Anchoring rod support quality detection method based on visual discrimination

Abstract

The present application relates to the field of engineering detection technology, more specifically to a kind of anchor rod support quality detection method based on visual discrimination.Based on the mobile platform of multiple-view color camera and depth camera, the image and pose data of anchor rod area are collected, after image preprocessing, the instance segmentation network with multi-modal feature fusion is used to realize anchor rod segmentation and three-dimensional positioning;Relying on point cloud extraction, the surface parameters such as anchor rod shape, arrangement and tray are extracted, and the deformation of tray and surrounding rock is obtained by sub-pixel DIC technology;Through simulation and test calibration of deformation pre-tightening force model, the anchor rod pre-tightening force is inversed and the parameters are corrected;Relying on the analytic hierarchy process, a multi-dimensional evaluation model is established to complete quality classification, and finally three-dimensional modeling is used to mark defective anchor rod and output detection report.The present application is non-contact and non-destructive testing, with high degree of automation, solving the problems of traditional anchor rod detection sampling damage and large artificial determination error, and can be used for batch quality detection of roadway and tunnel anchor rod support.

Description

Technical Field

[0001] This invention relates to the field of engineering testing technology, and more specifically to a visual discrimination-based method for testing the quality of anchor bolt support. Background Technology

[0002] Anchor bolt support is one of the most widely used support methods in geotechnical engineering, and its quality directly affects the overall safety and stability of the project. Currently, the quality inspection of anchor bolt support mainly relies on traditional methods. Pull-out test: This is a destructive test that can only be performed on samples and cannot achieve comprehensive testing. In addition, the testing process is time-consuming and labor-intensive and poses safety hazards. Ultrasonic testing method: Its accuracy is greatly affected by the surrounding rock conditions, anchor bolt material and anchoring agent distribution. The manual visual inspection method is highly subjective and can only detect surface features such as the exposed length and number of anchor bolts, but cannot detect key quality indicators such as preload and anchorage length.

[0003] In recent years, computer vision-based inspection technology has developed rapidly. However, existing visual inspection methods can only identify basic parameters such as the number of anchor bolts, exposed length, and spacing. There are no effective visual inspection methods for the core indicators that determine the anchor bolt support effect—preload and anchoring quality.

[0004] Therefore, how to provide a visual discrimination-based method for detecting the quality of anchor bolt support is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a visual discrimination-based anchor bolt support quality inspection method, which aims to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A visual-based method for detecting the quality of anchor bolt support includes the following steps: S1. Using a mobile inspection platform equipped with multiple industrial cameras, color images of the anchor bolt support in the area to be inspected are acquired from multiple different perspectives. Simultaneously, the three-dimensional depth information of the area is acquired, and the real-time position and attitude data of the inspection platform are recorded. S2. The acquired color image and depth image are sequentially processed with distortion correction, denoising, enhancement and spatial registration to obtain an aligned multi-view fused color image and corresponding depth image; S3. Using a trained multimodal instance segmentation neural network model, the fused color image and depth image are jointly processed to identify and segment each anchor rod. The three-dimensional spatial position and spatial attitude of each anchor rod are determined by combining depth information and platform attitude data. S4. Based on the segmentation mask and 3D point cloud data of the anchor bolt, extract multiple surface quality features of the anchor bolt; S5. Based on digital image correlation technology, compare the baseline image and the detection image after the anchor bolt is installed, extract the displacement field and strain field distribution of the anchor bolt tray and the surrounding rock surface, and obtain the tray warping deformation, the surrounding rock surface settlement and the width of the surrounding rock crack. S6. Based on the extracted pallet warping deformation, call the pre-established mapping model from pallet deformation to preload to calculate the actual preload of each anchor bolt. S7. Based on the surface quality characteristics of the anchor bolt, the actual preload and the deformation characteristics of the surrounding rock, establish a multi-dimensional evaluation model to classify and evaluate the support quality of each anchor bolt. S8. Generate an inspection report that includes anchor bolt locations, various quality indicators, and evaluation levels, and mark the anchor bolt locations of different quality levels in the three-dimensional space model.

[0007] Furthermore, the multiple industrial cameras mentioned in S1 include at least three color industrial cameras arranged from different angles and one depth camera. The mobile detection platform is equipped with a three-axis stabilized gimbal, an odometer, and an inertial measurement unit to compensate for platform vibration and provide accurate position and attitude information.

[0008] Furthermore, the multimodal instance segmentation neural network model described in S3 adds a multimodal feature fusion module after the backbone network, which fuses the texture features of the color image with the spatial features of the depth image before inputting them into subsequent network branches. The model is trained using a labeled image dataset containing different lighting conditions, different rock types, and different anchor bolt states.

[0009] Furthermore, the determination of the three-dimensional spatial position and spatial orientation of each anchor rod in S3 specifically involves: generating a three-dimensional point cloud of the anchor rod based on the segmentation mask of the anchor rod and the corresponding depth image; performing linear fitting on the three-dimensional point cloud to obtain the anchor rod axis equation; and then calculating the inclination angle and azimuth angle of the anchor rod, using the center coordinates of the tray as the three-dimensional spatial position of the anchor rod.

[0010] Furthermore, the multiple surface quality characteristics described in S4 include the exposed length of the anchor bolt, the diameter of the bolt body, the spacing between adjacent anchor bolts, the row spacing between adjacent anchor bolts, the flatness of the tray, the number of turns of the nut tightening, and the length of the exposed thread.

[0011] Furthermore, the digital image correlation technology described in S5 employs a sub-pixel-level matching algorithm, which obtains the displacement value with sub-pixel precision by iteratively solving the extreme value of the correlation coefficient function, thereby detecting the minute deformation of the tray and the surrounding rock surface.

[0012] Furthermore, the mapping model from pallet deformation to preload described in S6 obtains a theoretical relationship curve by simulating the deformation law of the pallet under different preloads through finite element analysis. The theoretical curve is then corrected using indoor pull-out test data. The model input includes the amount of pallet warping deformation and the material and size parameters of the pallet.

[0013] Furthermore, S6 also includes correcting the initial preload obtained by inversion, adjusting the initial preload according to the spatial inclination angle of the anchor bolt and the mechanical parameters of the surrounding rock, to obtain the final actual preload.

[0014] Furthermore, the multi-dimensional evaluation model described in S7 uses the analytic hierarchy process to determine the weight of each evaluation index, calculates the comprehensive score of each anchor bolt, and classifies the anchor bolt support quality into multiple levels based on the comprehensive score.

[0015] Furthermore, the three-dimensional spatial model described in S8 is reconstructed based on the collected depth data and platform attitude data, and different colors are used on the model to distinguish anchor bolts of different quality levels.

[0016] Compared to existing technologies, this invention eliminates the need for contact with the anchor bolts, making the inspection process safe, efficient, and without damaging the anchor bolt support structure. It not only detects the surface quality characteristics of the anchor bolts but also calculates their preload, enabling a comprehensive assessment of the anchor bolt support quality. Employing sub-pixel-level digital image correlation technology, it can detect minute deformations, achieving a preload inversion accuracy of ±5%. The entire inspection process is automated by computer, significantly improving efficiency and reducing labor costs. Furthermore, it generates a three-dimensional spatial inspection model, visually displaying the location of substandard anchor bolts, facilitating rectification by engineers. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a visual discrimination-based anchor support quality inspection method according to the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0020] See appendix Figure 1 This invention discloses a visual discrimination-based anchor bolt support quality inspection method. Its core principle is to acquire complete geometric information of the anchor bolt support system through multi-view visual acquisition, use deep learning to achieve accurate identification and positioning of the anchor bolt, extract minute deformation features based on digital image correlation technology, and then invert the anchor bolt preload, a core quality indicator. Finally, a comprehensive and objective evaluation of the anchor bolt support quality is achieved through a multi-dimensional evaluation model.

[0021] This step uses a track-type mobile detection platform, which consists of a walking mechanism, a pan-tilt unit, an image acquisition unit, a lighting unit, a control unit, and a data storage unit.

[0022] It adopts a wheeled track walking mechanism with a maximum walking speed of 1.5m / s and a positioning accuracy of ±5mm. It is equipped with an odometer and an IMU (Inertial Measurement Unit) to achieve real-time measurement of the platform's position and attitude. It uses a three-axis stabilized gimbal with a horizontal rotation range of ±180°, a vertical rotation range of ±90°, and a stabilization accuracy of ±0.01° to compensate for vibration and tilt during platform movement.

[0023] Three high-resolution color industrial cameras: global shutter CMOS cameras with a resolution of 2448×2048 pixels, a frame rate of 30fps, and a lens focal length of 16mm, are respectively mounted on the front, left, and right sides of the pan-tilt unit, with angles of 0°, 30°, and -30° to the vertical direction. The optical axes of the three cameras intersect on the detection plane 2m in front of the platform.

[0024] A structured light depth camera with a resolution of 1280×720 pixels, a depth measurement range of 0.5-5m, and a depth measurement accuracy of ±2mm is used. It is mounted in the center of the pan-tilt unit and acquires data synchronously with a color camera. Four sets of high-brightness LED strip lights, each with a power of 50W and a color temperature of 6500K, are evenly distributed around the image acquisition unit to provide shadowless illumination and avoid the impact of shadows on image quality. An industrial control computer, equipped with an Intel Core i7-12700H processor, 32GB of RAM, and a 1TB solid-state drive, running Ubuntu 20.04, is used to control the entire system and handle data processing. A 2TB high-speed solid-state drive is used to store the acquired image data and detection results.

[0025] The image acquisition process is as follows: The camera is preheated, and the gimbal is calibrated, the odometer is zeroed, and the IMU is initialized. Within 24 hours of the anchor bolt installation, and provided the surrounding rock has not undergone significant deformation, a baseline image of the area to be inspected is acquired. When quality inspection is required, the mobile inspection platform is controlled to move at a constant speed of 0.5 m / s along the tunnel axis. When the odometer records a movement distance of 0.2 m, the control unit simultaneously triggers three color cameras and one depth camera to acquire images, while recording the current platform's position and attitude information. The acquired image data and corresponding position and attitude information are stored in the data storage unit.

[0026] This step involves a series of preprocessing operations on the acquired color and depth images to eliminate image noise and distortion, improve image quality, and lay the foundation for subsequent anchor bolt recognition and feature extraction.

[0027] Before system use, the Zhang Zhengyou calibration method was used to calibrate three color cameras and a depth camera, obtaining the intrinsic parameter matrix, distortion coefficients, and extrinsic parameter matrix between the cameras. A 12×9 checkerboard calibration board was used, with each square measuring 30mm×30mm. At least 20 images of the calibration board were acquired at different angles and distances, and calibration was performed using the `calibrateCamera` function from the OpenCV library. After calibration, reprojection error verification was performed to ensure that the reprojection error was less than 0.5 pixels.

[0028] The acquired color and depth images are corrected using the calibrated intrinsic parameter matrix and distortion coefficients. For radial distortion, a second-order polynomial model is used for correction; for tangential distortion, a first-order polynomial model is used. The correction formulas are as follows: , , in, These are the pixel coordinates in the distorted image. These are the corrected pixel coordinates. The distance from a pixel to the center of the image. , The radial distortion coefficient is... , denoted as the tangential distortion coefficient.

[0029] A Gaussian filtering algorithm was used to denoise the corrected image, removing Gaussian noise and salt-and-pepper noise. The Gaussian filter kernel size was 5×5, and the standard deviation σ=1.5. For invalid depth values ​​(black holes) in the depth image, a neighborhood-based interpolation algorithm was used to fill them.

[0030] The Contrast-Limited Adaptive Histogram Equalization (CLAHE) algorithm is used to enhance the contrast of color images, highlighting the edges and details of anchor bars. The CLAHE algorithm divides the image into 8×8 blocks, performs histogram equalization on each block separately, and then stitches the blocks together using bilinear interpolation. Simultaneously, the contrast amplification factor is limited to 4.0 to avoid noise amplification caused by over-enhancement.

[0031] Based on the feature point matching algorithm, images acquired by three color cameras and images acquired by a depth camera are registered. The specific steps are as follows: The SIFT algorithm was used to extract feature points and feature descriptors for each image.

[0032] The feature descriptors are matched using a brute-force matching algorithm to obtain initial matching pairs.

[0033] The RANSAC algorithm is used to remove mismatched pairs, and the homography matrix between images is calculated.

[0034] Based on the homography matrix, the color images of the left and right sides are transformed into the coordinate system of the front color image.

[0035] The three transformed color images are then weighted and averaged to obtain a multi-view fused color image.

[0036] The depth image is transformed to the coordinate system of the frontal color image to obtain a depth image aligned with the color image.

[0037] This step uses an improved Mask R-CNN model to process multi-view fused images, achieving accurate instance segmentation and 3D spatial localization of anchor bolts; Based on the original Mask R-CNN model, a multimodal feature fusion module was added to fuse the texture features of color images and the spatial features of depth images, thereby improving the model's robustness to different lighting conditions and surrounding rock types.

[0038] ResNet-50-FPN is used as the backbone network to extract multi-scale features of the image. The feature maps of the color image and the depth image are input into this module, the number of channels is adjusted by 1×1 convolution, and then element-wise addition is performed to fuse them to obtain the fused feature map. Candidate regions are generated on the fused feature map. Features are extracted from each candidate region to obtain a fixed-size feature map. It is divided into three branches for object classification, bounding box regression and mask generation, respectively.

[0039] A total of 10,000 anchor bolt images were collected from different mines, roadways, lighting conditions, surrounding rock types, and anchor bolt states. The LabelMe tool was used to manually annotate the anchor bolts in the images, including their bounding boxes and segmentation masks. The dataset was then divided into training, validation, and test sets in a 7:2:1 ratio.

[0040] Data augmentation operations such as random flipping, random rotation, random scaling, random cropping, brightness adjustment, and contrast adjustment are performed on the training set images to expand the dataset size and improve the model's generalization ability.

[0041] An improved Mask R-CNN model was trained using the PyTorch deep learning framework. The training parameters were set as follows: an initial learning rate of 0.001, a learning rate decay policy of decreasing to 0.1 every 10 epochs, a batch size of 8, 30 training epochs, an AdamW optimizer, and loss functions consisting of classification loss, bounding box regression loss, and mask loss. After training, the model was evaluated on the test set, achieving an anchor recognition accuracy of 98.5% and a mean intersection-union (mIoU) of 0.92 for the segmented masks.

[0042] For each segmented anchor bolt, the 3D world coordinates of each point on the anchor bolt are calculated based on its pixel coordinates in the color image and the corresponding depth image. The Random Sample Consensus (RANSAC) algorithm is used to fit a straight line to the 3D point cloud of the anchor bolt to obtain the anchor bolt's axis equation. Based on the anchor bolt's axis equation, the spatial attitude parameters of the anchor bolt are calculated, including the anchor bolt's tilt angle (the angle with the vertical direction) and azimuth angle (the angle with the tunnel axis). The center coordinates of the anchor bolt tray are calculated as the 3D spatial position of the anchor bolt.

[0043] Based on the segmented anchor mask and 3D point cloud data, the following surface quality features are accurately extracted: Separate the anchor rod body point cloud and the tray point cloud from the 3D point cloud of the anchor rod. Calculate the average plane equation of the tray point cloud. Find the point in the anchor rod body point cloud that is farthest from the tray plane. This point is the exposed end of the anchor rod. Calculate the vertical distance from the exposed end point to the tray plane, which is the exposed length of the anchor rod.

[0044] In the point cloud of the anchor rod body, 10 sections perpendicular to the anchor rod axis are selected uniformly. The point cloud of each section is fitted with a circle to obtain the diameter of the section. The average value of the diameters of the 10 sections is calculated, which is the average diameter of the anchor rod.

[0045] Based on the three-dimensional spatial position of all anchors, establish the spatial coordinate matrix of the anchors. For each anchor, find the adjacent anchors in the same row and column, calculate the horizontal distance between the centers of the adjacent anchors, which is the anchor spacing, and calculate the vertical distance between the centers of the adjacent anchors, which is the row spacing of the anchors.

[0046] Plane fitting is performed on the pallet point cloud to obtain the reference plane equation of the pallet. The distance from each point in the pallet point cloud to the reference plane is calculated, and the difference between the maximum and minimum distances is the flatness error of the pallet.

[0047] The image of the nut region is extracted from the anchor mask. Hough transform is used to detect the edge and center of the nut. A pre-marked white scale line on the nut is identified. The angle between the scale line and the horizontal direction is calculated. Based on this angle, the number of turns to tighten the nut is calculated using the following formula: Number of turns = integer number of turns + included angle / 360° The integer number of turns is obtained by recording the initial and final positions of the nut during installation.

[0048] Extract the image of the threaded area at the outer end of the nut from the anchor bolt mask, use an edge detection algorithm to detect the edge of the thread, and calculate the length of the threaded area, which is the exposed thread length.

[0049] All extracted surface quality features are compared with the design values, the deviation values ​​are calculated, and it is determined whether they meet the design requirements.

[0050] This step is based on digital image correlation (DIC) technology, which accurately extracts the deformation field distribution characteristics of the anchor plate and the surrounding rock surface by comparing the baseline image and the detection image.

[0051] Digital image correlation techniques calculate the displacement of a sub-region by selecting a sub-region (called the reference sub-region) in a reference image and then finding the sub-region (called the target sub-region) in the detection image that has the highest correlation with the reference sub-region. By performing point-by-point calculations on the entire image, the displacement field of the entire image can be obtained.

[0052] To improve displacement measurement accuracy, this invention employs a sub-pixel-level matching algorithm based on Newton-Raphson iteration. This algorithm iteratively solves for the sub-pixel values ​​of the displacement by performing a Taylor expansion of the correlation coefficient function. The specific steps of the algorithm are as follows: A reference sub-region of size (2M+1)×(2M+1) is selected in the baseline image, where M is typically 15-31. In the detection image, a larger search region is selected, centered on the pixel position corresponding to the center of the reference sub-region. An integer pixel search algorithm is used to find the integer pixel position with the highest correlation coefficient. Using this integer pixel position as the initial value, the Newton-Raphson iterative method is used to solve for the sub-pixel value of the displacement until the iteration converges. The above steps are repeated to calculate the displacement field of the entire image. The sub-pixel-level DIC algorithm used in this invention can detect minute displacements at the 0.01 pixel level. Under the imaging conditions of this system, the deformation detection accuracy can reach 0.001 mm.

[0053] Displacement data of the pallet area is extracted from the displacement field, and the displacement difference between the pallet center and the pallet edge is calculated, which is the amount of pallet warping deformation. Displacement data of the surrounding rock area is extracted from the displacement field, and the vertical displacement of each point on the surrounding rock surface is calculated. The maximum value is taken as the settlement of the surrounding rock surface. Cracks on the surrounding rock surface are detected using the Canny edge detection algorithm, and the cracks are refined to obtain the centerline of the crack. The distance between corresponding points on both sides of the crack is calculated, which is the width of the crack.

[0054] This step uses the extracted pallet warping deformation amount and a pre-established "pallet deformation-preload" mapping model to accurately calculate the actual preload of each anchor bolt.

[0055] Finite element models of pallets made of different materials (Q235 steel, Q355 steel) and with different sizes (diameters of 150mm, 200mm, and 250mm, thicknesses of 8mm, 10mm, and 12mm) were established using ANSYS finite element analysis software. Solid186 elements were used for mesh generation with a mesh size of 2mm. The boundary condition was that the bottom of the pallet was in contact with the surrounding rock, with a surface-to-surface contact and a friction coefficient of 0.3. Anchor bolts passed through the center of the pallet, and an axial tensile force was applied to the top of the anchor bolts. The warping deformation of the pallets under different preloads (10kN, 20kN, 30kN, 40kN, 50kN, 60kN, 70kN, and 80kN) was simulated, and the theoretical relationship curve between pallet deformation and preload was obtained.

[0056] An indoor anchor pull-out test bench was constructed, consisting of a reaction frame, hydraulic jacks, force sensors, and displacement sensors. Pull-out tests were conducted on pallets of different materials and sizes, and the actual warping deformation of the pallets under different preloads was recorded. The test data were compared with the finite element simulation results, and the theoretical relationship curve was corrected to obtain the final mapping model of pallet deformation to preload. The model was fitted using a third-order polynomial, in the following form: , Where F is the anchor bolt preload. This represents the amount of warpage deformation of the pallet. , , , These are polynomial coefficients, with different coefficients corresponding to different materials and sizes of pallets.

[0057] For each anchor bolt, a mapping model for the deformation of the anchor plate to the preload is selected based on the material and size of its support plate. The extracted anchor plate warpage deformation is substituted into the mapping model to calculate the initial preload of the anchor bolt. The initial preload is then corrected based on the anchor bolt's inclination angle and the elastic modulus of the surrounding rock to obtain the final actual preload. The correction formula is as follows: in, This is the actual preload. This is the initial preload. Let be the inclination angle of the anchor bolt. Let be the elastic modulus of the surrounding rock. The correction factor was determined experimentally.

[0058] This step combines the surface quality characteristics, preload, and surrounding rock deformation characteristics of the anchor bolts to establish a multi-dimensional anchor bolt support quality evaluation model, and conducts a scientific and objective graded assessment of the support quality of each anchor bolt.

[0059] Exposed length deviation (weight 0.1), diameter deviation (weight 0.05), spacing deviation (weight 0.05), row spacing deviation (weight 0.05), pallet flatness (weight 0.05), preload compliance rate (weight 0.5), surrounding rock settlement (weight 0.1), surrounding rock crack width (weight 0.1).

[0060] The weights of each evaluation index were determined using the Analytic Hierarchy Process (AHP). The specific steps are as follows: A hierarchical model was established, with the evaluation objective, primary indicators, and secondary indicators serving as the objective layer, criterion layer, and scheme layer, respectively. A judgment matrix was constructed, and 10 experts in the field of geotechnical engineering were invited to score the relative importance of each indicator. The maximum eigenvalue and eigenvector of the judgment matrix were calculated, and the weights of each indicator were obtained. A consistency check was performed to ensure that the consistency ratio CR of the judgment matrix was less than 0.1.

[0061] Quantification methods and scoring standards for each evaluation indicator were developed, and a 100-point scale was used for scoring, as shown in Table 1: Table 1

[0062] The overall score for each anchor bolt is calculated using the following formula: Overall score = Σ(indicator score × indicator weight).

[0063] Based on the overall score, the quality of anchor bolt support is divided into three levels: Excellent: Overall score ≥ 90 points, support quality is good, and it can meet the safety requirements of the project.

[0064] Pass: 60 points ≤ overall score < 90 points. The support quality basically meets the requirements, but monitoring needs to be strengthened.

[0065] Unqualified: Overall score <60 points, indicating serious hidden dangers in the quality of the support, requiring immediate rectification.

[0066] Using depth data collected by a depth camera and attitude data collected by an IMU, a 3D point cloud model of the area to be inspected is reconstructed. The 3D position, attitude, and quality inspection results of each anchor are superimposed on the 3D point cloud model. Different colors are used to mark anchors of different quality levels: green indicates excellent, yellow indicates qualified, and red indicates unqualified.

[0067] This embodiment takes the quality inspection of anchor bolt support in an underground roadway of a coal mine as an example to illustrate the specific implementation process of the present invention.

[0068] The underground roadway of this coal mine has a rectangular cross-section, with a width of 4.5m and a height of 3.5m. It is supported by anchor bolts with a specification of φ20mm×2400mm and a pallet specification of 200mm×200mm×10mm. The design preload is 50kN, the anchor bolt spacing is 800mm, the row spacing is 800mm, and the design value of the exposed length is 150mm.

[0069] Using the mobile detection platform of this invention, the tunnel was moved at a speed of 0.5 m / s along the tunnel axis, and a set of images was collected at 0.2 m intervals. A total of 100 m of tunnel images were collected, resulting in 500 sets of image data.

[0070] After image preprocessing, anchor bolt instance segmentation, feature extraction, preload inversion, and quality evaluation, a total of 1250 anchor bolts were detected. The detection results are summarized below: Excellent anchor bolts: 875, accounting for 70%; Qualified anchor bolts: 300, accounting for 24%; 75 anchor bolts were substandard, accounting for 6% of the total. The main problems with substandard anchor bolts include: Insufficient preload: 45 anchor bolts, accounting for 60% of the substandard anchor bolts; Exposed length exceeds standard: 15 bolts, accounting for 20% of the substandard anchor bolts; Pallet flatness exceeded the standard: 10 pieces, accounting for 13.3% of the substandard anchor bolts; Five anchor bolts had rock fissures exceeding the standard width, accounting for 6.7% of the substandard anchor bolts. Twenty anchor rods were randomly selected for pull-out tests. The results showed that the relative error between the preload inverted by the present invention and the actual preload measured by the pull-out test was less than 5%, which verified the accuracy of the test results of the present invention.

[0071] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0072] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting the quality of anchor bolt support based on visual discrimination, characterized in that, Includes the following steps: S1. Using a mobile inspection platform equipped with multiple industrial cameras, color images of the anchor bolt support in the area to be inspected are acquired from multiple different perspectives. Simultaneously, the three-dimensional depth information of the area is acquired, and the real-time position and attitude data of the inspection platform are recorded. S2. The acquired color image and depth image are sequentially processed with distortion correction, denoising, enhancement and spatial registration to obtain an aligned multi-view fused color image and corresponding depth image; S3. Using a trained multimodal instance segmentation neural network model, the fused color image and depth image are jointly processed to identify and segment each anchor rod. The three-dimensional spatial position and spatial attitude of each anchor rod are determined by combining depth information and platform attitude data. S4. Based on the segmentation mask and 3D point cloud data of the anchor bolt, extract multiple surface quality features of the anchor bolt; S5. Based on digital image correlation technology, compare the baseline image and the detection image after the anchor bolt is installed, extract the displacement field and strain field distribution of the anchor bolt tray and the surrounding rock surface, and obtain the tray warping deformation, the surrounding rock surface settlement and the width of the surrounding rock crack. S6. Based on the extracted pallet warping deformation, call the pre-established mapping model from pallet deformation to preload to calculate the actual preload of each anchor bolt. S7. Based on the surface quality characteristics of the anchor bolt, the actual preload and the deformation characteristics of the surrounding rock, establish a multi-dimensional evaluation model to classify and evaluate the support quality of each anchor bolt. S8. Generate an inspection report that includes anchor bolt locations, various quality indicators, and evaluation levels, and mark the anchor bolt locations of different quality levels in the three-dimensional space model.

2. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The multiple industrial cameras mentioned in S1 include at least three color industrial cameras arranged from different angles and one depth camera. The mobile detection platform is equipped with a three-axis stabilized gimbal, an odometer, and an inertial measurement unit to compensate for platform vibration and provide accurate position and attitude information.

3. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The multimodal instance segmentation neural network model described in S3 adds a multimodal feature fusion module after the backbone network, which fuses the texture features of the color image with the spatial features of the depth image before inputting them into subsequent network branches. The model is trained using a labeled image dataset containing different lighting conditions, different rock types, and different anchor bolt states.

4. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The determination of the three-dimensional spatial position and spatial orientation of each anchor rod in S3 specifically involves: generating a three-dimensional point cloud of the anchor rod based on the segmentation mask of the anchor rod and the corresponding depth image; performing linear fitting on the three-dimensional point cloud to obtain the anchor rod axis equation; and then calculating the inclination angle and azimuth angle of the anchor rod, using the center coordinates of the tray as the three-dimensional spatial position of the anchor rod.

5. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The surface quality characteristics described in S4 include the exposed length of the anchor bolt, the diameter of the bolt body, the spacing between adjacent anchor bolts, the row spacing between adjacent anchor bolts, the flatness of the tray, the number of turns of the nut tightening, and the length of the exposed thread.

6. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The digital image correlation technology described in S5 employs a sub-pixel-level matching algorithm, which obtains sub-pixel precision displacement values ​​by iteratively solving the extreme values ​​of the correlation coefficient function, thereby detecting minute deformations on the tray and surrounding rock surfaces.

7. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The mapping model from pallet deformation to preload described in S6 is established by simulating the deformation law of the pallet under different preloads through finite element analysis to obtain the theoretical relationship curve. The theoretical curve is then corrected by indoor pull-out test data. The model input includes the pallet warping deformation amount and the material and size parameters of the pallet.

8. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, S6 also includes correcting the initial preload obtained by inversion, adjusting the initial preload according to the spatial inclination angle of the anchor bolt and the mechanical parameters of the surrounding rock, to obtain the final actual preload.

9. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The multi-dimensional evaluation model described in S7 uses the analytic hierarchy process to determine the weight of each evaluation indicator, calculates the comprehensive score of each anchor bolt, and classifies the anchor bolt support quality into multiple levels based on the comprehensive score.

10. The method for detecting the quality of anchor bolt support based on visual discrimination according to claim 1, characterized in that, The three-dimensional spatial model described in S8 is reconstructed based on the collected depth data and platform attitude data, and different colors are used on the model to distinguish anchor bolts of different quality levels.

Images