Method, device and medium for three-dimensional geotechnical net laying quality evaluation and process optimization facing complex slope surface

Abstract

The application discloses a three-dimensional geotechnical net laying quality evaluation and process optimization method and device for complex slope surfaces and a medium, and relates to the technical field of artificial intelligence, and comprises the following steps: collecting laying area data in real time through a visual sensing device, generating a three-dimensional point cloud of a geotechnical net surface and a slope surface, and calculating a fitting deviation of a fitting degree; obtaining mechanical state information in laying, reversely solving a tension distribution based on a net mat mechanical constitutive relation, and then calculating a distribution index reflecting tension uniformity; fusing the fitting deviation and the tension distribution uniformity index, and outputting a comprehensive real-time construction quality score by using a pre-established quality evaluation model; taking the maximum score as a target, dynamically adjusting laying process parameters through an online optimization algorithm, and controlling construction equipment to execute, so that a perception-evaluation-optimization closed loop is formed. The application realizes online quantitative evaluation of laying quality and self-adaptive optimization of process parameters.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, equipment, and medium for quality assessment and process optimization of three-dimensional geonet laying on complex slopes. Background Technology

[0002] Three-dimensional geonets are widely used in slope protection projects. The quality of their installation, especially the tightness of the mesh against the slope and the uniformity of its own tension, directly affects the long-term protective effect of the project.

[0003] Currently, the monitoring and evaluation of such key quality indicators during construction mainly rely on on-site visual inspection and experience-based judgment by construction personnel. This method is highly subjective and makes it difficult to form a precise and unified online quantitative evaluation of requirements such as close contact with the slope and design tension, resulting in a lack of objective and consistent standards for measuring construction quality.

[0004] Furthermore, the key process parameters of the laying equipment, such as the travel speed, the movement trajectory of the laying mechanism, and the tension setting, also rely heavily on the experience of the operators for debugging. They cannot be adaptively and dynamically adjusted according to the real-time undulations of the slope and the actual mechanical state of the geonet, making it difficult to guarantee a stable and optimal laying effect under various complex working conditions. Summary of the Invention

[0005] This invention provides a method, equipment, and medium for evaluating the quality and optimizing the process of three-dimensional geonet laying on complex slopes. The technical problem it aims to solve is: how to provide a solution that can quantitatively evaluate the laying quality online and adaptively optimize the process parameters accordingly.

[0006] In a first aspect, embodiments of the present invention provide a method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes, comprising: Visual data of the paving area is collected in real time by visual sensing devices deployed on construction equipment. Based on the visual data, a first three-dimensional point cloud of the three-dimensional geonet surface and a second three-dimensional point cloud of the designed or measured slope are generated. Based on the spatial relationship between the first three-dimensional point cloud and the second three-dimensional point cloud, the fit deviation, which characterizes the degree of fit between the three-dimensional geonet and the slope, is calculated. Obtain the mechanical state information of the three-dimensional geonet during the laying process; Based on the mechanical state information and the preset mechanical constitutive relation of the geonet, the tension distribution of the three-dimensional geonet is obtained by inverse solution, and the tension distribution uniformity index, which characterizes the tension uniformity, is calculated according to the tension distribution. Based on the fit deviation and the tension distribution uniformity index, a comprehensive real-time construction quality score is output using a pre-established quality assessment model. With the goal of maximizing the real-time construction quality score, the laying process parameters are dynamically adjusted based on the current real-time construction quality score using a preset online optimization algorithm, and the adjusted process parameters are output to the construction equipment to control the execution of the construction equipment.

[0007] Optionally, the calculation of the fit deviation, which characterizes the degree of fit between the three-dimensional geonet and the slope, includes: Calculate the average Hausdorff distance between the first 3D point cloud and the second 3D point cloud; For each point in the first three-dimensional point cloud, find the nearest neighbor in the second three-dimensional point cloud, and calculate the directed distance from each point in the first three-dimensional point cloud to the corresponding nearest neighbor in the local slope normal direction. Integrate all the directed distances that are greater than zero to obtain the total volume of the gap between the first 3D point cloud and the second 3D point cloud; The normalized average Hausdorff distance is weighted and fused with the total void volume to generate the fit deviation.

[0008] Optionally, the step of weightedly fusing the normalized average Hausdorff distance with the total void volume to generate the fit deviation includes: The average curvature and rate of change of curvature of the current construction slope area are calculated based on the second three-dimensional point cloud. Determine whether the average curvature is greater than a first preset threshold and whether the rate of change of curvature is greater than a second preset threshold; If so, based on the magnitude of the curvature change rate, the weighting coefficient of the total volume of the void is dynamically calculated according to a preset linear mapping relationship, wherein the larger the curvature change rate, the larger the weighting coefficient; If not, assign a fixed basic weighting coefficient to the total volume of the voids; The normalized total void volume is weighted using the determined weighting coefficients, and the normalized average Hausdorff distance is weighted using a preset fixed weighting coefficient. The two weighted results are added together to generate the fit deviation.

[0009] Optionally, obtaining the mechanical state information of the three-dimensional geonet during the laying process includes: The visual sensing device is used to identify and track the image coordinates of the grid intersections of the three-dimensional geonet. The three-dimensional spatial displacement of the grid intersection node is calculated based on the changes in the image coordinates between consecutive frames and the calibration parameters of the visual sensing device. The three-dimensional spatial displacement is used as the mechanical state information.

[0010] Optionally, identifying and tracking the image coordinates of the grid intersections of the three-dimensional geonet includes: The original image acquired by the visual sensing device is preprocessed by grayscale conversion and contrast enhancement. On the preprocessed image, the scale-invariant feature transform algorithm is used to detect and generate key points and descriptors of the key points; Based on the preset mesh geometry and spatial topological relationship of the three-dimensional geonet, the theoretical image spacing range of the intersection nodes of adjacent meshes is calculated; Based on the theoretical image spacing range, spatial clustering and filtering are performed on all detected key points, and key points that conform to the grid distribution rules are retained as candidate grid intersection nodes; Between consecutive image frames acquired by the visual sensing device, the Euclidean distance between the descriptors of candidate grid intersection nodes is calculated, and the inter-frame correspondence is established through the nearest neighbor matching algorithm; For successfully matched point pairs, perform bidirectional consistency verification and outlier removal to obtain the image coordinate sequence of grid intersection nodes that are stably tracked.

[0011] Optionally, the step of obtaining the tension distribution of the three-dimensional geonet by inverse solution based on the mechanical state information and the preset constitutive relation of the geonet includes: The three-dimensional spatial displacement of the grid intersection nodes is used as the known displacement boundary condition; Based on the material parameters and mesh topology of the three-dimensional geonet, the system equilibrium equation with the internal forces of the elements as unknowns is established using the finite element method. Solving the equilibrium equations of the system yields the internal forces of each element in the three-dimensional geonet, thus forming the tension distribution.

[0012] Optionally, the step of calculating the tension distribution uniformity index, which characterizes the degree of tension uniformity, based on the tension distribution includes: Calculate the standard deviation of the tension distribution; Calculate the average value of the tension difference between adjacent grid cells in the tension distribution, and use it as the tension gradient index; The tension distribution uniformity index is calculated according to the formula: uniformity index = K / (1+α*standard deviation+β*tension gradient index), where K is a normalization constant and α and β are preset adjustment coefficients.

[0013] Optionally, the step of calculating and outputting a comprehensive real-time construction quality score based on the fit deviation and the tension distribution uniformity index using a pre-established quality assessment model includes: The fit deviation and the tension distribution uniformity index are normalized respectively to form an input vector; The input vector is input into a multilayer perceptron neural network model that includes an attention mechanism module; the attention mechanism module calculates the attention weight values ​​corresponding to the fit deviation and the tension distribution uniformity index, respectively. The calculated attention weight values ​​are used to weight the corresponding input features to form a weighted feature vector; The weighted feature vector is input into at least one fully connected hidden layer of the neural network model for nonlinear transformation; The transformed result is mapped to the real-time construction quality score through the output layer of the neural network model.

[0014] Optionally, the neural network model updates the weight parameters using an online learning method with feedback verification, including: During construction, data sets including the current fit deviation, the current tension distribution uniformity index, and the current process parameters are collected in real time and used as new input samples. The current fit deviation and the current tension distribution uniformity index in the new input sample are input into the neural network model to obtain the corresponding first real-time construction quality score; With the goal of maximizing the first real-time construction quality score, the current execution process parameters in the new input sample are used as initial values, and the step of dynamically adjusting the paving construction process parameters through online optimization algorithm is executed to obtain the optimized process parameters; Based on the optimized process parameters, the current fit deviation, and the current tension distribution uniformity index, the second real-time construction quality score is calculated using the quality assessment model. When the difference between the second real-time construction quality score and the first real-time construction quality score is greater than a set threshold, the current fit deviation, the current tension distribution uniformity index, and the second real-time construction quality score are used to form a new training sample pair for incremental training of the neural network model.

[0015] Optionally, the step of dynamically adjusting the paving construction process parameters based on the current real-time construction quality score, with the goal of maximizing the real-time construction quality score, includes: A kinematic and dynamic model of the construction equipment is established as a prediction model. The model predictive control problem is constructed with the optimization objective of maximizing the sum of the real-time construction quality scores within a future prediction time domain and the physical execution limit of the construction equipment as the constraint. In each control cycle, the model predictive control problem is solved in a rolling manner to obtain the optimal process parameter adjustment sequence, and the first control command in the sequence is sent to the controller of the construction equipment.

[0016] Optionally, when constructing the model to predict the control problem, the optimization objective function includes a penalty term for the rate of change of process parameters. The coefficient of the penalty term is adaptively adjusted based on the changing trend of the real-time construction quality score, including: Monitor the changes in the real-time construction quality score over multiple consecutive control cycles and calculate the slope of the score change trend. If the slope of the change trend is less than zero and the number of consecutive periods exceeds the third preset threshold, the penalty coefficient is reduced proportionally according to the magnitude of the score decrease. If the score fluctuates within a stable range and the number of fluctuation cycles exceeds a fourth preset threshold, then the penalty coefficient is increased.

[0017] Optionally, the method further includes: During construction, timestamps, fit deviations, tension distribution uniformity index, real-time construction quality scores, and actual process parameters are recorded simultaneously. The synchronously recorded data is associated with and stored in conjunction with the corresponding geographical information of the construction location and the three-dimensional model of the slope to form a digital quality archive of the three-dimensional geonet laying section.

[0018] Secondly, embodiments of the present invention also provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0019] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0020] This invention provides a method for quality assessment and process optimization of three-dimensional geonet laying on complex slopes. The method includes: real-time acquisition of visual data of the laying area using a visual sensing device deployed on construction equipment; generation of a first three-dimensional point cloud of the geonet surface and a second three-dimensional point cloud of the designed or measured slope based on the visual data; calculation of the fit deviation, representing the degree of fit between the three-dimensional geonet and the slope, based on the spatial relationship between the first and second three-dimensional point clouds; acquisition of the mechanical state information of the three-dimensional geonet during the laying process; obtaining the tension distribution of the three-dimensional geonet through inverse solving based on the mechanical state information and a preset constitutive relation of the geonet, and calculating a tension distribution uniformity index, representing the degree of tension uniformity, based on the tension distribution; calculation using a pre-established quality assessment model based on the fit deviation and the tension distribution uniformity index, and outputting a comprehensive real-time construction quality score; and dynamically adjusting the laying process parameters using a preset online optimization algorithm, with the goal of maximizing the real-time construction quality score, and outputting the adjusted process parameters to the construction equipment to control its execution. This invention achieves online synchronous quantification of core quality indicators for 3D geonet laying by real-time fusion of visual and mechanical sensor data. Furthermore, it calculates the fit deviation using the generated 3D point cloud spatial relationships and obtains the tension distribution and its uniformity index through mechanical inversion, thus transforming the subjective experience-based judgments of close contact with the slope and design tension into objective and accurate digital measurements. Further, this method constructs a closed-loop control system with real-time quality scoring as the optimization objective. This system dynamically adjusts process parameters such as laying speed, robotic arm trajectory, and tension based on the score, replacing manual adjustments relying on worker experience and achieving adaptive optimization based on local slope undulations and the actual stress state of the geonet. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a method for quality assessment and process optimization of three-dimensional geonet laying on complex slopes, provided as an embodiment of the present invention; Figure 2 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0025] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0026] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0028] Please see Figure 1 This invention provides a method for quality assessment and process optimization of three-dimensional geonet laying on complex slopes. The implementation of this invention relies on an embedded intelligent control system integrated into the construction equipment. This system includes, but is not limited to: a high-performance industrial control computer mounted on the construction equipment, a binocular stereo vision sensor, an inertial measurement unit, a GNSS positioning module, and a PLC or motion controller connected to the geonet laying actuator. The modules communicate with each other via a real-time data bus to ensure data synchronization and low-latency control. Specifically, the method includes the following steps: S1 collects visual data of the paving area in real time through a visual sensing device deployed on the construction equipment.

[0029] In practice, the visual sensing device typically employs a high-resolution binocular stereo camera, rigidly mounted on the gantry or robotic arm of the geonet-laying equipment, ensuring its field of view covers the three-dimensional geonet being laid and the slope below. The visual sensing device simultaneously acquires images from both eyes at a fixed frequency (e.g., 10Hz) to form visual data.

[0030] S2, based on the visual data, generate a first three-dimensional point cloud of the surface of the three-dimensional geonet and a second three-dimensional point cloud of the designed or measured slope, and calculate the fit deviation, which characterizes the degree of fit between the three-dimensional geonet and the slope, based on the spatial relationship between the first three-dimensional point cloud and the second three-dimensional point cloud.

[0031] In practice, the preferred visual sensing device is a calibrated binocular stereo camera. First, distortion and epipolar correction are performed on the acquired left and right eye images. Then, a semi-global stereo matching algorithm is used for pixel-level matching. The three-dimensional coordinates of each matching point are calculated using forward intersection triangulation, generating a dense first three-dimensional point cloud representing the instantaneous three-dimensional morphology of the geonet. Further, the second three-dimensional point cloud can be obtained through two methods: one is by importing a slope digital elevation model (DEM) or three-dimensional design model based on the engineering design, and aligning and cropping it according to the real-time Global Navigation Satellite System (GNSS) position of the construction equipment; the other is by scanning the exposed slope surface before construction to obtain a measured point cloud. When calculating the fit deviation, the first and second three-dimensional point clouds need to be converted to the same coordinate system, and a spatial correspondence established. Further, the fit deviation, characterizing the degree of fit between the three-dimensional geonet and the slope surface, is calculated.

[0032] In some preferred embodiments, the calculation of the fit deviation, which characterizes the degree of fit between the three-dimensional geonet and the slope, includes: calculating the average Hausdorff distance between the first three-dimensional point cloud and the second three-dimensional point cloud; for each point in the first three-dimensional point cloud, finding the nearest neighbor in the second three-dimensional point cloud, and calculating the directed distance from each point in the first three-dimensional point cloud to the corresponding nearest neighbor in the local slope normal direction; integrating all directed distances greater than zero to obtain the total void volume between the first three-dimensional point cloud and the second three-dimensional point cloud; and weightedly fusing the normalized average Hausdorff distance with the total void volume to generate the fit deviation.

[0033] In practice, the specific steps for calculating the average Hausdorff distance are as follows: For each point p in the first 3D point cloud P, calculate its Euclidean distance to all points in the second 3D point cloud Q, and take the minimum value as the distance from point p to point cloud Q. Similarly, calculate the minimum distance from each point in point cloud Q to point cloud P. Finally, take the average of these two sets of distance values ​​as the average Hausdorff distance, which reflects the degree of spatial deviation between the two point clouds as a whole.

[0034] Furthermore, when calculating the directed distance, for each point p in the first 3D point cloud, its spatial nearest neighbor q is first quickly found in the second 3D point cloud using data structures such as KD-Tree. Then, using methods such as principal component analysis, the unit normal vector n of the local slope at point q (a small plane fitted by point q and its neighboring points) is calculated. The directed distance d from point p to this local slope is the projection of vector pq onto the normal vector n, i.e., d = (pq)·n. A positive d indicates that point p is located outside the slope (i.e., in an unfitted gap), while a negative or zero d indicates that it has contacted or embedded in the slope.

[0035] Furthermore, when calculating the total void volume, all directed distance values ​​greater than zero are collected. Since the point cloud is discrete, each point represents a small region. By calculating the sum of all positive directed distance values ​​and multiplying it by the average slope area represented by a unit point estimated based on the point cloud density, the total void volume in which the first 3D point cloud is suspended above the second 3D point cloud can be approximately obtained. Alternatively, by reconstructing the second 3D point cloud as a triangular mesh surface and calculating the integral of the positive directed distance from each point in the first 3D point cloud to this surface over the corresponding projection area, the total void volume can be estimated more accurately. This is an approximate quantification of the volume of the unfitted space.

[0036] Furthermore, when generating the fit deviation, the average Hausdorff distance and the total void volume are first normalized by dividing them by their respective empirical maximum values, resulting in two dimensionless values ​​between 0 and 1. Then, preset weights are assigned to these two values ​​(e.g., distance weight 0.4, volume weight 0.6), and a weighted sum is performed. The sum is the fit deviation value; the smaller the value, the better the fit quality.

[0037] This embodiment combines two complementary geometric metrics, the average Hausdorff distance and the total void volume, to quantify bonding defects in a refined manner from different dimensions. The average Hausdorff distance effectively captures the overall offset and parallel separation between point clouds, such as when the entire mesh pad is suspended. The total void volume, on the other hand, focuses on quantifying the actual three-dimensional spatial separation volume, and is more sensitive to situations such as local bulges or the mesh pad failing to sink into the pits. By weighting and fusing the two metrics, the generated bonding deviation index can more comprehensively and robustly characterize various complex forms of non-bonding problems.

[0038] In some preferred embodiments, the step of weightedly fusing the normalized average Hausdorff distance with the total void volume to generate the fit deviation includes: calculating the average curvature and rate of change of curvature of the current construction slope area based on the second three-dimensional point cloud; determining whether the average curvature is greater than a first preset threshold and whether the rate of change of curvature is greater than a second preset threshold; if so, dynamically calculating the weight coefficient of the total void volume according to a preset linear mapping relationship based on the magnitude of the rate of change of curvature, wherein the larger the rate of change of curvature, the larger the weight coefficient; if not, assigning a fixed basic weight coefficient to the total void volume; weighting the normalized total void volume using the determined weight coefficient, and simultaneously weighting the normalized average Hausdorff distance using a preset fixed weight coefficient, and adding the two weighted results to generate the fit deviation.

[0039] In specific implementation, the steps for calculating the average curvature and rate of curvature change based on the second 3D point cloud are as follows: First, the second 3D point cloud is reconstructed or triangulated to form a slope triangular mesh model. Then, for each vertex, its Gaussian curvature or average curvature is calculated using the geometric information of its surrounding neighboring vertices. Finally, the arithmetic mean of the average curvature of all vertices within the current construction area (e.g., a 1-square-meter area below the robotic arm) is taken to obtain the average curvature of the area. Further, the rate of curvature change is obtained by calculating the difference in average curvature between adjacent sampling areas and dividing it by the time or spatial interval of area movement.

[0040] Further, the steps for determining the threshold and dynamically calculating the weight coefficients are as follows: A first threshold (e.g., a curvature value representing moderately undulating terrain) and a second threshold (representing the rate of curvature change for drastically altered terrain) are preset. The system determines in real time whether the average curvature of the current area is greater than the first threshold and whether the rate of curvature change is greater than the second threshold. If the conditions are met, it indicates that the current area is a region with complex and rapidly changing terrain. At this point, the system initiates dynamic weight calculation: a linear mapping function is established to map the rate of curvature change (between the second threshold and an upper limit) to a higher weight coefficient range (e.g., linearly increasing from a base weight of 0.6 to 0.9). This means that in areas with rapidly changing terrain, the evaluation will focus more on void volume, because the mesh is more prone to local levitation or folding due to abrupt terrain changes, and volume indicators are more sensitive to this.

[0041] Furthermore, the step of assigning basic weight coefficients is as follows: if the above judgment condition is not met, i.e., the terrain is relatively flat or changes slowly, then the importance of the average Hausdorff distance and the total void volume is considered to be relatively balanced. In this case, a fixed basic weight coefficient (e.g., 0.5) is assigned to the total void volume, while a complementary fixed weight (e.g., 0.5) is used for the average Hausdorff distance.

[0042] Furthermore, the final step in generating the fit deviation is to weight the normalized total void volume using determined weighting coefficients (dynamic or basic), weight the normalized average Hausdorff distance using corresponding fixed weights, and add the two weighted values ​​to obtain the final fit deviation.

[0043] This embodiment introduces an adaptive weight fusion mechanism based on terrain features. On gentle slopes, overall offset and local voids may be equally important. However, in complex and variable slope areas, the three-dimensional voids formed by localized gaps and folds in the mesh due to its difficulty in conforming to the terrain are a more critical factor affecting bonding quality and long-term stability. By dynamically increasing the weight of the total void volume index, the system can automatically adjust the evaluation focus to the most likely and more harmful bonding defect types in complex terrain sections.

[0044] S3, Obtain the mechanical state information of the three-dimensional geonet during the laying process.

[0045] In practical implementation, there are two parallel and mutually verifiable methods for acquiring mechanical state information. First, during the fabrication of the geonet, passive RFID tags integrating miniature strain sensors are pre-embedded at key grid nodes. During construction, the tags are read by an RFID reader on the equipment to wirelessly acquire the real-time strain values ​​of the nodes, which serve as the mechanical state information. Second, node displacement is measured using high-precision visual measurement, and the mechanical state information is calculated. It is understood that those skilled in the art can select appropriate methods for acquiring mechanical state information based on actual needs; this invention is not specifically limited to these methods.

[0046] It should be noted that, for the first embodiment based on pre-embedded sensor tags, its specific implementation method is as follows: The passive RFID tag integrating a miniature strain sensor comprises: a flexible substrate, a miniature resistance strain gauge fabricated on the substrate, and an ultra-high frequency (UHF) passive RFID chip electrically connected to and encapsulated therewith. During the geonet manufacturing stage, the tag is fixed to predetermined key grid intersections via adhesive or stitching. During construction, a UHF RFID reader antenna deployed on the construction equipment transmits radio frequency energy and scans the laying area. Upon receiving radio frequency energy, the tag activates, and its built-in sensing circuit converts the resistance change value measured by the strain gauge (corresponding to the strain at the grid node) into a digital signal, modulates it, and backscatters it back to the reader via the tag antenna. The reader demodulates the digital signal to obtain the real-time strain value at that node.

[0047] The strain value is converted into standard micro-strain (με) unit data through pre-stored tag calibration coefficients (the strain-resistance-digital value relationship obtained through calibration). This micro-strain data directly reflects the degree of deformation of the mesh line at the node and can be used as part of the mechanical state information. Furthermore, by combining the strain data of two or more tags arranged on the same mesh line with the elastic modulus of the mesh line material, the internal force (tension) of the mesh line segment can be calculated; or, by monitoring the strain changes of the mesh line tags in different directions at the same node, the displacement trend of the node can be indirectly assessed. All these strain, calculated internal force, and displacement trend information obtained from the pre-embedded tag network together constitute the mechanical state information used for subsequent reverse engineering. This embodiment is independent of the aforementioned visual measurement method and can also be used as a redundant or verification data source, fused with the visual measurement results to improve the robustness and reliability of mechanical state perception.

[0048] Furthermore, in some preferred embodiments, obtaining the mechanical state information of the three-dimensional geonet during the laying process includes: identifying and tracking the image coordinates of the grid intersection nodes of the three-dimensional geonet through the visual sensing device; calculating the three-dimensional spatial displacement of the grid intersection nodes based on the changes in the image coordinates between consecutive frames and the calibration parameters of the visual sensing device; and using the three-dimensional spatial displacement as the mechanical state information.

[0049] In practice, when identifying and tracking the image coordinates of grid intersections, the visual sensing device (such as the aforementioned binocular camera) continuously acquires high-frame-rate video streams of the laid area. First, image processing algorithms (such as edge detection and corner detection) are used to initially identify candidate pixel regions that may be grid intersections in each frame. These intersections typically exhibit regular, high-contrast "+" or "X" shaped features. Then, using feature tracking algorithms (such as KLT optical flow), based on the similarity of image grayscale or feature descriptors between consecutive frames, the pixel coordinate positions of the same grid intersection in different frames are predicted and determined, forming an image coordinate trajectory that changes over time.

[0050] Furthermore, when calculating the 3D spatial displacement, since a calibrated binocular vision system is used, the internal parameters (focal length, principal point, etc.) and relative positional relationship (extrinsic parameters) of the two cameras are known. For the same grid intersection node that is successfully matched in the left and right image, the 3D spatial coordinates of the node at the current moment can be calculated based on its left and right image coordinates using the forward intersection principle. By comparing the 3D coordinates of the node at adjacent time points (e.g., 0.1 seconds apart), its 3D spatial displacement vector (Δx, Δy, Δz) within this time interval can be calculated.

[0051] Furthermore, using three-dimensional spatial displacement as mechanical state information means that the system records and outputs not just single image coordinates, but displacement quantities with physical length units obtained through geometric calculations. This displacement quantity directly reflects the motion of the geonet nodes in three-dimensional space and is the essential basic data for subsequent mechanical inversion analysis.

[0052] This invention provides an effective, non-contact, and low-cost method for acquiring the mechanical state information of geonets. Utilizing existing visual sensing devices on construction equipment, it eliminates the need for additional physical sensors (such as strain gauges) on the geonet, avoiding the increased costs, construction interference, and potential damage associated with sensor deployment. By tracking specific feature points from image sequences and calculating their three-dimensional displacement, this method transforms the visually visible deformation information of the geonet into precise, quantifiable three-dimensional motion data. This displacement data directly reflects the dynamic deformation response of the geonet during installation and is a key input for revealing its internal stress state.

[0053] In some preferred embodiments, the identification and tracking of the image coordinates of the grid intersection nodes of the three-dimensional geonet includes: preprocessing the original image acquired by the visual sensing device by converting it to grayscale and enhancing its contrast; detecting and generating key points and descriptors of the key points on the preprocessed image using a scale-invariant feature transformation algorithm; calculating the theoretical image spacing range of adjacent grid intersection nodes based on the preset mesh geometry and spatial topological relationship of the three-dimensional geonet; performing spatial clustering and filtering on all detected key points according to the theoretical image spacing range, retaining key points that conform to the grid distribution rules as candidate grid intersection nodes; calculating the Euclidean distance between the descriptors of candidate grid intersection nodes between consecutive image frames acquired by the visual sensing device, and establishing inter-frame correspondence through a nearest neighbor matching algorithm; performing bidirectional consistency verification and outlier removal on successfully matched point pairs to obtain a sequence of image coordinates of stably tracked grid intersection nodes.

[0054] In practice, the image preprocessing steps include converting the original color image into a grayscale image and using histogram equalization or contrast-limited adaptive histogram equalization algorithms to enhance image contrast, thereby improving the robustness of subsequent feature detection under uneven lighting conditions.

[0055] Furthermore, when using the SIFT algorithm to detect keypoints, the algorithm searches for extreme points in the image at different scales and filters out low-contrast points and unstable edge response points. For each retained keypoint, the SIFT algorithm calculates its position, scale, and orientation, and generates a 128-dimensional feature descriptor vector that describes the gradient direction distribution in the keypoint's neighborhood.

[0056] Furthermore, when calculating the theoretical image spacing range, it is necessary to know the actual physical dimensions (such as side length) of the 3D geonet mesh and the imaging scale (pixels / meter) of the visual sensing device at the current distance. Based on the physical dimensions of the mesh and the imaging scale, the theoretical pixel distance range that adjacent mesh intersections should appear on the image plane can be calculated (for example, under normal laying tension, the distance should be between 30 and 50 pixels).

[0057] Further, the spatial clustering and filtering steps are as follows: For all SIFT keypoints in the image, cluster analysis is performed based on their pixel coordinates and the theoretical image spacing range. Keypoint groups whose distances conform to the theoretical range and can form a regular grid pattern are filtered out. For keypoints within each group, their spatial arrangement is further verified to ensure it conforms to the grid topology (e.g., approximately in rows and columns). Finally, keypoints that pass the verification are retained as high-confidence candidate grid intersection nodes.

[0058] Furthermore, when establishing inter-frame correspondences, in two consecutive frames, the Euclidean distance between the SIFT descriptor of a candidate node in the current frame and the descriptors of all candidate nodes in the next frame is calculated. The nearest neighbor matching principle is used; if the ratio of the nearest neighbor distance to the second nearest neighbor distance is less than a threshold (e.g., 0.8), the match is considered successful. Bidirectional consistency verification means that matching is performed not only from frame A to frame B, but also from frame B to frame A, retaining only mutually matched point pairs to eliminate false matches.

[0059] This invention significantly improves the robustness of accurate identification and stable tracking of grid intersection nodes in complex construction scenarios by introducing spatial regularity constraints based on known mesh geometry. In construction environments, there may be interfering objects (such as stones, shadows, and vegetation) that resemble the visual features of grid nodes. Simply relying on general feature matching algorithms like SIFT can easily lead to a large number of false detections and mismatches. By utilizing prior knowledge of mesh size and topology, spatial clustering and regularity filtering of detected key points can effectively filter out most false feature points that do not conform to the mesh distribution pattern, thus focusing candidate targets on genuine grid intersection nodes. This greatly improves the purity and reliability of the feature point set.

[0060] S4. Based on the mechanical state information and the preset mechanical constitutive relationship of the geonet, the tension distribution of the three-dimensional geonet is obtained by inverse solution, and the tension distribution uniformity index, which characterizes the tension uniformity, is calculated according to the tension distribution.

[0061] In practice, the pre-defined constitutive relation of the geogrid is a physical model describing the stress-strain relationship of the three-dimensional geogrid wires, typically simplified to a linear elastic model. Using the read nodal strain information as input, and combining it with the known material elastic modulus, cross-sectional area of ​​the geogrid wires, and mesh topology, a system of linear equations can be constructed with the internal forces of the geogrid wires as unknowns. By solving this system of equations, the internal forces of all geogrid elements, i.e., the tension distribution, can be derived.

[0062] In some preferred embodiments, the step of obtaining the tension distribution of the three-dimensional geonet by inverse solution based on the mechanical state information and the preset constitutive relation of the geonet includes: using the three-dimensional spatial displacement of the grid intersection nodes as known displacement boundary conditions; establishing a system equilibrium equation with the internal forces of the elements as unknowns using the finite element method based on the material parameters and grid topology of the three-dimensional geonet; solving the system equilibrium equation to obtain the internal forces of each grid element of the three-dimensional geonet, thus constituting the tension distribution.

[0063] In practice, taking the three-dimensional spatial displacement of the grid intersection nodes as a known displacement boundary condition means that when constructing the mechanical model, these displacement components of specific nodes in three-dimensional space obtained by visual measurement (e.g., 2 mm in the X direction and -1 mm in the Z direction) are taken as known conditions that must be satisfied when solving the model.

[0064] Furthermore, when establishing the system equilibrium equations, the structure of the three-dimensional geonet is first discretized. Based on its mesh topology (such as quadrilateral or hexagonal mesh), each segment of the mesh is considered as a rod element that can only withstand axial forces or a beam element that considers bending. According to the parameters of the geonet material, such as the elastic modulus E and Poisson's ratio μ (for the isotropic assumption), combined with the cross-sectional area A of the mesh, the stiffness matrix of each element can be calculated. Then, according to the assembly rules of the finite element method, the stiffness matrices of all elements are integrated according to their connection relationship to form the overall stiffness matrix [K] describing the entire geonet structure. The equilibrium equation of the system is [K]{U}={F}, where {U} is the displacement vector of all nodes (partially known, partially unknown), and {F} is the external force vector of the nodes (usually the unknown internal force to be solved or known to be zero). Since visual measurements provide displacement values ​​of some nodes (at displacement boundary conditions), these known displacements can be substituted into the equations, thereby modifying the equations into a solvable form with the remaining unknown displacements and all element internal forces as unknowns.

[0065] Furthermore, after solving the system equilibrium equations, the internal force values ​​of each geogrid element can be obtained. Organizing and visualizing the internal force values ​​of all elements according to their spatial location creates a tension distribution map reflecting the stress state inside the geogrid. This distribution map visually shows which areas are under high tension (potentially too tight) and which areas have low tension (potentially loose).

[0066] The core of this embodiment lies in applying the finite element method from solid mechanics to the construction process analysis of flexible mesh structures. By using the displacement data of a finite number of discrete nodes obtained through visual measurement as boundary conditions, it constructs a mechanical model covering the entire mesh using the physical constraints of the material constitutive relationship and topological connectivity of the mesh itself. Solving this model allows for the deduction of globally invisible tension from locally visible displacements. This method achieves a state mapping from apparent deformation to internal force, overcoming the difficulties of directly measuring mesh tension, which requires densely deployed sensors and struggles to obtain continuously distributed information.

[0067] In some preferred embodiments, the step of calculating the tension distribution uniformity index, which characterizes the uniformity of tension, based on the tension distribution includes: calculating the standard deviation of the tension distribution; calculating the average value of the tension difference between adjacent grid cells in the tension distribution as a tension gradient index; and calculating the tension distribution uniformity index according to the formula: uniformity index = K / (1 + α * standard deviation + β * tension gradient index), where K is a normalization constant and α and β are preset adjustment coefficients.

[0068] In practice, when calculating the standard deviation of the tension distribution, the internal force values ​​(tension values) of all the network cable elements obtained from the solution are first treated as a dataset. The arithmetic mean μ of this dataset is then calculated. Next, the square of the deviation of each tension value from the mean μ is calculated, the average of these squared values ​​is taken, and finally the square root is taken to obtain the standard deviation σ. The standard deviation reflects the degree of dispersion of all network cable tension values ​​around the mean; the larger the standard deviation, the more uneven the tension distribution.

[0069] Furthermore, when calculating the tension gradient index, it is necessary to determine which grid elements are adjacent (e.g., elements sharing the same node) based on the geonet's grid topology. For each pair of adjacent grid elements, the absolute difference in their internal force values ​​is calculated. Then, the arithmetic mean of the tension differences for all such pairs of adjacent elements is calculated; this average is the tension gradient index. This index reflects the degree of tension variation in the geonet space; a larger index value indicates a greater tension difference between adjacent areas, and a possible abrupt change in the distribution.

[0070] Furthermore, when calculating the tension distribution uniformity index using the formula, K is a normalization constant, usually set to 1, to ensure the final index falls within the range of 0 to 1; a larger value indicates greater uniformity. α and β are two preset positive adjustment coefficients used to balance the influence of the standard deviation and gradient index on the final uniformity index. For example, if more attention is paid to overall dispersion, α can be increased appropriately; if more attention is paid to local abrupt changes, β can be increased appropriately. Substituting the calculated standard deviation σ and the tension gradient index into the formula yields a value between 0 and 1, which is the tension distribution uniformity index.

[0071] This embodiment constructs a comprehensive evaluation index for tension distribution uniformity by combining two statistical measures: standard deviation and tension gradient index. Standard deviation alone can only measure the overall dispersion of tension values, but it cannot distinguish between two different non-uniformity patterns: one where tension values ​​are generally high or low but change gradually overall; and the other where tension values ​​experience drastic jumps in local areas, forming obvious stress concentration zones. The latter is more detrimental to the local damage and long-term performance of the mattress. The tension gradient index is introduced precisely to capture the latter non-uniformity pattern. By calculating the average tension difference between adjacent units, it can effectively quantify the smoothness or abruptness of the tension distribution. These two complementary indices are fused into a single uniformity index through a formula, enabling this index to more comprehensively and sensitively reflect the quality of the tension distribution. It not only indicates whether the tension is uniform but also reveals, to some extent, the specific form of non-uniformity, providing the quality assessment model with richer and more discriminative characteristic information about the tension state.

[0072] S5. Based on the fit deviation and the tension distribution uniformity index, a pre-established quality assessment model is used to calculate and output a comprehensive real-time construction quality score.

[0073] In practice, when establishing and utilizing a quality assessment model, this model can be a multilayer perceptron pre-trained using historical construction data. The inputs to the quality assessment model are the normalized fit deviation and tension distribution uniformity index. By learning the complex nonlinear relationship between these two indicators and expert experience scores, the quality assessment model can output a real-time construction quality score between 0 and 100.

[0074] Furthermore, the pre-established quality assessment model is trained by collecting historical construction or simulation data. Specifically, a large amount of data covering fit deviation and tension distribution uniformity index under different slope topography and paving techniques is collected, and experts or those based on preset quality standards comprehensively score these data to form a training sample set. This sample set is then used to supervise the training of a multilayer perceptron neural network incorporating an attention mechanism until the error between its predicted score and the standard score converges to a predetermined range.

[0075] In some preferred embodiments, the step of calculating and outputting a comprehensive real-time construction quality score based on the fit deviation and the tension distribution uniformity index using a pre-established quality assessment model includes: normalizing the fit deviation and the tension distribution uniformity index to form input vectors; inputting the input vectors into a multilayer perceptron neural network model containing an attention mechanism module; calculating the attention weight values ​​corresponding to the fit deviation and the tension distribution uniformity index through the attention mechanism module; weighting the corresponding input features using the calculated attention weight values ​​to form a weighted feature vector; inputting the weighted feature vector into at least one fully connected hidden layer of the neural network model for nonlinear transformation; and mapping the transformed result to the real-time construction quality score through the output layer of the neural network model.

[0076] In practice, the fit deviation and tension distribution uniformity index are normalized by linearly mapping their values ​​to the standard intervals of [0,1] or [-1,1] to eliminate the influence of differences in dimensions and numerical ranges on the model, forming a two-dimensional input vector, such as [N(T),N(S)], where N() represents the normalization function, T is the fit deviation, and S is the tension distribution uniformity index.

[0077] Furthermore, when the input vector is fed into an MLP containing an attention mechanism module, the input layer of that MLP receives this two-dimensional vector. The attention mechanism module is typically located after the input layer and before the first fully connected layer. This module contains a trainable parameter matrix used to calculate the importance weights of the two input features. Specifically, it calculates an attention weight W_T for the normalized fit bias N(T) and an attention weight W_S for the normalized tension distribution uniformity index N(S). The sum of these two weights is typically 1, and its magnitude is automatically learned by the model during training based on which feature contributes more to the final score prediction.

[0078] Furthermore, when forming the weighted feature vector, the original input features are multiplied by their corresponding attention weights: N'(T) = W_T*N(T), N'(S) = W_S*N(S). The vector composed of [N'(T), N'(S)] is the weighted feature vector. This means that the model dynamically adjusts the contribution ratio of different quality dimensions in this score.

[0079] Furthermore, during the nonlinear transformation, the weighted feature vectors are fed into one or more fully connected hidden layers. Each hidden layer consists of multiple neurons, which apply a nonlinear activation function (such as the ReLU function) to the weighted sum of the inputs. Through multiple layers of nonlinear transformation, the model is able to learn the influence of the complex, nonlinear coupling between fit and tension uniformity on the final quality.

[0080] Furthermore, during the output layer mapping, the output of the last hidden layer is connected to an output layer neuron. This neuron typically uses a linear activation function or a sigmoid function (constraining the output between 0 and 1, then mapping it to a score between 0 and 100) to map the high-dimensional features into a scalar value, i.e., a real-time construction quality score.

[0081] This embodiment introduces an attention mechanism, enabling the quality assessment model to dynamically perceive the importance of features. Under different construction scenarios, the weights of fit and tension uniformity on overall quality are not constant. For example, on extremely steep slopes, ensuring tight fit to prevent slippage may be more crucial than absolute tension uniformity; while in flat areas, uniform tension distribution may be key to preventing localized bulging. Traditional fixed-weight fusion models cannot adapt to such variations. The attention mechanism allows the model to automatically infer and assign the relative importance of two features in the current score based on the specific values ​​of each input. This allows the model to more flexibly and intelligently integrate multi-dimensional quality information, and its output score better reflects the core quality contradictions under specific working conditions, resulting in more reasonable and targeted assessment results.

[0082] In some preferred embodiments, the neural network model updates its weight parameters using an online learning method with feedback verification, including: during construction, collecting data sets containing the current fit deviation, the current tension distribution uniformity index, and the current execution process parameters in real time as new input samples; inputting the current fit deviation and the current tension distribution uniformity index from the new input samples into the neural network model to obtain a corresponding first real-time construction quality score; with the goal of maximizing the first real-time construction quality score, using the current execution process parameters from the new input samples as initial values, executing the step of dynamically adjusting the paving construction process parameters through an online optimization algorithm to obtain optimized process parameters; based on the optimized process parameters, the current fit deviation, and the current tension distribution uniformity index, calculating a second real-time construction quality score through the quality assessment model; when the difference between the second real-time construction quality score and the first real-time construction quality score is greater than a set threshold, constructing a new training sample pair using the current fit deviation, the current tension distribution uniformity index, and the second real-time construction quality score, and incrementally training the neural network model.

[0083] In practice, when collecting new input samples in real time, the system packages the fit deviation value, tension distribution uniformity index value, and process parameters (such as speed v, robotic arm posture x, y, z, tension F) calculated at the current moment into a data set (T, S, v, x, y, z, F) in each control cycle or every few cycles.

[0084] Furthermore, when the first real-time construction quality score is obtained, T and S from this data set are input into the currently deployed quality assessment model (MLP), and the model outputs a score value Q1.

[0085] Furthermore, when performing the optimization steps to obtain the optimized parameters, Q1 is used as the current quality feedback, and the model predictive control algorithm is invoked. The algorithm takes the currently executed process parameters (v,x,y,z,F) as the starting point for optimization, predicts how the Q value will change if the parameters are adjusted in the future, and solves for a set of new process parameters (v,x,y,z,F*) that maximizes the total predicted score. This set of parameters represents a better solution calculated through simulation.

[0086] Furthermore, when calculating the second real-time construction quality score, since this is a simulation optimization process, the actual T and S values ​​after adopting the new parameters cannot be obtained. Therefore, in a simplified implementation, it can be assumed that the terrain and mattress state remain unchanged in a short period of time. The new parameters (v,x,y,z,F*) are combined with the original T and S and input again into the quality assessment model to obtain a new, simulated score value Q2. Q2 reflects the expected quality that the model believes can be achieved after optimization.

[0087] Furthermore, when constructing new training samples and performing incremental training, the system compares the difference between Q2 and Q1. If the difference is greater than a set threshold (e.g., 5 points), it indicates a significant inconsistency between the model's current evaluation and its self-driven optimization results, meaning the model believes it "can do better." In this case, the original T, S, and the higher score Q2 are used as a new training sample pair (input: [T, S], output label: Q2). Using this sample pair, a mini-batch incremental training is performed on the MLP model (e.g., fine-tuning the weights using stochastic gradient descent), allowing the model to gradually learn that a higher quality score Q2 is possible and reasonable given feature inputs like (T, S). This process enables the model to use its own generated optimization results as supervisory signals, continuously improving and calibrating itself.

[0088] This embodiment creates a self-evolving model optimization mechanism. Traditional machine learning models are typically static after deployment, and their performance is limited by the coverage of the training data. In complex and ever-changing construction environments, new working conditions are often encountered that are not fully covered by the training data, leading to potential biases or inaccuracies in model evaluation. This embodiment generates a high-quality feedback signal by comparing the model's own output (score) with the "better" expected result generated after optimization based on that output. When the model underestimates its optimization potential (i.e., Q2 is much larger than Q1), this signal drives the model to perform online incremental learning. This essentially allows the model to continuously "self-criticize" and "self-improve" during the construction process. In this way, the model can gradually adapt to new and unknown construction conditions, correct its evaluation bias, and make its scoring criteria increasingly closer to the optimal construction results that can be achieved in practice. This significantly improves the adaptability, accuracy, and reliability of the quality assessment model in long-term use.

[0089] S6, with the goal of maximizing the real-time construction quality score, based on the current real-time construction quality score, dynamically adjust the paving construction process parameters through a preset online optimization algorithm, and output the adjusted process parameters to the construction equipment to control the execution of the construction equipment.

[0090] In practice, when adjustments are made using a preset online optimization algorithm, this online optimization algorithm can be model predictive control. The online optimization algorithm aims to maximize the sum of predicted quality scores over several future control cycles, using adjustable parameters of the construction equipment (such as tracked chassis and multi-axis robotic arms) (such as travel speed, joint angles, and end effector tension) as optimization variables, and the physical limits of the equipment (maximum speed, torque limits) as constraints. It solves for the optimal combination of process parameters in real time and sends the solution to the equipment controller for execution.

[0091] This invention, by integrating visual and mechanical sensor information, achieves for the first time online synchronous quantitative evaluation of core quality indicators (fit and tension uniformity) in three-dimensional geonet construction. Visual data directly reflects the macroscopic spatial fit between the geonet and the slope, while mechanical information directly relates to the internal stress state of the geonet. This combination overcomes the limitations of single-sensory evaluation. Based on this quantitative evaluation result, the method constructs a closed-loop control system with real-time construction quality scoring as the optimization objective. The system no longer relies on the subjective judgment and experience of operators, but instead automatically and continuously fine-tunes process parameters through online optimization algorithms, driving the construction quality score closer to the optimal value. This closed-loop process essentially transforms construction quality control from an open-loop, experience-dependent model to a closed-loop, data-driven automatic optimization model, thereby ensuring high-quality and highly consistent laying results dynamically maintained on slopes of varying complexity. This effectively solves the core technical problems of inaccurate manual evaluation, delayed parameter adjustment, and reliance on experience.

[0092] In some preferred embodiments, the step of dynamically adjusting the laying construction process parameters based on the current real-time construction quality score, with the goal of maximizing the real-time construction quality score, includes: establishing a kinematic and dynamic model of the construction equipment as a prediction model; constructing a model predictive control problem with the physical execution limit of the construction equipment as a constraint condition, taking the maximization of the sum of the real-time construction quality scores in a future prediction time domain as the optimization objective; and solving the model predictive control problem in each control cycle to obtain the optimal process parameter adjustment sequence, and sending the first control command in the sequence to the controller of the construction equipment.

[0093] In practical implementation, establishing kinematic and dynamic models of construction equipment requires system modeling of the equipment. For example, for a net-laying machine consisting of a tracked chassis, a multi-joint robotic arm, and an end effector tensioning roller, its kinematic model describes the geometric relationship between the angles of each joint of the robotic arm and the spatial position and attitude of the end effector. Its dynamic model further describes the relationship between the acceleration, velocity, and displacement of each joint and chassis under the action of the motor driving torque, and considers factors such as the mass, inertia, and friction of the mechanism. These models can be obtained through theoretical derivation combined with system identification experiments.

[0094] Furthermore, when constructing the model predictive control problem, the objective function J is defined as the negative sum of the predicted real-time construction quality scores over the next N control cycles (prediction time domain) (because a minimization problem is usually solved), i.e., J = -Σ(Q_pred_k), k = 1 to N. Here, Q_pred_k is the score predicted in the k-th step, which needs to be predicted through the correlation between the equipment model, process parameters, and quality assessment model. Constraints include: upper and lower limits of chassis speed, angle and angular velocity limits of each joint of the robotic arm, and physical execution limits such as the maximum and minimum tension that the end effector can provide.

[0095] Furthermore, during the rolling solution and instruction issuance, in each control cycle (e.g., 0.1 seconds), the MPC controller, starting from the currently measured construction quality score and equipment status, solves an optimization problem online under the constraints, obtaining the optimal process parameter adjustment sequence [u_1, u_2, ..., u_N] for the next N steps. Then, only the first control instruction u_1 in this sequence (e.g., increasing the speed by 0.05 m / s, increasing the angle of robotic arm joint 2 by 0.5 degrees) is immediately issued to the underlying controller of the construction equipment (e.g., PLC or motion control card) for execution. In the next control cycle, the system starts from the new state and repeats the above prediction and optimization process, realizing "rolling optimization and real-time control".

[0096] This embodiment applies advanced model predictive control theory to the optimization of geonet laying process, realizing proactive control based on forward-looking prediction. Unlike classical feedback control (such as PID) that only reacts to the current error, MPC utilizes the dynamic model of the equipment to predict the trajectory of construction quality scores under different control strategies over a future period. By solving a finite-time optimization problem online, it can find a suitable set of control commands, considering not only immediate improvement but also the overall optimal effect over multiple future steps. This forward-looking approach makes control actions smoother and more coordinated, avoiding overshoot or oscillations caused by short-sighted decisions. For example, when a pothole is predicted ahead, MPC may slightly reduce tension and adjust the robotic arm trajectory in advance for a smooth transition, rather than reacting drastically after the score drops sharply due to hitting the pothole.

[0097] In some preferred embodiments, when constructing the model predictive control problem, the optimization objective function includes a penalty term for the rate of change of process parameters. The coefficient of the penalty term is adaptively adjusted based on the changing trend of the real-time construction quality score, including: monitoring the changes in the real-time construction quality score over multiple consecutive control cycles and calculating the slope of the score's changing trend; if the slope of the changing trend is less than zero and the number of cycles exceeds a third preset threshold, the coefficient of the penalty term is reduced proportionally according to the magnitude of the score decrease; if the score fluctuates within a stable range and the number of cycles of fluctuation exceeds a fourth preset threshold, the coefficient of the penalty term is increased.

[0098] In practice, when monitoring score changes and calculating the trend slope, the system records the real-time construction quality scores Q(t) for the most recent M control cycles (e.g., M=10). Linear regression is performed on these data points to fit a straight line, and the slope k of this line represents the trend slope of the score change. k>0 indicates that the score is rising, k<0 indicates that the score is falling, and the magnitude of |k| indicates the rate of change.

[0099] Furthermore, the process of determining the conditions and adjusting the penalty term coefficient is as follows: First, a negative third threshold (e.g., -0.5 points / cycle) and a fourth threshold (e.g., 5 consecutive cycles) are preset. The additional penalty term in the optimization objective function is usually in the form of λ*Σ(Δu)^2, where Δu is the change in process parameters and λ is the penalty term coefficient.

[0100] Furthermore, when the calculated trend slope k is less than the third threshold (i.e., significantly negative) and this downward trend persists for more than the third preset threshold number of cycles (e.g., k is less than -0.5 for three consecutive cycles), the system determines that the current score is rapidly deteriorating. At this point, based on the magnitude of the score decline (e.g., the total score decrease over the past three cycles), the value of λ is reduced by a preset scaling factor (e.g., λ decreases by 0.1 for every 1-point decrease). Reducing λ means relaxing the restrictions on the magnitude of parameter changes, allowing the controller to take larger, more aggressive adjustments to reverse the trend of quality deterioration as quickly as possible.

[0101] Furthermore, when the score fluctuates within a stable range (e.g., [85, 90] points) and this fluctuation persists for more than four preset threshold cycles (e.g., eight consecutive cycles), the system determines that it is currently near a relatively optimal stable operating point. At this point, the value of λ is increased. Increasing λ means strengthening the penalty for parameter changes, encouraging the controller to make smaller, more cautious adjustments, avoiding disturbances introduced by unnecessary parameter fluctuations near the stable point, thereby maintaining control smoothness and system stability.

[0102] This embodiment cleverly resolves the inherent contradiction between response speed and stability in a control system by adapting the penalty term coefficient λ to the changing trend of construction quality. A fixed λ value leads to a trade-off: if λ is too large, the system response is sluggish, making it difficult to quickly correct quality declines; if λ is too small, the system is overly sensitive and prone to unnecessary oscillations when approaching the optimal value. This method intelligently switches control strategies by analyzing the scoring trend in real time. When a significant deterioration in quality is detected, it automatically reduces λ, essentially granting the controller temporary emergency handling authority, enabling it to take stronger intervention measures and respond quickly to adverse changes, demonstrating control agility. When the system is operating in a high-quality, stable region, it automatically increases λ, equivalent to entering a fine-tuning mode, suppressing frequent and large changes in control commands, and avoiding disruption of the achieved good balance due to over-adjustment, demonstrating control robustness.

[0103] In some preferred embodiments, the method further includes: during construction, simultaneously recording timestamps, fit deviations, tension distribution uniformity index, real-time construction quality scores, and actual process parameters; and associating and storing the synchronously recorded data with the corresponding construction location geographic information and slope three-dimensional model to form a digital quality archive of the three-dimensional geonet laying section.

[0104] In practice, a data logging service runs on the industrial control computer or embedded system of the construction equipment. This service is synchronized with modules such as vision processing, mechanical inversion, quality assessment, and optimization control. At the end of each data processing cycle (e.g., every 0.1 seconds or after completing one laying action cycle), it collects and packages the following data: the current timestamp, the calculated fit deviation value, the tension distribution uniformity index value, the real-time construction quality score value, and the process parameter values ​​(such as speed, position, and tension setpoints) issued to the equipment and actually executed at this moment. All of this data is packaged into a structured data record.

[0105] Furthermore, when forming a digital quality archive through associated storage, each of the aforementioned data records is also bound to the current latitude and longitude coordinates provided by the GNSS system of the construction equipment. Simultaneously, this record is also indexed and associated with the corresponding second-dimensional point cloud (i.e., the 3D model of the slope at the current construction location) used to calculate the fit deviation at that moment. All this information—time-series data, geographical location, and slope model—is stored in a structured database or file system. Each roll or section of the completed 3D geonet corresponds to a complete digital quality archive composed of multiple consecutive data records. This archive can be exported to a common format (such as JSON or database backup) for long-term storage and subsequent analysis.

[0106] This invention constructs a complete digital twin record of three-dimensional geonet laying by systematically collecting, associating, and storing key data throughout the entire construction process. Traditional construction only retains the final acceptance results, completely lacking process data. The archive recorded by this method not only includes the final quality score but also covers the entire dynamic process involved in achieving that quality: including quantitative quality indicators at each moment (fit deviation, tension distribution uniformity index), specific process parameters adopted to achieve that quality, and the precise spatial location of these events and the corresponding three-dimensional slope morphology. This high-fidelity, multi-dimensional, spatiotemporally correlated data record completely makes the previously invisible, black-box construction process transparent and data-driven.

[0107] Please see Figure 2 , Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0108] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0109] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it enables the processor 502 to execute a method for quality assessment and process optimization of three-dimensional geonet laying on complex slopes.

[0110] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0111] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for quality assessment and process optimization of three-dimensional geonet laying on complex slopes.

[0112] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0113] The processor 502 is used to run the computer program 5032 stored in the memory to implement the steps of the three-dimensional geonet laying quality assessment and process optimization method for complex slopes proposed in any of the above embodiments.

[0114] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0115] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0116] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the steps of the method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes proposed in any of the above embodiments.

[0117] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0118] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0119] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0120] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0121] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0122] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0123] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for quality assessment and process optimization of three-dimensional geonet laying on complex slopes, characterized in that, include: Visual data of the paving area is collected in real time by visual sensing devices deployed on construction equipment. Based on the visual data, a first three-dimensional point cloud of the three-dimensional geonet surface and a second three-dimensional point cloud of the designed or measured slope are generated. Based on the spatial relationship between the first three-dimensional point cloud and the second three-dimensional point cloud, the fit deviation, which characterizes the degree of fit between the three-dimensional geonet and the slope, is calculated. Obtain the mechanical state information of the three-dimensional geonet during the laying process; Based on the mechanical state information and the preset mechanical constitutive relation of the geonet, the tension distribution of the three-dimensional geonet is obtained by inverse solution, and the tension distribution uniformity index, which characterizes the tension uniformity, is calculated according to the tension distribution. Based on the fit deviation and the tension distribution uniformity index, a comprehensive real-time construction quality score is output using a pre-established quality assessment model. With the goal of maximizing the real-time construction quality score, the laying process parameters are dynamically adjusted based on the current real-time construction quality score using a preset online optimization algorithm, and the adjusted process parameters are output to the construction equipment to control the execution of the construction equipment.

2. The method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes according to claim 1, characterized in that, The calculation of the fit deviation, which characterizes the degree of fit between the three-dimensional geonet and the slope, includes: Calculate the average Hausdorff distance between the first 3D point cloud and the second 3D point cloud; For each point in the first three-dimensional point cloud, find the nearest neighbor in the second three-dimensional point cloud, and calculate the directed distance from each point in the first three-dimensional point cloud to the corresponding nearest neighbor in the local slope normal direction. Integrate all the directed distances that are greater than zero to obtain the total volume of the gap between the first 3D point cloud and the second 3D point cloud; The normalized average Hausdorff distance is weighted and fused with the total void volume to generate the fit deviation.

3. The method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes according to claim 2, characterized in that, The step of weightedly fusing the normalized average Hausdorff distance with the total void volume to generate the fit deviation includes: The average curvature and rate of change of curvature of the current construction slope area are calculated based on the second three-dimensional point cloud. Determine whether the average curvature is greater than a first preset threshold and whether the rate of change of curvature is greater than a second preset threshold; If so, based on the magnitude of the curvature change rate, the weighting coefficient of the total volume of the void is dynamically calculated according to a preset linear mapping relationship, wherein the larger the curvature change rate, the larger the weighting coefficient; If not, assign a fixed basic weighting coefficient to the total volume of the voids; The normalized total void volume is weighted using the determined weighting coefficients, and the normalized average Hausdorff distance is weighted using a preset fixed weighting coefficient. The two weighted results are added together to generate the fit deviation.

4. The method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes according to claim 1, characterized in that, The acquisition of the mechanical state information of the three-dimensional geonet during the laying process includes: The visual sensing device is used to identify and track the image coordinates of the grid intersections of the three-dimensional geonet. The three-dimensional spatial displacement of the grid intersection node is calculated based on the changes in the image coordinates between consecutive frames and the calibration parameters of the visual sensing device. The three-dimensional spatial displacement is used as the mechanical state information.

5. The method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes according to claim 4, characterized in that, The process of identifying and tracking the image coordinates of the grid intersections of the three-dimensional geonet includes: The original image acquired by the visual sensing device is preprocessed by grayscale conversion and contrast enhancement. On the preprocessed image, the scale-invariant feature transform algorithm is used to detect and generate key points and descriptors of the key points; Based on the preset mesh geometry and spatial topological relationship of the three-dimensional geonet, the theoretical image spacing range of the intersection nodes of adjacent meshes is calculated; Based on the theoretical image spacing range, spatial clustering and filtering are performed on all detected key points, and key points that conform to the grid distribution rules are retained as candidate grid intersection nodes; Between consecutive image frames acquired by the visual sensing device, the Euclidean distance between the descriptors of candidate grid intersection nodes is calculated, and the inter-frame correspondence is established through the nearest neighbor matching algorithm; For successfully matched point pairs, perform bidirectional consistency verification and outlier removal to obtain the image coordinate sequence of grid intersection nodes that are stably tracked.

6. The method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes according to claim 4 or 5, characterized in that, The process of obtaining the tension distribution of the three-dimensional geonet through inverse solution based on the mechanical state information and the preset constitutive relation of the geonet includes: The three-dimensional spatial displacement of the grid intersection nodes is used as the known displacement boundary condition; Based on the material parameters and mesh topology of the three-dimensional geonet, the system equilibrium equation with the internal forces of the elements as unknowns is established using the finite element method. Solving the equilibrium equations of the system yields the internal forces of each element in the three-dimensional geonet, thus forming the tension distribution.

7. The method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes according to claim 1, characterized in that, The calculation of the tension distribution uniformity index, which characterizes the degree of tension uniformity, based on the tension distribution includes: Calculate the standard deviation of the tension distribution; Calculate the average value of the tension difference between adjacent grid cells in the tension distribution, and use it as the tension gradient index; The tension distribution uniformity index is calculated according to the formula: uniformity index = K / (1+α*standard deviation+β*tension gradient index), where K is a normalization constant and α and β are preset adjustment coefficients.

8. The method for quality assessment and process optimization of three-dimensional geonet laying for complex slopes according to claim 1, characterized in that, The method, based on the fit deviation and the tension distribution uniformity index, uses a pre-established quality assessment model to calculate and output a comprehensive real-time construction quality score, including: The fit deviation and the tension distribution uniformity index are normalized respectively to form an input vector; The input vector is input into a multilayer perceptron neural network model that includes an attention mechanism module; the attention mechanism module calculates the attention weight values ​​corresponding to the fit deviation and the tension distribution uniformity index, respectively. The calculated attention weight values ​​are used to weight the corresponding input features to form a weighted feature vector; The weighted feature vector is input into at least one fully connected hidden layer of the neural network model for nonlinear transformation; The transformed result is mapped to the real-time construction quality score through the output layer of the neural network model.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-8.

Images