Pillar non-destructive climbing control method based on fusion of vision and force sensing

Abstract

This invention relates to the field of machine vision technology, and more particularly to a non-destructive climbing control method for piers based on the fusion of vision and force sensing. This method collects surface defect and material information to generate a three-dimensional state vector, which is then matched with a pressure threshold matrix. Force sensing feedback enables distributed extrusion to form initial adhesion. Deformation monitoring identifies crack and spalling damage features and generates a three-level risk vector field. Based on the risk level, multi-objective optimization determines the initial control strategy. Secondary visual scanning and pressure monitoring verify the optimization effect, using the modulus change rate and pressure standard deviation as dual indicators to determine convergence; if convergence fails, feedback iteration occurs. After the current layer stabilizes, reinforcement learning generates displacement trajectories, activating the next layer and repeating the entire process to achieve collaborative climbing. This invention, through a closed-loop mechanism of visual prediction, force control execution, and dynamic verification, effectively avoids pier surface damage and improves climbing efficiency and safety under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and in particular to a non-destructive climbing control method for piers based on the fusion of vision and force sensing. Background Technology

[0002] With the continuous expansion of bridge construction in my country, the need for surface defect detection and maintenance of bridge piers, as key load-bearing structures, is becoming increasingly urgent. Traditional manual inspection methods are not only inefficient and costly, but also pose safety risks associated with working at heights. In recent years, climbing robot technology has been gradually applied to the field of pier inspection. Existing technologies mainly use rigid mechanical clamps or pneumatic suction cup structures, combined with a preset constant compressive force for climbing and fixation. While such solutions can achieve basic climbing functions, they have the following prominent problems in practical applications:

[0003] First, existing climbing robots lack the ability to perceive the surface condition of piers and cannot identify defects such as concrete cracks, spalling, and honeycomb pitting. Applying uniform extrusion pressure can easily lead to secondary damage to weak areas. Although some studies have introduced a single vision module for surface observation, image acquisition and force control execution are independent of each other, failing to form a closed-loop feedback and making it difficult to achieve adaptive pressure adjustment.

[0004] Secondly, traditional control methods often employ open-loop or simple PID regulation, which fails to adequately consider the pressure balance at each climbing node. When the pier cross-section has varying diameters, curvatures, or irregular shapes, localized stress concentrations are highly likely to occur, leading to the propagation of microcracks in the concrete surface or coating peeling, thus violating the original purpose of non-destructive testing.

[0005] Third, existing technologies generally adopt a serial operation mode of "fixing layer by layer - lifting as a whole", which lacks coordinated planning between climbing layers, is time-consuming, and has the risk of instantaneous instability during the transition between layers. Although some solutions attempt multi-layer linkage, trajectory planning relies on manual preset and cannot dynamically adjust climbing speed and posture according to real-time risk level, making it difficult to adapt to complex and ever-changing on-site conditions.

[0006] In summary, existing technologies are insufficient to simultaneously meet the comprehensive requirements of adapting to complex pier shapes, real-time surface condition perception, intelligent control of compressive pressure, and multi-layer collaborative optimization. There is an urgent need for a non-destructive climbing control method that integrates visual-force sensing information and has autonomous decision-making and dynamic verification capabilities to achieve safety, efficiency, and intelligence in bridge pier inspection. Summary of the Invention

[0007] Therefore, the present invention provides a non-destructive climbing control method for piers based on the fusion of vision and force sensing, in order to solve the aforementioned problems existing in the prior art.

[0008] To achieve the above objectives, the present invention provides a non-destructive climbing control method for piers based on the fusion of vision and force sensing, comprising:

[0009] Step S1: Collect the defect distribution information and material strength information of the pier surface in real time to construct a three-dimensional surface state vector, and generate the extrusion pressure threshold matrix of each target extrusion point by matching based on the preset surface state-pressure response relation database.

[0010] Step S2: Perform extrusion control on each target extrusion point according to the extrusion pressure threshold matrix to obtain a preliminary adhesion state;

[0011] Step S3: Monitor the surface deformation of the contact area of ​​each extrusion point according to the preliminary adhesion state, analyze and compare the images before and after extrusion based on the surface deformation to identify the failure evolution characteristics, and classify the failure evolution characteristics according to the degree of strain concentration and the rate of expansion to generate a risk level and its failure risk vector field;

[0012] Step S4: Based on the destruction risk vector field, determine the priority and scheduling according to the risk level, and start a multi-objective optimization loop to determine the initial control strategy;

[0013] Step S5: Based on the initial control strategy, perform a secondary scan of each extrusion point using vision and monitor the pressure distribution using a force sensor to obtain a convergence determination result.

[0014] Step S6: Based on the convergence determination result and the risk level, after the current layer stabilizes, a displacement control trajectory corresponding to the risk level is generated through a reinforcement learning algorithm, and the next execution layer is activated simultaneously to repeat the process from step S1 to step S5 to obtain the target control strategy.

[0015] Furthermore, the process of step S1 includes:

[0016] The entire surface of the pier is scanned from a distance, while close-range image acquisition is performed on the target areas of each compression point to obtain multi-scale image data.

[0017] Defect semantic segmentation is performed on the multi-scale image data to identify defect types and output defect confidence vectors. At the same time, the concrete grade or coating type is determined based on the surface material texture and reflective properties to determine the surface bearing strength parameters.

[0018] The defect confidence vector and the bearing strength parameter are concatenated along the channel dimension, and the local curvature change features calculated from the surface point cloud data are fused to construct the three-dimensional surface state vector.

[0019] The three-dimensional surface state vector is input into the relational database, and the optimal extrusion pressure threshold for the current region is matched using a nonlinear interpolation algorithm to obtain the extrusion pressure threshold matrix.

[0020] Furthermore, the process of performing defect semantic segmentation on the multi-scale image data to identify defect types and output defect confidence vectors includes:

[0021] A pre-trained convolutional neural network is used to perform pixel-level segmentation on the multi-scale image data to obtain a probability heatmap of the defect type;

[0022] The probability heatmap is input into the defect connected component analysis algorithm to calculate the area, aspect ratio, and edge roughness parameters of each defect region.

[0023] Based on the area, aspect ratio, and edge roughness parameters, a support vector machine classifier is used to determine the defect type and output the corresponding confidence vector.

[0024] Furthermore, the process of step S2 includes:

[0025] Read the extrusion pressure threshold matrix to obtain the target pressure value corresponding to each target extrusion point in order to start the distributed collaborative extrusion process;

[0026] According to the target pressure value, progressive pressurization is performed synchronously at each target extrusion point, while force sensor feedback data is collected in real time and the pressure deviation between the current pressure and the target pressure is calculated.

[0027] When the pressure deviation value enters the preset transition range, the pressurization rate is automatically reduced and the micro-motion control mode is switched. When the pressure deviation value approaches zero, pressure locking is performed.

[0028] The force balance index is calculated based on the pressure difference between each extrusion point. The output of each point is adjusted by pressure compensation so that the force balance index converges to a stable range to obtain the initial adhesion state.

[0029] Furthermore, the process of calculating the force balance index based on the pressure difference between each extrusion point and adjusting the output of each point through pressure compensation includes:

[0030] The average pressure value of all extrusion points is calculated as the force balance benchmark. The deviation between the real-time pressure of each extrusion point and the average pressure value is calculated. The root mean square value of the deviation of all extrusion points is used as the force balance index.

[0031] When the force balance index exceeds the preset stable range, the extrusion point with the largest pressure deviation is identified as the adjustment target; a pressure retraction operation is performed on the extrusion point with pressure higher than the average value, and a pressure compensation operation is performed on the extrusion point with pressure lower than the average value.

[0032] Based on the pressure distribution of adjacent extrusion points, the retraction pressure value is allocated to the extrusion points with lower pressure according to the distance weight ratio. The above calculation and adjustment process is repeated until the force balance index converges to the stable range.

[0033] Furthermore, the process of step S3 includes:

[0034] Acquire a reference image before extrusion and a real-time image after extrusion, and perform affine registration on the reference image and the real-time image to obtain a registered image;

[0035] The image is subjected to pixel-level difference operations to calculate the surface displacement vector field and convert it into a strain field distribution;

[0036] In the strain field distribution, strain concentration regions are identified, and principal strain regions exceeding the material's elastic limit are marked as crack initiation zones, while shear strain localization zones are marked as spalling risk zones.

[0037] The strain amplitude and regional expansion rate of the strain concentration region are quantitatively classified to generate a damage risk vector field containing three levels of risk (high, medium, and low) and spatial coordinate information.

[0038] Furthermore, the process of performing pixel-level difference operations on the image to calculate the surface displacement vector field and convert it into a strain field distribution includes:

[0039] The displacement vector of each pixel in the registered image is calculated using the optical flow method to generate a dense displacement field.

[0040] The displacement gradient tensor is obtained by performing spatial gradient calculation on the dense displacement field;

[0041] The displacement gradient tensor is symmetrically decomposed to extract principal strain components and shear strain components to generate a strain field distribution characterizing the degree of surface deformation.

[0042] Furthermore, the process of step S4 includes:

[0043] Based on the risk level and spatial distribution information of each region, high-risk areas are marked as priority adjustment targets to obtain the marking results;

[0044] A multi-objective optimization scheduling queue is established based on the labeling results, and higher adjustment priority and greater adjustment weight are assigned to high-risk areas;

[0045] An optimization model is constructed using a non-dominated sorting genetic algorithm, with the first objective being to maximize the distance to avoid weak areas and the second objective being to minimize the pressure variance of force distribution uniformity.

[0046] The optimization model is iteratively solved, and the solution that best matches the current risk level is selected from the obtained frontier solution set as the initial control strategy.

[0047] Furthermore, the process of step S5 includes:

[0048] According to the initial control strategy, each extrusion point is driven to adjust its attitude according to the reconstructed pressure field and topological configuration;

[0049] After the posture adjustment is completed, secondary image data of the contact area of ​​each extrusion point is acquired by visual acquisition, and real-time pressure distribution data is obtained by force sensor.

[0050] Deformation analysis is performed on the secondary image data to calculate the modulus change rate of the damage risk vector field, and the standard deviation of the pressure distribution at all extrusion points is calculated on the real-time pressure distribution data.

[0051] When the magnitude of the risk vector field approaches zero and the standard deviation of the pressure distribution is less than the preset convergence threshold, the convergence is determined to be successful and the convergence determination result is output; when the convergence condition is not met, the current pressure distribution data is fed back to step S4 for further optimization.

[0052] Furthermore, step S6 includes the following process:

[0053] Construct a reinforcement learning state space, including the relative position information of the current layer, the pressure distribution vector of each extrusion point, the surface defect distribution map of the pier column, and environmental disturbance parameters;

[0054] Construct a reinforcement learning action space, including motion velocity, motion acceleration, and squeeze timing control parameters at each level;

[0055] The reward function weights are set according to the risk level, and the reward function is constructed by weighting climbing speed, energy consumption and damage level.

[0056] A proximal policy optimization algorithm is used to iterate the policy in the state space and action space to generate a motion trajectory that matches the current risk level.

[0057] After the current layer completes the convergence determination, the next layer is driven to repeat the process of steps S1 to S5 to obtain the target control strategy.

[0058] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention dynamically generates the extrusion pressure threshold by visually perceiving defects and material characteristics, utilizes a force sensing closed loop to achieve progressive, impact-free adhesion, and quantitatively assesses the risk of failure based on two dimensions: strain concentration and propagation rate. This makes the risk level positively correlated with the optimization weight, driving a multi-objective genetic algorithm to find the optimal balance between obstacle avoidance distance and pressure uniformity. It also ensures that the convergence process follows the material's mechanical response characteristics through dual feedback iteration correction using modulus change rate and pressure standard deviation. Reinforcement learning adaptively adjusts the speed and safety weights according to the risk level to generate the optimal displacement trajectory, achieving interlayer spatiotemporal interleaving and coordination, forming a complete causal chain of perception-decision-execution-verification. This effectively avoids the propagation of concrete microcracks caused by pressure overload, significantly improving climbing efficiency and pier structural safety in complex environments. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the non-destructive climbing control method for piers based on the fusion of vision and force sensing provided by the present invention.

[0060] Figure 2 This is a flowchart illustrating the non-destructive climbing control method for piers based on the fusion of vision and force sensing provided by the present invention.

[0061] Figure 3 This is a flowchart illustrating the non-destructive climbing control method for piers based on the fusion of vision and force sensing provided by the present invention.

[0062] Figure 4 This is a flowchart illustrating the non-destructive climbing control method for piers based on the fusion of vision and force sensing provided by the present invention. Detailed Implementation

[0063] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0064] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0065] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0066] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0067] Please see Figure 1 As shown, this invention provides a non-destructive climbing control method for piers based on the fusion of vision and force sensing, comprising:

[0068] Step S1: Collect the defect distribution information and material strength information of the pier surface in real time to construct a three-dimensional surface state vector, and generate the extrusion pressure threshold matrix of each target extrusion point by matching based on the preset surface state-pressure response relation database.

[0069] Specifically, such as Figure 2 As shown, the process of step S1 includes:

[0070] Step S11: Scan the entire surface of the pier from a distance, and at the same time, perform close-range image acquisition on the target area of ​​each compression point to obtain multi-scale image data.

[0071] Specifically, a zoom-vision acquisition method is employed. First, a wide-angle long-distance scan of the entire pier is performed to acquire a low-resolution image containing the entire pier, used to identify the macroscopic distribution patterns of defects over a large area. Simultaneously, close-range images are acquired using optical zoom of the local area that each extrusion point is about to contact, obtaining high-resolution detail images for precise identification of minute defects in the contact area. The extrusion point target area refers to the pre-defined range surrounding the actual contact point between each extrusion foot of the climbing robot and the pier surface, typically defined as a circular area with a diameter of 5-10 cm extending outward from the center of the contact point. This area is the key focus area for subsequent pressure control.

[0072] Step S12: Perform defect semantic segmentation on the multi-scale image data to identify the defect type and output the defect confidence vector. At the same time, determine the concrete grade or coating type based on the surface material texture and reflective properties to determine the surface bearing strength parameters.

[0073] Specifically, surface texture analysis extracts texture roughness and contrast parameters using a gray-level co-occurrence matrix, while reflectivity is determined by analyzing the skewness and kurtosis of the image brightness histogram. Combined with a pre-defined optical property database of concrete grades and coating types, the surface load-bearing strength parameters are retrieved. The pre-defined optical property database refers to a pre-established database of correspondences between material optical characteristics and physical properties, calibrated through experiments. This database stores standard reflectivity parameters for different concrete grades (e.g., C30, C40, C50) and coating types (e.g., epoxy resin, polyurethane) under specific lighting conditions, including statistical quantities such as mean brightness, standard deviation, skewness, and kurtosis, as well as corresponding reference values ​​for surface load-bearing strength.

[0074] Specifically, the process of performing defect semantic segmentation on the multi-scale image data to identify defect types and output defect confidence vectors includes:

[0075] A pre-trained convolutional neural network is used to perform pixel-level segmentation on the multi-scale image data to obtain a probability heatmap of the defect type;

[0076] Specifically, after inputting multi-scale image data into the network, the probability distribution of each pixel in four categories—cracks, spalling, honeycomb pitting, and intact surfaces—is output, forming a four-channel probability heatmap. The pixel value of each channel represents the probability that the location belongs to the corresponding defect type, and the value ranges from 0 to 1.

[0077] Specifically, a probability heatmap is a two-dimensional image data that uses color depth or grayscale values ​​to represent probability. In this method, four grayscale images of the same size as the original image are used, each corresponding to a defect type. The closer the grayscale value of a pixel in the image is to 1 (represented by a warm color), the higher the probability that the location belongs to this type of defect. This representation method preserves the gradation information of the defect boundary and internal region.

[0078] The probability heatmap is input into the defect connected component analysis algorithm to calculate the area, aspect ratio, and edge roughness parameters of each defect region.

[0079] Specifically, a confidence threshold (e.g., 0.7) is set for each heatmap, and the pixels with a probability higher than the threshold are marked as candidate defect pixels. Then, the eight-neighbor connectivity rule is used to cluster adjacent candidate pixels into independent connected regions, each region representing a potential defect instance. The area is obtained by counting the number of pixels in the region and multiplying it by the actual pixel size; the aspect ratio is obtained by fitting the ratio of the long side to the short side of the minimum bounding rectangle of the region; the edge roughness parameter is obtained by calculating the standard deviation of the Fret diameter of the boundary pixels or the boundary fractal dimension. The edge roughness parameter is a quantitative indicator describing the irregularity of the defect boundary; the larger the value, the more tortuous the boundary and the more complex the defect morphology.

[0080] Based on the area, aspect ratio, and edge roughness parameters, a support vector machine classifier is used to determine the defect type and output the corresponding confidence vector.

[0081] Specifically, a multi-class support vector machine (SVM) is employed. A three-dimensional feature vector is constructed from the area, aspect ratio, and edge roughness of each defect's connected component, and this vector is input into a pre-trained SVM model. The SVM maps the features to a high-dimensional space using a kernel function (such as a radial basis function), and then searches for the optimal classification hyperplane in this space that best distinguishes different defect types. For each defect instance, the classifier outputs its decision function values ​​for the four categories. The softmax function is then used to convert these decision values ​​into a probability distribution, forming a confidence vector.

[0082] Step S13: The defect confidence vector and the bearing strength parameter are concatenated by channel dimension, and the local curvature change features calculated from the surface point cloud data are fused to construct the three-dimensional surface state vector.

[0083] Specifically, channel-dimensional stitching refers to stacking and combining feature data from different sources along the feature channel direction, similar to the RGB three-channel synthesis method in image processing. Surface point cloud data is a 3D coordinate dataset of the pier surface obtained through structured light or binocular vision principles, containing the spatial location information of each point. Local curvature variation features are obtained by calculating the principal curvature of the local surface fitted to each point and its neighboring points in the point cloud data, reflecting the degree of surface unevenness. When constructing the 3D surface state vector, the defect confidence vector (4D), bearing strength parameter (1D), and local curvature features (2D, including maximum and minimum principal curvature) are stitched together along the channel dimension to form a 7D feature vector. This vector has a corresponding value at each spatial coordinate position on the pier surface, forming a 3D tensor structure.

[0084] Step S14: Input the three-dimensional surface state vector into the relational database, and use a nonlinear interpolation algorithm to match the optimal extrusion pressure threshold for the current region to obtain the extrusion pressure threshold matrix.

[0085] Specifically, the relational database is a pre-established surface state-pressure response mapping library built through experiments, storing the safe pressure thresholds corresponding to different combinations of surface state characteristics. Since the database cannot cover all possible surface state combinations, a nonlinear interpolation algorithm is used to perform smooth transition predictions between existing data points. During interpolation, the Euclidean distance between the 3D state vector of the query point and each sample point in the database is used as the weight. The optimal pressure threshold for the current region is calculated through radial basis function or cubic spline interpolation, ultimately generating a corresponding threshold for each spatial location on the pier surface, forming a compressive pressure threshold matrix.

[0086] Specifically, the optimal compressive force threshold refers to the maximum safe compressive force value that ensures stable adhesion of the climbing robot without damaging the pier surface under the current surface conditions. Physically, it represents a critical pressure value; below this value, adhesion is insufficient, while above it may cause surface damage. This threshold is determined comprehensively based on the degree of surface defects, material strength, and geometry. The range of the optimal compressive force threshold is set based on concrete material characteristics and engineering experience as follows: 400-600N for intact concrete surfaces, 250-350N for slightly cracked surfaces, and 100-200N for severely defective surfaces. This range is obtained through offline calibration using indentation and shear tests.

[0087] Step S2: Perform extrusion control on each target extrusion point according to the extrusion pressure threshold matrix to obtain a preliminary adhesion state;

[0088] Specifically, such as Figure 3 As shown, the process of step S2 includes:

[0089] Step S21: Read the extrusion pressure threshold matrix and obtain the target pressure value corresponding to each target extrusion point to start the distributed collaborative extrusion process;

[0090] Specifically, the initial compressive pressure threshold matrix is ​​a two-dimensional data structure. Row coordinates correspond to the spatial location information on the pier surface, column coordinates correspond to the numbers of each compressive point, and the matrix element values ​​represent the target pressure value for that compressive point. Each element in the matrix is ​​traversed sequentially, and the triplet of row coordinates, column coordinates, and a numerical value is parsed. The numerical value is then used as the target pressure value for the corresponding numbered compressive point.

[0091] Step S22: Gradual pressurization is applied to each target extrusion point simultaneously according to the target pressure value, while force sensor feedback data is collected in real time and the pressure deviation between the current pressure and the target pressure is calculated.

[0092] Specifically, progressive pressurization means that the pressure output is not achieved in one step, but rather gradually increased according to a preset pressurization curve. The initial stage uses a rapid pressurization phase (with a steep slope), and when the real-time pressure approaches 70% of the target value, it switches to a slow pressurization phase (with a gentler slope). The pressure deviation is calculated using the formula: ΔP = P_target - P_current, where P_target is the target pressure value and P_current is the real-time pressure value fed back by the force sensor. The force sensor feedback data acquisition frequency is set to 100Hz-1000Hz to ensure that the dynamic pressure change process can be captured.

[0093] Step S23: When the pressure deviation value enters the preset transition range, the pressurization rate is automatically reduced and the micro-motion control mode is switched. When the pressure deviation value approaches zero, pressure locking is performed.

[0094] Specifically, the preset transition range is defined as the range where the absolute value of the pressure deviation is less than 10%-15% of the target pressure value. This range is obtained through experimental calibration and represents a stable margin zone where the pressure is close to the target value but has not yet been precisely reached. The micro-motion control mode refers to reducing the pressurization rate to 5%-10% of the normal rate and shortening the control cycle to half its original length. A finer pulse width modulation signal is used to drive the actuator, achieving micrometer-level step adjustments to the pressure. The condition for the pressure deviation to approach zero is set as: |ΔP| < 0.5% × P_target and the duration exceeds 200ms. Pressure locking refers to the control system sending a holding command to the actuator after the target pressure is reached, cutting off the power input and activating a mechanical locking device or electromagnetic brake to stabilize the pressure value near the target value, preventing pressure drop or overshoot.

[0095] Step S24: Calculate the force balance index based on the pressure difference between each extrusion point, and adjust the output of each point through pressure compensation to make the force balance index converge to a stable range to obtain the initial adhesion state.

[0096] Specifically, the process of calculating the force balance index based on the pressure difference between each extrusion point and adjusting the output of each point through pressure compensation includes:

[0097] The average pressure value of all extrusion points is calculated as the force balance benchmark. The deviation between the real-time pressure of each extrusion point and the average pressure value is calculated. The root mean square value of the deviation of all extrusion points is used as the force balance index.

[0098] Specifically, the force balance index is a proprietary parameter unique to this method, used to quantitatively evaluate the uniformity of pressure distribution at multiple extrusion points. Its physical meaning is the degree of dispersion of pressure at each point relative to the overall average. Here, the force balance index is a numerical indicator that can be calculated in real time and dynamically tracked, rather than a qualitative description. The calculation formula is P_avg = (P_1 + P_2 + ... + P_n) / n, where n is the number of extrusion points and P_avg is the force balance benchmark. The pressure deviation refers to the difference between the real-time pressure value P_i at each extrusion point and the force balance benchmark P_avg, i.e., ΔP_i = P_i - P_avg. The force balance index is a comprehensive evaluation value obtained by summarizing the pressure deviations of all extrusion points using the root mean square algorithm. The smaller the index value, the more uniform the pressure distribution at each extrusion point. A sliding window averaging method is used in the calculation, with the window length set to 50-100 sampling periods to filter out sensor noise and transient disturbances.

[0099] When the force balance index exceeds the preset stable range, the extrusion point with the largest pressure deviation is identified as the adjustment target; a pressure retraction operation is performed on the extrusion point with pressure higher than the average value, and a pressure compensation operation is performed on the extrusion point with pressure lower than the average value.

[0100] Specifically, the preset stability range is a manually set allowable range of force balance indicators based on climbing safety requirements, typically set to [0.05×P_avg, 0.15×P_avg], meaning that the root mean square value of the pressure deviation at each point is allowed to be between 5% and 15% of the average pressure. When the real-time calculated force balance indicator exceeds the upper limit of this range (0.15×P_avg), the system enters active adjustment mode. When identifying the adjustment target, the absolute value of the pressure deviation ΔP_i at all squeezing points is taken and sorted, and the squeezing point with the largest absolute value is selected as the primary adjustment target. Pressure rollback operation refers to reducing the target pressure setting value of the squeezing point at a preset rate (e.g., reducing the target pressure by 5% per second); pressure compensation operation refers to increasing the target pressure setting value of the low-pressure squeezing point at a preset rate (e.g., increasing the target pressure by 5% per second).

[0101] Based on the pressure distribution of adjacent extrusion points, the retraction pressure value is allocated to the extrusion points with lower pressure according to the distance weight ratio. The above calculation and adjustment process is repeated until the force balance index converges to the stable range.

[0102] Specifically, adjacent compression points refer to the 2-3 compression points that are closest to the object being adjusted in terms of geometric location within the climbing loop. The distance weight ratio is calculated as follows: Let the distances between the object being adjusted and its neighboring points B and C be d_AB and d_AC, respectively. Then, the compensation weight obtained by point B is w_B = (1 / d_AB) / (1 / d_AB + 1 / d_AC), and the weight of point C is w_C = (1 / d_AC) / (1 / d_AB + 1 / d_AC), meaning that the closer the distance, the greater the weight. The retreat pressure value ΔP_retreat is distributed to each low-pressure point according to the weight: ΔP_B = ΔP_retreat × w_B, ΔP_C = ΔP_retreat × w_C. The criterion for convergence to the stable interval is that the force balance index remains above the lower limit (0.05×P_avg) and below the upper limit (0.15×P_avg) of the preset stable interval for 5-10 consecutive calculation cycles, indicating that the system has reached a dynamic equilibrium state.

[0103] Step S3: Monitor the surface deformation of the contact area of ​​each extrusion point according to the preliminary adhesion state, analyze and compare the images before and after extrusion based on the surface deformation to identify the failure evolution characteristics, and classify the failure evolution characteristics according to the degree of strain concentration and the rate of expansion to generate a risk level and its failure risk vector field;

[0104] Specifically, such as Figure 4 As shown, the process of step S3 includes:

[0105] Step S31: Obtain the reference image before extrusion and the real-time image after extrusion, and perform affine registration on the reference image and the real-time image to obtain the registered image;

[0106] Specifically, the process of performing pixel-level difference operations on the image to calculate the surface displacement vector field and convert it into a strain field distribution includes:

[0107] The displacement vector of each pixel in the registered image is calculated using the optical flow method to generate a dense displacement field.

[0108] The displacement gradient tensor is obtained by performing spatial gradient calculation on the dense displacement field;

[0109] The displacement gradient tensor is symmetrically decomposed to extract principal strain components and shear strain components to generate a strain field distribution characterizing the degree of surface deformation.

[0110] Step S32: Perform pixel-level difference operation on the image to calculate the surface displacement vector field and convert it into strain field distribution;

[0111] Specifically, optical flow is used to calculate the displacement vector of each pixel in the registered image, generating a dense displacement field. Based on the assumptions of constant image brightness and local smoothness, optical flow solves the motion equations of pixels between adjacent image frames to obtain the displacement components of each pixel in the X and Y directions. Next, spatial gradient operations are performed on the dense displacement field to obtain the displacement gradient tensor, i.e., calculating the partial derivatives of the displacement components with respect to the X and Y coordinates at each pixel location, forming a 2×2 derivative matrix. Finally, the displacement gradient tensor is symmetrically decomposed to extract the principal strain components and shear strain components to generate a strain field distribution characterizing the degree of surface deformation. Specifically, the symmetric decomposition divides the displacement gradient tensor into a symmetric part and an antisymmetric part, where the eigenvalues ​​of the symmetric part are the principal strain components (magnitude of tensile and compressive deformation), and the off-diagonal elements of the symmetric part are the shear strain components (magnitude of shear deformation).

[0112] Step S33: Identify strain concentration regions in the strain field distribution, mark the principal strain regions exceeding the material's elastic limit as crack initiation zones, and mark the shear strain localization zones as spalling risk zones.

[0113] Specifically, strain concentration regions refer to localized areas in the strain field where the strain value is significantly higher than the surrounding background value. These are identified through threshold segmentation and connected component analysis: principal strain thresholds and shear strain thresholds are set; pixels exceeding the thresholds are marked as candidate regions; and adjacent candidate pixels are clustered to form connected regions. The elastic limit of a material refers to the maximum strain value at which concrete can recover deformation. Exceeding this value will result in irreversible plastic deformation or cracking. This limit is determined through uniaxial compression tests on concrete, typically 0.15%-0.2% strain for C30 concrete. Principal strain regions refer to connected regions where the principal strain component exceeds the elastic limit; these are marked as crack initiation zones, indicating that microcracks have already appeared or macrocracks are about to occur in these areas. Shear strain localization zones refer to areas where the shear strain component forms a narrow, elongated band, characterized by high shear strain values ​​and linear or arc-shaped extension; these are marked as spalling risk zones, indicating shear slippage in these areas, and the surface concrete may peel off.

[0114] Specifically, the principal strain threshold is a key criterion for determining whether tensile cracks are about to form on the concrete surface. When the principal strain component in a localized area exceeds this threshold, it indicates that the material has exceeded its elastic deformation range and entered the plastic damage stage, where micro-cracks begin to initiate and propagate. Continued pressure will lead to macroscopically visible cracks. The principal strain threshold is set according to the following rules: Intact concrete surface: 0.12% - 0.15% (strain value); Concrete surface with micro-cracks: 0.08% - 0.10%; Coating or repair layer surface: 0.05% - 0.08%. The shear strain threshold is a key criterion for determining whether shear slip occurs on the concrete surface, leading to spalling failure. When shear strain is concentrated in a localized area and forms a banded distribution exceeding this threshold, it indicates the presence of a shear slip surface within the material, resulting in relative displacement between the surface concrete and the bulk, posing a risk of spalling. The shear strain threshold is set according to the following rules: ordinary concrete surface: 0.10% - 0.12% (shear strain value); inferior concrete with honeycomb pitting: 0.06% - 0.08%; smooth coated surface: 0.04% - 0.06%.

[0115] Step S34: Quantitatively classify the strain amplitude and regional expansion rate of the strain concentration region to generate a damage risk vector field containing three levels of risk (high, medium, and low) and spatial coordinate information.

[0116] Specifically, strain amplitude refers to the maximum principal strain or maximum shear strain value among all pixels within the strain concentration region, representing the severity of damage in that region. Region expansion rate refers to the speed at which the area or boundary of the strain concentration region expands outward compared to the previous detection time, calculated as: Expansion rate = (Current area - Previous area) / Time interval, reflecting the degree of active evolution of damage. The quantification and grading rules are as follows: High risk level indicates that the strain amplitude exceeds the elastic limit by more than 50% and the expansion rate is greater than the expansion rate threshold; Medium risk level indicates that the strain amplitude exceeds the elastic limit but the expansion rate is lower than the expansion rate threshold, or the strain amplitude is close to the elastic limit but the expansion rate is relatively high; Low risk level indicates that the strain amplitude is close to the elastic limit and the expansion rate is very low. The damage risk vector field is a three-dimensional data structure: the first dimension is spatial coordinates (pixel positions in the image), the second dimension is the risk level (high, medium, low), and the third dimension is the vector direction (pointing to the direction of the damage evolution gradient), used to comprehensively describe the spatial distribution and evolution trend of damage risk.

[0117] Specifically, the propagation rate thresholds are set according to the following rules: general propagation rate threshold: 0.5 mm² / s (region area growth rate); crack propagation linear rate threshold: 0.1 mm / s (crack tip advance speed); boundary propagation rate threshold: 0.3 mm / s (damage region boundary expansion speed).

[0118] Step S4: Based on the destruction risk vector field, determine the priority and scheduling according to the risk level, and start a multi-objective optimization loop to determine the initial control strategy;

[0119] Specifically, step S4 includes the following process:

[0120] Based on the risk level and spatial distribution information of each region, high-risk areas are marked as priority adjustment targets to obtain the marking results;

[0121] Specifically, the damage risk vector field is analyzed to extract all spatial coordinates of points with a risk level of "high". Simultaneously, the spatial distribution information of each high-risk area is analyzed, calculating the area, centroid position, and geometry of each area. High-risk areas with an area greater than 5 cm² or adjacent to the contact area of ​​the compression point are marked as priority adjustment targets, with a value of 1 assigned to the corresponding position in the mask matrix, and 0 assigned to the remaining positions. The marking process employs a morphological dilation operation, extending the boundary of the high-risk area outward by 3-5 pixels to ensure complete coverage of the hazardous impact range.

[0122] A multi-objective optimization scheduling queue is established based on the labeling results, and higher adjustment priority and greater adjustment weight are assigned to high-risk areas;

[0123] Specifically, the multi-objective optimization scheduling queue is a priority-sorted list of tasks. Each element contains three attributes: the coordinates of the region to be adjusted, the risk level, and the adjustment weight. All regions marked as 1 in the mask matrix are traversed and inserted into the queue head according to their risk level from high to low, forming a priority order of high → medium → low. Adjustment weights of 1.5-2.0 are assigned to high-risk regions, 1.0 to medium-risk regions, and 0.5-0.7 to low-risk regions. These weights are directly used as penalty coefficients for distance calculations in the subsequent objective function. The adjustment priority determines the order in which regions are processed in the optimization calculation, with high-risk regions being given priority consideration in the early iterations. The adjustment weight is an amplification factor; a larger weight means that the optimization algorithm has stricter distance requirements for that region when calculating avoidance distances, i.e., it requires the squeeze point to be farther away from that region.

[0124] An optimization model is constructed using a non-dominated sorting genetic algorithm, with the first objective being to maximize the distance to avoid weak areas and the second objective being to minimize the pressure variance of force distribution uniformity.

[0125] Specifically, an optimization model with two objectives is constructed: the first objective function is: Where N represents the total number of high-risk areas (areas marked as 1). This represents the Euclidean distance between the i-th compression point and the i-th high-risk area. The goal of assigning adjustment weights is to maximize this weighted sum, i.e., to keep the squeeze point as far away from the danger zone as possible. The second objective function is: Where n is the total number of compression points. , representing the average pressure of all squeezing forces. During algorithm initialization, a set of control strategies (called the population, containing 50-100 individuals, each representing a set of candidate schemes for squeezing point locations and pressures) is randomly generated. Through iterative evolution, each generation of individuals undergoes crossover mutation to produce offspring. Then, based on the objective function value, non-dominated sorting and crowding calculation are performed to select high-quality individuals that simultaneously achieve both objectives, ultimately forming the Pareto front solution set.

[0126] The optimization model is iteratively solved, and the solution that best matches the current risk level is selected from the obtained frontier solution set as the initial control strategy.

[0127] Specifically, the iterative solution process is set to a maximum of 100-150 generations, terminating when the generation count reaches the upper limit or the objective function value shows no significant improvement for 20 consecutive generations. After iteration, a set of non-dominant solutions is obtained, forming the "Pareto front solution set." Each solution in this set cannot improve either objective without compromising the other. When the high-risk area accounts for more than 30% of the damage risk vector field, the solution with the optimal first objective value (obstacle avoidance distance) (i.e., the solution furthest from the danger zone) is selected, prioritizing safety. When the high-risk area accounts for 10%-30%, a solution that balances the two objectives is selected (by calculating the normalized distance of the objective function for each solution and selecting the solution closest to the ideal point), balancing safety and stability. When the high-risk area accounts for less than 10%, the solution with the minimum second objective value (pressure variance) (i.e., the solution with the most uniform pressure distribution) is selected, as optimal mechanical performance can be pursued at this point. The selected solution is decoded into the specific location coordinates of each compression point, the ball joint rotation angle, and the pressure setpoint, forming the initial control strategy.

[0128] Step S5: Based on the initial control strategy, perform a secondary scan of each extrusion point using vision and monitor the pressure distribution using a force sensor to obtain a convergence determination result.

[0129] Specifically, step S5 includes the following process:

[0130] According to the initial control strategy, each extrusion point is driven to adjust its attitude according to the reconstructed pressure field and topological configuration;

[0131] Specifically, the initial control strategy comprises three sets of instructions: the target pressure value for each compression point (from the reconstructed pressure field), the spatial coordinates of each compression point (from the customized topology), and the rotation angle of the ball joint mechanism at each compression point (from attitude optimization). The execution process is as follows: the control unit parses the strategy instructions sequentially. First, it drives the scissor folding mechanism to adjust the spatial position of each compression point to the target coordinates, controlling the adjustment speed at 5-10 mm / s to avoid mechanical impact. After the position is adjusted, it drives the ball joint mechanism to rotate to the target angle, controlling the rotation speed at 2-5° / s to ensure smooth contact surface fit. Finally, it activates the pressure control loop, applying pressure to each compression point according to the reconstructed pressure field settings. The timing of attitude adjustment follows the principle of position first, then angle, then pressure, with a 50-100 ms delay between each step to ensure the preceding action is stable before executing subsequent actions.

[0132] After the posture adjustment is completed, secondary image data of the contact area of ​​each extrusion point is acquired by visual acquisition, and real-time pressure distribution data is obtained by force sensor.

[0133] Specifically, after attitude adjustment, the system waits 200-500 ms for the structural vibration to completely decay, ensuring stability. Then, a secondary scan is triggered: the visual acquisition frequency is set to 10-20 Hz, continuously capturing 5-10 frames of images at each compression point contact area. Clear secondary image data is obtained through multi-frame averaging and noise reduction. The force sensor samples at a high speed of 1000 Hz, continuously acquiring data for 1 second and calculating the time average to obtain stable real-time pressure distribution data. The secondary image data specifically refers to the images acquired after implementing the optimization strategy, with a different time point than the real-time images after compression, used to verify the adjustment effect. Visual and force sensor data acquisition are strictly synchronized, with timestamp alignment accuracy better than 1 ms, ensuring that the images and pressure data correspond to the same mechanical state.

[0134] Deformation analysis is performed on the secondary image data to calculate the modulus change rate of the damage risk vector field, and the standard deviation of the pressure distribution at all extrusion points is calculated on the real-time pressure distribution data.

[0135] Specifically, after registering the secondary image with the pre-extrusion reference image, the displacement vector and strain field distribution of each pixel are calculated. Strain concentration areas are re-identified and quantified, generating a new failure risk vector field. The formula for calculating the modulus change rate is: R_change = (||V_new|| - ||V_old||) / ||V_old||, where ||V_new|| is the modulus of the risk vector field after the secondary scan (the modulus is calculated as the square root of the sum of the squares of the areas of all high-risk regions and their risk level weights), and ||V_old|| is the modulus of the risk vector field before adjustment. The standard deviation of the pressure distribution reflects the dispersion of the current pressure distribution relative to the mean; the smaller the value, the more uniform the distribution.

[0136] When the magnitude of the risk vector field approaches zero and the standard deviation of the pressure distribution is less than the preset convergence threshold, the convergence is determined to be successful and the convergence determination result is output; when the convergence condition is not met, the current pressure distribution data is fed back to step S4 for further optimization.

[0137] Specifically, the condition for the modulus length to approach zero is: ||V_new|| < 0.1 × ||V_initial||, meaning the new risk modulus length is less than 10% of the initial risk modulus length, indicating that the high risk has been largely eliminated. The convergence threshold for the pressure standard deviation is set as: σ_threshold = max(0.03×P_avg, 0.5 N), taking the larger value between 3% of the average pressure and 0.5 N to ensure a reasonable convergence standard under both low-pressure and high-pressure conditions. Both conditions must be met simultaneously for successful convergence. If not, the real-time pressure values ​​{P_1, P_2, ..., P_n} of each extrusion point are used as new pressure constraints and fed back to the optimization model in step four. The multi-objective optimization calculation is re-executed to generate a corrected control strategy, forming a small closed-loop iteration of "optimization-execution-verification-re-optimization". A maximum of 3-5 iterations are allowed. If convergence is still not achieved, the current area is determined to be unclimbable and an alarm is triggered.

[0138] Step S6: Based on the convergence determination result and the risk level, after the current layer stabilizes, a displacement control trajectory corresponding to the risk level is generated through a reinforcement learning algorithm, and the next execution layer is activated simultaneously to repeat the process from step S1 to step S5 to obtain the target control strategy.

[0139] Specifically, step S6 includes the following process:

[0140] Construct a reinforcement learning state space, including the relative position information of the current layer, the pressure distribution vector of each extrusion point, the surface defect distribution map of the pier column, and environmental disturbance parameters;

[0141] Specifically, the relative position information is as follows: The height z of the current climbing layer relative to the bottom of the pier, the horizontal offsets Δx and Δy relative to the ideal climbing path, and the yaw angle θ relative to the pier axis are obtained through an encoder or displacement sensor, forming a 4-dimensional position sub-vector. The pressure distribution vector at each compression point is generated by real-time acquisition of pressure values ​​{P_1, P_2, ..., P_n} at all n compression points, calculating their average value P_avg and standard deviation σ. The original pressure values ​​are concatenated with the statistical values ​​to form an (n+2)-dimensional pressure sub-vector. The defect distribution map of the pier surface is generated by projecting the defect confidence portion (cracks, spalling, honeycomb pitting) from the 3D surface state vector onto a 2D polar coordinate grid, forming a defect distribution map with a resolution of 0.1 m / pixel. This map is then encoded into a 16-dimensional feature vector using a convolutional neural network. Environmental disturbance parameters include wind speed v_wind (obtained from an anemometer, unit: m / s), ambient temperature T (obtained from a temperature sensor, unit: ℃), and vibration amplitude a_vib (obtained from an accelerometer, unit: m / s²), forming a 3D environmental sub-vector. The final state space is the concatenation of the above sub-vectors, with a total dimension of 4 + (n+2) + 16 + 3 = n + 25 dimensions, where n is the number of compression points.

[0142] Construct a reinforcement learning action space, including motion velocity, motion acceleration, and squeeze timing control parameters at each level;

[0143] Specifically, two motion dimensions are defined: vertical lifting speed (v_lift) and horizontal adjustment speed (v_adj). Both values ​​are discretized into five levels: [-2 cm / s, 0, 2 cm / s, 5 cm / s, 10 cm / s]. Negative values ​​indicate downward or reverse adjustment. Acceleration (a) is defined as an independent motion dimension, with a discretized value range of [-10 cm / s², 0, 10 cm / s²], used to control start-stop smoothness and avoid impacts. For the double-layer climbing robot, two motion dimensions are defined: upper-layer compression timing (t_upper) and lower-layer compression timing (t_lower). These values ​​are relative time offsets (in ms), ranging from [-500 ms, 0, 500 ms]. Negative values ​​indicate premature compression, and positive values ​​indicate delayed compression. The total motion space has 5 dimensions, using discrete motion encoding, resulting in a total of 5 × 5 × 3 × 3 × 3 = 675 possible motions. Strategy selection is performed using motion index values.

[0144] The reward function weights are set according to the risk level, and the reward function is constructed by weighting climbing speed, energy consumption and damage level.

[0145] Specifically, R = w_speed × r_speed + w_energy × r_energy + w_damage × r_damage, where the three sub-reward functions are calculated as follows: Climbing speed reward r_speed: r_speed = Δz / Δt, which is the vertical displacement increment per unit time, in cm / s, encouraging rapid climbing. Energy consumption reward: ,in, This represents the energy consumption reward, where m represents the total number of squeeze points. This represents the average pressure at the i-th compression point within a time step. This represents the distance traveled by the i-th compression point within a time step, i.e., the negative sum of the products of the pressure at each compression point and the travel distance, encouraging low energy consumption. Damage reward r_damage: r_damage = -k × ||V||, i.e., the negative value of the magnitude of the risk vector field, where k is the risk level coefficient, encouraging low risk. Weight settings are dynamically adjusted according to the risk level.

[0146] High-risk level: w_speed = 0.2, w_energy = 0.3, w_damage = 0.5 (damage weight is the highest, safety first);

[0147] Medium risk level: w_speed = 0.4, w_energy = 0.3, w_damage = 0.3;

[0148] Low risk level: w_speed = 0.5, w_energy = 0.3, w_damage = 0.2 (speed has the highest weight, efficiency takes priority); the weight values ​​satisfy w_speed + w_energy + w_damage = 1.

[0149] A proximal policy optimization algorithm is used to iterate the policy in the state space and action space to generate a motion trajectory that matches the current risk level.

[0150] Specifically, the proximal policy optimization algorithm is a reinforcement learning algorithm based on policy gradients. Its core is the construction of two neural networks: a policy network and a value network. The policy network takes a state space vector as input and outputs the probability distribution of each action; the value network evaluates the value of the current state. A two-stage learning strategy of offline pre-training and online fine-tuning is employed: pre-training is performed in a simulation environment using a large number of pier models (containing different defect distributions and geometries), accumulating 1 million steps of interaction data; during actual climbing, the pre-trained model is used as the initial policy, and online fine-tuning (10-20 steps) is performed using real-world data after each layer is climbed. During policy iteration, a truncated policy loss is used: L_clip = E[min(r_t(θ)* A_t, clip(r_t(θ), 1-ε, 1+ε) * A_t)], where r_t(θ) = π_new(a|s) / π_old(a|s) is the probability ratio of the new policy to the old policy, A_t is the advantage function, and ε=0.2 is the truncation hyperparameter to prevent instability caused by excessive policy updates. In each iteration generating the motion trajectory, the current state is input into the policy network, and the 5 actions with the highest probabilities are selected for Monte Carlo tree search simulation. The cumulative reward for the next 3-5 steps is evaluated, and finally, the optimal action sequence is selected as the motion trajectory.

[0151] After the current layer completes the convergence determination, the next layer is driven to repeat the process of steps S1 to S5 to obtain the target control strategy.

[0152] Specifically, after the current layer completes the convergence determination, it sends a "stable layer" signal to the upper-level control unit, containing information such as the current layer's height, pressure distribution, and risk level. Upon receiving the signal, the upper-level control unit initiates steps one through five of the current layer's process, while the current layer begins its lifting action based on the motion trajectory generated in step four. A spatiotemporally staggered control sequence is formed between the two layers: when the current layer lifts, the next layer performs compression and fixation; after the current layer is fixed, the next layer begins lifting. This alternating cycle achieves a streamlined "climb-detection-adjustment" operation. The target control strategy is the final complete output plan, including the motion trajectory sequence of all climbing layers, the pressure timing of each layer, the risk level distribution map, the total energy consumption estimate, and the total time estimate, forming an executable multi-layer collaborative climbing plan.

[0153] Specifically, this invention dynamically generates a pressure threshold by visually perceiving defects and material characteristics, utilizes a force sensing closed loop to achieve progressive, impact-free adhesion, and quantitatively assesses the risk of failure based on two dimensions: strain concentration and propagation rate. This ensures a positive correlation between risk level and optimization weight, driving a multi-objective genetic algorithm to optimize between obstacle avoidance distance and pressure uniformity. It also uses dual feedback iterations of modulus change rate and pressure standard deviation to ensure that the convergence process follows the material's mechanical response characteristics. Reinforcement learning adaptively adjusts speed and safety weights according to risk level to generate the optimal displacement trajectory, achieving interlayer spatiotemporal collaboration and forming a complete causal chain of perception-decision-execution-verification. This effectively avoids the propagation of concrete microcracks caused by pressure overload, significantly improving climbing efficiency and pier structural safety in complex environments.

[0154] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0155] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-destructive climbing control method for piers based on the fusion of vision and force sensing, characterized in that, include: Step S1: Collect the defect distribution information and material strength information of the pier surface in real time to construct a three-dimensional surface state vector, and generate the extrusion pressure threshold matrix of each target extrusion point by matching based on the preset surface state-pressure response relation database. Step S2: Perform extrusion control on each target extrusion point according to the extrusion pressure threshold matrix to obtain a preliminary adhesion state; The process of step S2 includes: Read the extrusion pressure threshold matrix to obtain the target pressure value corresponding to each target extrusion point in order to start the distributed collaborative extrusion process; According to the target pressure value, progressive pressurization is performed synchronously at each target extrusion point, while force sensor feedback data is collected in real time and the pressure deviation between the current pressure and the target pressure is calculated. When the pressure deviation value enters the preset transition range, the pressurization rate is automatically reduced and the micro-motion control mode is switched. When the pressure deviation value approaches zero, pressure locking is performed. The force balance index is calculated based on the pressure difference between each extrusion point. The output of each point is adjusted by pressure compensation so that the force balance index converges to a stable range to obtain the initial adhesion state. The average pressure value of all extrusion points is calculated as the force balance benchmark. The deviation between the real-time pressure of each extrusion point and the average pressure value is calculated. The root mean square value of the deviation of all extrusion points is used as the force balance index. Step S3: Monitor the surface deformation of the contact area of ​​each extrusion point according to the preliminary adhesion state, analyze and compare the images before and after extrusion based on the surface deformation to identify the failure evolution characteristics, and classify the failure evolution characteristics according to the degree of strain concentration and the rate of expansion to generate a risk level and its failure risk vector field; Step S4: Based on the destruction risk vector field, determine the priority and scheduling according to the risk level, and start a multi-objective optimization loop to determine the initial control strategy; Step S5: Based on the initial control strategy, perform a secondary scan of each extrusion point using vision and monitor the pressure distribution using a force sensor to obtain a convergence determination result. Step S6: Based on the convergence determination result and the risk level, after the current layer stabilizes, a displacement control trajectory corresponding to the risk level is generated through a reinforcement learning algorithm, and the next execution layer is activated simultaneously to repeat the process from step S1 to step S5 to obtain the target control strategy.

2. The non-destructive climbing control method for piers based on vision and force sensing fusion as described in claim 1, characterized in that, The process of step S1 includes: The entire surface of the pier is scanned from a distance, while close-range image acquisition is performed on the target areas of each compression point to obtain multi-scale image data. Defect semantic segmentation is performed on the multi-scale image data to identify defect types and output defect confidence vectors. At the same time, the concrete grade or coating type is determined based on the surface material texture and reflective properties to determine the surface bearing strength parameters. The defect confidence vector and the bearing strength parameter are concatenated along the channel dimension, and the local curvature change features calculated from the surface point cloud data are fused to construct the three-dimensional surface state vector. The three-dimensional surface state vector is input into the relational database, and the optimal extrusion pressure threshold for the current region is matched using a nonlinear interpolation algorithm to obtain the extrusion pressure threshold matrix.

3. The non-destructive climbing control method for piers based on vision and force sensing fusion as described in claim 2, characterized in that, The process of performing defect semantic segmentation on the multi-scale image data to identify defect types and output defect confidence vectors includes: A pre-trained convolutional neural network is used to perform pixel-level segmentation on the multi-scale image data to obtain a probability heatmap of the defect type; The probability heatmap is input into the defect connected component analysis algorithm to calculate the area, aspect ratio, and edge roughness parameters of each defect region. Based on the area, aspect ratio, and edge roughness parameters, a support vector machine classifier is used to determine the defect type and output the corresponding confidence vector.

4. The non-destructive climbing control method for piers based on vision and force sensing fusion as described in claim 3, characterized in that, The process of calculating the force balance index based on the pressure difference between each extrusion point and adjusting the output of each point through pressure compensation includes: When the force balance index exceeds the preset stable range, the extrusion point with the largest pressure deviation is identified as the adjustment target; a pressure retraction operation is performed on the extrusion point with pressure higher than the average value, and a pressure compensation operation is performed on the extrusion point with pressure lower than the average value. Based on the pressure distribution of adjacent extrusion points, the retraction pressure value is allocated to the extrusion points with lower pressure according to the distance weight ratio. The above calculation and adjustment process is repeated until the force balance index converges to the stable range.

5. The non-destructive climbing control method for piers based on vision and force sensing fusion as described in claim 4, characterized in that, The process of step S3 includes: Acquire a reference image before extrusion and a real-time image after extrusion, and perform affine registration on the reference image and the real-time image to obtain a registered image; Pixel-level difference operations are performed on the image to calculate the surface displacement vector field and convert it into a strain field distribution; In the strain field distribution, strain concentration regions are identified, and principal strain regions exceeding the material's elastic limit are marked as crack initiation zones, while shear strain localization zones are marked as spalling risk zones. The strain amplitude and the rate of expansion of the strain concentration region are quantitatively classified to generate a damage risk vector field containing three levels of risk (high, medium, and low) and spatial coordinate information. The process of performing pixel-level differencing on an image to calculate the surface displacement vector field and convert it into a strain field distribution includes: The displacement vector of each pixel in the registered image is calculated using the optical flow method to generate a dense displacement field. The displacement gradient tensor is obtained by performing spatial gradient calculation on the dense displacement field; The displacement gradient tensor is symmetrically decomposed to extract principal strain components and shear strain components to generate a strain field distribution characterizing the degree of surface deformation.

6. The non-destructive climbing control method for piers based on vision and force sensing fusion as described in claim 5, characterized in that, The process of step S4 includes: Based on the risk level and spatial distribution information of each region, high-risk areas are marked as priority adjustment targets to obtain the marking results; A multi-objective optimization scheduling queue is established based on the labeling results, and higher adjustment priority and greater adjustment weight are assigned to high-risk areas; An optimization model is constructed using a non-dominated sorting genetic algorithm, with the first objective being to maximize the distance to avoid weak areas and the second objective being to minimize the pressure variance of force distribution uniformity. The optimization model is iteratively solved, and the solution that best matches the current risk level is selected from the obtained frontier solution set as the initial control strategy.

7. The non-destructive climbing control method for piers based on vision and force sensing fusion as described in claim 6, characterized in that, The process of step S5 includes: According to the initial control strategy, each extrusion point is driven to adjust its attitude according to the reconstructed pressure field and topological configuration; After the posture adjustment is completed, secondary image data of the contact area of ​​each extrusion point is acquired by visual acquisition, and real-time pressure distribution data is obtained by force sensor. Deformation analysis is performed on the secondary image data to calculate the modulus change rate of the damage risk vector field, and the standard deviation of the pressure distribution at all extrusion points is calculated on the real-time pressure distribution data. When the magnitude of the risk vector field approaches zero and the standard deviation of the pressure distribution is less than the preset convergence threshold, the convergence is determined to be successful and the convergence determination result is output; when the convergence condition is not met, the current pressure distribution data is fed back to step S4 for further optimization.

8. The non-destructive climbing control method for piers based on vision and force sensing fusion as described in claim 7, characterized in that, The process of step S6 includes: Construct a reinforcement learning state space, including the relative position information of the current layer, the pressure distribution vector of each extrusion point, the surface defect distribution map of the pier column, and environmental disturbance parameters; Construct a reinforcement learning action space, including motion velocity, motion acceleration, and squeeze timing control parameters at each level; The reward function weights are set according to the risk level, and the reward function is constructed by weighting climbing speed, energy consumption and damage level. A proximal policy optimization algorithm is used to iterate the policy in the state space and action space to generate a motion trajectory that matches the current risk level. After the current layer completes the convergence determination, the next layer is driven to repeat the process of steps S1 to S5 to obtain the target control strategy.

Images