square, and the edges are coherent. In order to have a coherent overall edge, the global geometric information of the entire contour is encoded by a minimum distance.

[0084] S223, extract the gray scale gradient amplitude of each pixel point in the inspection image based on the gradient operator.

[0085] The collected 2D inspection image is preprocessed to calculate its gray scale gradient amplitude. In a specific embodiment, the Sobel operator can be used to extract the intensity of the real edge that can be provided in the 2D inspection image. On the premise of obtaining global geometric information, the contour information of the pile is comprehensively judged by local edge intensity information.

[0086] S224, construct a joint observation probability of the pixel point based on the gray gradient amplitude of the pixel point and the minimum distance, and calculate a feature matching loss function between the predicted pile column pose and the inspection image based on the joint observation probability.

[0087] The joint observation probability of the pixel point is a probability that the pixel belongs to the foreground or the background, which is calculated by combining the three-dimensional point cloud data and the two-dimensional inspection image data.

[0088] As a preferred embodiment, the joint observation probability of the pixel point can be represented by the following formula:

[0089] ;

[0090] wherein, represents the pixel point in the binary mask the minimum distance from the predicted pile column pose to the inspection image; represents a smoothing function, which can smoothly judge the probability that the pixel point belongs to the "foreground" according to the minimum distance value.

[0091] In the present embodiment, the smoothing function adopts the following form: which can adapt to the edge blur caused by slight jitter, so that the contour description of the laser point cloud can make accurate foreground and background judgments. Wherein, represents a hyperbolic tangent function, is an amplitude, is a slope.

[0092] represents the probability that the pixel point is both foreground in 3D and 2D, represents the probability that the pixel point is both background in 3D and 2D, when the predicted pile column pose of the point cloud perfectly coincides with the real edge in the inspection image, the joint observation probability is maximum, that is, the required accurate matching state.

[0093] is a foreground gradient model of the inspection image, specifically:

[0094] ;

[0095] wherein, represents a normalized gray gradient amplitude; is a smoothing factor.

[0096] The foreground gradient model is used to describe the probability distribution function of the gradient of the foreground region in the inspection image, represents the gradient of the pixel point in the foreground region, represents the foreground region of the image.

[0097] When the gradient approaches 1 (strong edge), the function value approaches 1, i.e. in the foreground region (near the edge of the pile), the gradient should be as large as possible. When the gradient of a certain point in the image is 0 (for example on a uniform wall), if the foreground gradient model is 0, then in the calculation of the loss function, log(0) is undefined and tends to negative infinity, which will cause program crash or meaningless results in numerical calculation. Therefore, a smoothing factor is set to prevent the foreground gradient model from becoming 0, ensuring that even in the area without gradient, there is a very small probability value, which can smoothly express the possibility that each pixel point is the foreground of the pile. That is, when the gradient is 0, the function value is very small but not zero .

[0098] is the background gradient model of the inspection image, specifically:

[0099] ;

[0100] wherein, represents the normalized gray gradient amplitude, ranging in [0, 1]; is a smoothing factor.

[0101] The background gradient model is used to describe the probability distribution function of the gradient of the background region in the inspection image, represents the gradient of the pixel point in the background region, represents the image background region. When the gradient approaches 0 (smooth region), the function value approaches 1. When the gradient is 1, the function value is minimum, i.e. in the background region, the smaller the gradient is, the better.

[0102] In this embodiment, the negative log-likelihood estimation is used to describe the feature matching loss function, i.e.

[0103] ;

[0104] wherein, represents the predicted pile pose obtained by projecting the 3D contour description into the inspection image, represents the minimum distance under the predicted pile pose . The other parameter meanings are the same as the joint observation probability formula of the pixel point, which will not be repeated here.

[0105] When the 3D contour description perfectly coincides with the real edge in the inspection image, for each pixel point :

[0106] Pixels near the contour (foreground), close to 1, and the gradient is large, so the foreground probability dominates in the joint observation probability.

[0107] Pixels far from the contour (background), close to 0, and the gradient is small, so the background probability dominates in the joint observation probability.

[0108] The true pose that makes the 3D contour description best match the true edge maximizes the overall likelihood probability, and the negative log-likelihood estimate is minimized, at which time the contour description best coincides with the edge in the image.

[0109] The present application comprehensively measures whether a pixel point is foreground (i.e., a pile) or background by combining the distance geometric features of 3D laser point cloud data and the physical gradient features of 2D inspection image data, and measures the matching degree of the edge in the inspection image and the contour description obtained by the point cloud, and the feature matching loss function is used to minimize the distance between the actual edge and the contour description, so that the 3D image and the 2D image can be accurately matched.

[0110] S225, iteratively calculate the minimum value of the feature matching loss function, and output the offset update amount of the predicted pile pose in the inspection image when the feature matching loss function reaches the minimum value.

[0111] Optimizing the predicted pile pose , that is, continuously adjusting the projection position of the 3D contour description in the inspection image, so that the contour coincides with the actual pile edge in the image.

[0112] In the specific implementation process, the minimum value of the feature matching loss function can be represented in the form of least squares, and the regression parameters are iteratively solved by using a nonlinear regression model, for example, the Levenberg-Marquardt (LM) method can be used. The real pile engineering image inevitably has noise (such as illumination variation, sensor noise) and outliers (such as weeds and cracks that are incorrectly detected as edges), and the damping mechanism of the LM method makes it less sensitive to outliers, and the optimization process is more stable.

[0113] Based on the LM method, the offset update amount .

[0114] In the formula, H represents the Hessian matrix obtained by the LM method, denotes a damping coefficient, denotes a coefficient matrix, This represents the Jacobian matrix obtained by the LM method. For the specific solution process, please refer to existing LM methods; this invention is not limited thereto.

[0115] S226, Adjust the projection position of the 3D contour description in the inspection image based on the offset update amount to obtain the real pose image of the pile, and output the real pose of the pile in real time.

[0116] The predicted pile pose is adjusted by translation and rotation based on the offset update, so that the features of the 3D image data can match the features of the 2D image data, thereby obtaining the true real-time pile pose. It should be understood that the output real-time true pose of the pile is the three-dimensional coordinates of the pile feature points.

[0117] S23 inputs real-time pose, contact measurement data, and temperature and humidity data into the AI ​​intelligent model for data fusion, and outputs real-time structural characteristic data of pile-slab roadbed.

[0118] This invention integrates the advantages of both 3D and 2D images, closely linking the 3D geometric shape information of the pile with the edge information of the 2D image. Minimum distance information provides absolute geometric constraints, clearly defining the internal / external position information of the pile, limiting the algorithm's search range to a reasonable geometric space, and preventing the image algorithm from losing its way due to texture loss. Gradient magnitude provides local, high-resolution edge responses, accurately indicating the true edge strength information in the image. The combination of these two methods effectively solves the recognition accuracy problem caused by weak texture on concrete surfaces, avoiding interference from changes in lighting, surface contamination, and motion blur in single-dimensional data. By pursuing the minimization of the matching loss function through an iterative optimization process, it can provide higher accuracy pose estimation than traditional feature point matching, thus keenly capturing subtle vertical displacement deviations and settlement.

[0119] The real-time structural feature data includes pile top coordinates, pile bottom coordinates, stress conditions, verticality data, strain data, settlement data, and structural deformation data. This data is then compared with historical data and the dimensional parameters of the pile-slab structure to perform verticality and settlement detection. Verticality detection includes angle deviation detection and displacement deviation detection; specifically, the angle deviation is measured by the pile inclination angle, and the displacement deviation is measured by the pile top coordinate deformation. Using the pile top coordinates (x0, y0, z0) as marker points, assuming a total of n pile marker points, the initial parameters for each pile i are... , For tilt sensor data, This represents the pile length under healthy baseline conditions. The parameters for this pile point at the corresponding testing time t-th testing cycle are: ;when and When the inclination angle is 1, only settlement occurs for pile i; when the inclination angle is 1, the settlement occurs for pile i. and the pile top coordinate deformation amount , the corresponding pile i has an inclination condition; when the inclination angle and the pile top coordinate deformation amount , the corresponding pile i has both settlement and inclination conditions. Among them, the pile top coordinate deformation amount .

[0120] Thus, the first real-time state data of the detection period , wherein, is the vertical angle deviation, is the vertical displacement deviation, i.e. the aforementioned pile top coordinate deformation amount, is the settlement amount, is the settlement rate. Combined with computer vision, the accurate position of the pile plate structure can be visualized, as shown in Figure 4 , which can output the verticality heat map of the pile plate structure and the verticality overrun warning and other results. In the figure, the four pile columns of the pile plate foundation are sequentially labeled as 1#, 2#, 3#, and 4#, indicates the inclination angle of the pile column, is the vertical displacement deviation.

[0121] As a preferred embodiment, the state intelligent prediction model in step S3 is:

[0122] ;

[0123] In the formula, indicates the optimal estimation based on the first detection period to predict the state of the pile plate foundation in the first detection period; indicates the real-time state data of the first detection period; indicates the real-time state data control input matrix. indicates the iterative optimization data of the historical record data based on the first detection period to the state of the pile foundation in the first detection period, i.e. on the basis of the optimal state estimation of the last period, a theoretical prediction value of the current period state is calculated through the physical evolution model of the system. This prediction value is the best guess of the current state before fusing new sensor data. indicates the historical record data state transition matrix; indicates the intelligent model calculation data after maintenance and repair in the first detection period; indicates the calculation matrix; indicates the deviation control matrix of the first detection period.

[0124] The application fuses real-time detection data, historical iterative optimization data and intelligent model calculation data of information from three different sources and different time scales, greatly overcomes the limitations of single data source or single model, and significantly improves the accuracy and reliability of the prediction of the future state of the complex roadbed system. The long-term trend prediction ability of the AI intelligent model is embedded in the short-term state prediction based on the physical model, which can actively receive external input (real-time detection data) and auxiliary information calculated by the intelligent model (deep learning prediction) to correct the prediction, realizing the effective combination of short-term physical model and long-term data-driven model.

[0125] As a preferred embodiment, the intelligent model calculation data is:

[0126] ;

[0127] In the formula, represents the real-time state data of the first detection cycle. represents the real-time state data of the first detection cycle. represents the adjustment matrix of the first detection cycle. The adjustment matrix of the first detection cycle is used to learn and extract features from historical maintenance data, and is used to calculate the system state after the last cycle of maintenance. represents the deviation control matrix of the first detection cycle. The deviation control matrix of the first detection cycle is used to correct the deviation and uncertainty in the calculation process.

[0128] The roadbed verticality detection and settlement prediction involves multiple strong nonlinear factors such as soil creep and groundwater change, and the traditional physical model is difficult to accurately describe. The AI model represented by the adjustment matrix can better capture and fit these complex nonlinear relationships and provide more accurate trend calculation data. The AI model can be independently optimized and updated without changing the core state prediction framework, improving the maintainability of the system.

[0129] As a preferred embodiment, referring to Figure 5 The risk assessment of the pile-slab roadbed state of the next detection cycle is to take the real-time state data of the pile-slab roadbed of different detection cycles as input, train an optimized neural network model to perform short-term and long-term risk trend prediction. Maintenance decisions are made based on the evaluation results, and countermeasures are provided: high risk should take emergency measures (48h), medium risk should take planned maintenance (within 2 weeks), and low risk should take routine maintenance; different risk countermeasures, the system can provide information such as required labor, equipment, budget, construction period, etc. for user selection and reference. The risk assessment level in this embodiment is shown in Table 1:

[0130] Table 1 Risk assessment level

[0131]

[0132] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of the steps is not strictly limited in order, and the steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of the steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or steps or stages in other steps.

[0133] Based on the same inventive concept, the embodiments of the present application also provide a system for implementing the above-mentioned AI analysis-based pile-slab subgrade detection and predictive maintenance. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in the following embodiments of the AI analysis-based pile-slab subgrade detection and predictive maintenance system can be referred to the limitations of the method described above, which will not be repeated here.

[0134] As Figure 6 shown, the present application also provides an AI analysis-based pile-slab subgrade detection and predictive maintenance system, comprising:

[0135] a data acquisition module for acquiring historical record data, multi-source real-time data and engineering profile data of the pile-slab subgrade in response to a preset detection period; the multi-source real-time data includes inspection image data, laser point cloud data, contact measurement data, temperature and humidity data;

[0136] an AI analysis module for analyzing and fusing the multi-source real-time data based on an AI intelligent model, extracting real-time structural feature data of the pile-slab subgrade; and acquiring real-time state data of the pile-slab subgrade based on the real-time structural feature data and the engineering profile data;

[0137] an intelligent prediction module for establishing a state intelligent prediction model, taking the real-time state data and the historical record data as inputs, and predicting the state of the pile-slab subgrade in the next detection period;

[0138] a maintenance decision module for risk assessment on the state of the pile-slab subgrade in the next detection period, and selecting a corresponding maintenance decision to be executed based on the risk assessment result.

[0139] It should be noted that the pile plate type roadbed detection and prediction maintenance system based on AI analysis provided in the embodiment is only exemplarily illustrated by the above-mentioned division of the functional modules in processing the roadbed maintenance problem, and in actual application, the above-mentioned functions can be completed by different functional modules according to needs, that is, the internal structure of the system is divided into different functional modules. Each functional module can be composed of a single execution unit, or two or more execution units can be integrated into a functional module to realize the overall function of the functional module.

[0140] Those skilled in the art can understand that each of the above modules can be realized by software, hardware and a combination thereof in whole or in part. Each of the above modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to call and execute the operations corresponding to each of the above modules by the processor.

[0141] The present application is not limited to the above specific embodiments, and various modifications made by those skilled in the art without creative labor, based on the above concept, all fall within the protection scope of the present application.

Claims

1. A pile-plate subgrade detection and predictive maintenance method based on AI analysis, characterized in that, Includes the following steps: In response to a preset detection cycle, historical data, multi-source real-time data, and project overview data of the pile-slab roadbed are acquired; the multi-source real-time data includes inspection image data, laser point cloud data, contact measurement data, and temperature and humidity data. Based on the AI ​​intelligent model analysis and fusion of the multi-source real-time data, real-time structural feature data of the pile-slab roadbed is extracted; real-time status data of the pile-slab roadbed is obtained based on the real-time structural feature data and the project overview data. A state-intelligent prediction model is established, using the real-time state data and the historical data as inputs, to predict the state of the pile-slab subgrade in the next detection cycle; A risk assessment is conducted on the condition of the pile-slab subgrade for the next inspection cycle, and maintenance decisions are selected based on the risk assessment results. The real-time structural feature data of the pile-slab subgrade extracted by analyzing and fusing the multi-source real-time data based on the AI ​​intelligent model includes: The pre-created offline 3D model template of the pile is projected onto the space of the real-time laser point cloud. Point cloud registration is performed based on the iterative nearest point algorithm to obtain the 3D contour description of the pile in the real-time laser point cloud. Based on the projection of the 3D contour description into the inspection image and the gradient change of the inspection image, the laser point cloud and the piles in the inspection image are matched, and the real-time true pose of the piles is obtained through iterative optimization. Real-time pose, contact measurement data, and temperature and humidity data are input into the AI ​​intelligent model to perform data fusion and output real-time structural feature data of pile-slab roadbed. The process of matching the laser point cloud with the piles in the inspection image based on the projection of the 3D contour description into the inspection image and the gradient changes of the inspection image, and iteratively optimizing to obtain the real-time true pose of the piles includes: The 3D contour description is projected onto the inspection image to obtain the predicted pile pose, and a binary mask of the inspection image with respect to the predicted pile pose is generated. Calculate the minimum distance from each pixel in the binary mask to the predicted pose of the pile in the inspection image, wherein the sign of the minimum distance is inverted for different binary value regions; Extract the grayscale gradient magnitude of each pixel in the inspection image based on the gradient operator; The joint observation probability of a pixel is constructed based on the gray-level gradient magnitude and the minimum distance of the pixel, and the feature matching loss function between the predicted pile pose and the inspection image is calculated based on the joint observation probability. Iteratively calculate the minimum value of the feature matching loss function, and output the offset update amount of the predicted pile pose in the inspection image when the feature matching loss function reaches the minimum value; The projection position of the 3D contour description in the inspection image is adjusted based on the offset update amount to obtain the real pose image of the pile, and the real pose of the pile is output.

2. The method for detection and predictive maintenance of pile-slab subgrade based on AI analysis according to claim 1, characterized in that, The historical data includes historical testing data, historical maintenance data, and historical stress data; the project overview data includes geological environment data and dimensional parameters of the pile-slab structure.

3. The method for detection and predictive maintenance of pile-slab subgrade based on AI analysis according to claim 2, characterized in that, The joint observation probability of a pixel constructed based on the grayscale gradient magnitude of the pixel and the minimum distance is specifically as follows: ； in, Represents pixels in a binary mask The minimum distance to predict the pose of the pile in the inspection image; This represents a smoothing function used to describe the minimum distance. With pixels The probability distribution function belonging to the foreground region; This is a foreground gradient model for inspected images, used to describe the probability distribution function of the gradient in the foreground region of the inspected image. This represents the gradient of pixels in the foreground region. Indicates the foreground region of the image; This is a background gradient model for inspection images, used to describe the probability distribution function of the gradient in the background region of the inspection image. This represents the gradient of pixels in the background region. This represents the background area of ​​the image.

4. The method for detection and predictive maintenance of pile-slab subgrade based on AI analysis according to claim 3, characterized in that, The foreground gradient model is specifically as follows: ； The background gradient model is specifically as follows: ； in, This represents the normalized grayscale gradient magnitude; This is a smoothing factor.

5. The method for detection and predictive maintenance of pile-slab subgrade based on AI analysis according to claim 1, characterized in that, The state intelligent prediction model is as follows: ； In the formula, Indicates based on the first The optimal estimate of the detection cycle for the first Prediction of the condition of pile-slab subgrade during the inspection cycle; Indicates the first Real-time status data during the detection cycle; The control input matrix represents the real-time status data. Indicates based on the first Historical data of the detection cycle for the first Iterative optimization data for detecting the subgrade condition during the detection cycle; Represents the state transition matrix of historical data; Indicates the first Data calculated by the intelligent model after maintenance and repair during the inspection cycle; Represents the calculation matrix; Indicates the first Deviation control matrix for the detection cycle.

6. The method for detection and predictive maintenance of pile-slab subgrade based on AI analysis according to claim 5, characterized in that, The first Real-time status data of the detection cycle ,in, This is the vertical angle deviation. This refers to the verticality displacement deviation. For settlement, Settlement rate; The data calculated by the intelligent model is as follows: ； In the formula, Indicates the first Real-time status data during the detection cycle; Indicates the first The adjustment matrix for the detection cycle; Indicates the first Deviation control matrix for the detection cycle.

7. The method for detection and predictive maintenance of pile-slab subgrade based on AI analysis according to claim 1, characterized in that, The risk assessment of the pile-slab subgrade status for the next inspection cycle involves using real-time status data of the pile-slab subgrade from different inspection cycles as input to train and optimize a neural network model for short-term and long-term risk trend prediction, and then implementing maintenance decisions based on the risk level: emergency measures for high-risk levels, planned maintenance for medium-risk levels, and routine maintenance for low-risk levels.

8. A pile-slab subgrade detection and predictive maintenance system based on AI analysis, characterized in that, include: The data acquisition module is used to acquire historical data, multi-source real-time data, and project overview data of the pile-slab roadbed in response to a preset detection cycle; the multi-source real-time data includes inspection image data, laser point cloud data, contact measurement data, and temperature and humidity data. The AI ​​analysis module is used to analyze and fuse the multi-source real-time data based on the AI ​​intelligent model to extract real-time structural feature data of the pile-slab roadbed; and to obtain real-time status data of the pile-slab roadbed based on the real-time structural feature data and the project overview data. The real-time structural feature data of the pile-slab subgrade extracted by analyzing and fusing the multi-source real-time data based on the AI ​​intelligent model includes: The pre-created offline 3D model template of the pile is projected onto the space of the real-time laser point cloud. Point cloud registration is performed based on the iterative nearest point algorithm to obtain the 3D contour description of the pile in the real-time laser point cloud. Based on the projection of the 3D contour description into the inspection image and the gradient change of the inspection image, the laser point cloud and the piles in the inspection image are matched, and the real-time true pose of the piles is obtained through iterative optimization. Real-time pose, contact measurement data, and temperature and humidity data are input into the AI ​​intelligent model to perform data fusion and output real-time structural feature data of pile-slab roadbed. The process of matching the laser point cloud with the piles in the inspection image based on the projection of the 3D contour description into the inspection image and the gradient changes of the inspection image, and iteratively optimizing to obtain the real-time true pose of the piles includes: The 3D contour description is projected onto the inspection image to obtain the predicted pile pose, and a binary mask of the inspection image with respect to the predicted pile pose is generated. Calculate the minimum distance from each pixel in the binary mask to the predicted pose of the pile in the inspection image, wherein the sign of the minimum distance is inverted for different binary value regions; Extract the grayscale gradient magnitude of each pixel in the inspection image based on the gradient operator; The joint observation probability of a pixel is constructed based on the gray-level gradient magnitude and the minimum distance of the pixel, and the feature matching loss function between the predicted pile pose and the inspection image is calculated based on the joint observation probability. Iteratively calculate the minimum value of the feature matching loss function, and output the offset update amount of the predicted pile pose in the inspection image when the feature matching loss function reaches the minimum value; The projection position of the 3D contour description in the inspection image is adjusted based on the offset update amount to obtain the real pose image of the pile, and the real pose of the pile is output. The intelligent prediction module is used to establish a state intelligent prediction model, taking the real-time state data and the historical data as input, to predict the state of the pile-slab subgrade in the next detection cycle. The maintenance decision module is used to conduct a risk assessment of the condition of the pile-slab subgrade in the next inspection cycle, and select and execute maintenance decisions based on the risk assessment results.

Images