Automatic engineering crack detection equipment and method based on bionic quadruped robot

Abstract

The invention relates to an automatic engineering crack detection device and method based on a bionic quadruped robot. The method comprises the steps that the distance between two cameras of a binocular camera and lens parameters are set according to the size of a to-be-detected structure; the 3D-DIC binocular detection equipment is adjusted to be in a horizontal state by adjusting the binocular detection equipment horizontal adjusting support; the 3D-DIC binocular detection equipment is moved to the front of the to-be-detected structure through the bionic quadruped robot, and the posture of the bionic quadruped robot is adjusted by using the quadruped robot overall adjustment gradienter to ensure that the detected area of the to-be-detected structure is located in the field of view; collecting a speckle image sequence of the surface of the to-be-measured structure through a binocular camera under the assistance of a light supplementing lamp; according to a preset strain threshold value, 3D-DIC analysis is carried out on the speckle image sequence through a crack recognition algorithm, a crack area is obtained, and parameters of cracks in the crack area are quantified. Compared with the prior art, the reliability of crack detection is improved, and accurate detection of crack parameters is realized.

Description

Technical Field

[0001] This invention relates to the field of optical measurement technology, and in particular to an automated engineering crack detection device and method based on a bionic quadruped robot. Background Technology

[0002] Three-dimensional digital image correlation (3D-DIC) is a non-contact optical measurement technology that uses binocular cameras to simultaneously acquire speckle images of an object's surface and calculates displacement and strain fields using digital image correlation algorithms. Current technologies apply 3D-DIC to crack detection, initially covering areas such as teeth, track beds, and anti-icing structures. Although 3D-DIC technology is mature, its implementation in complex engineering scenarios such as underground areas and ruins relies on the performance of mobile carriers. Existing mobile carrier solutions mainly rely on fixed rail installations and manual hand-held movement, resulting in single, fixed application scenarios and significant limitations in mobility. Few technologies can overcome the limitations of fixed engineering scenarios in crack detection while simultaneously achieving automation and intelligence in mobile monitoring. For example, patent application CN117495758A discloses a method for detecting track bed cracks and fissures based on a 3D camera. This method uses a 3D camera to move along a preset track to acquire track color and depth images. It then eliminates interference from the boundaries of auxiliary equipment such as cables and signal devices on the track bed, and finally uses local dense sampling to identify real cracks or fissures, statistically analyzing the number, length, and maximum width parameters of the cracks. This method can only move along a preset track, resulting in fixed scenarios and severe limitations in mobility. Therefore, there is an urgent need to develop a new type of automated mobile carrier platform and a system deployment method to ensure its safe development in engineering crack detection. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art by providing an automated engineering crack detection device and method based on a bionic quadruped robot, which improves the reliability of crack detection and achieves accurate detection of crack parameters.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] An automated engineering crack detection device based on a bionic quadruped robot includes a 3D-DIC binocular inspection device and a bionic quadruped robot platform. The 3D-DIC binocular inspection device includes a slide rail support plate and a binocular camera and a supplementary light mounted on the slide rail support plate. The bionic quadruped robot platform includes a bionic quadruped robot and a quadruped robot overall level adjustment device. The 3D-DIC binocular inspection device is connected to the bionic quadruped robot through a binocular inspection device level adjustment bracket. The quadruped robot overall level adjustment device is located at the connection point between the binocular inspection device level adjustment bracket and the bionic quadruped robot.

[0006] Furthermore, the two cameras of the binocular camera are respectively mounted on the slide rail support plate via camera sliding supports.

[0007] Furthermore, the binocular detection device is equipped with an adjustment rod on its horizontal adjustment bracket.

[0008] According to another aspect of the present invention, an automated engineering crack detection method is provided that applies the automated engineering crack detection device based on a bionic quadruped robot as described above, comprising the following steps:

[0009] The spacing between the two cameras of the binocular camera and the lens parameters are set according to the size of the structure to be measured.

[0010] Adjust the 3D-DIC binocular inspection device to a horizontal position by adjusting the horizontal adjustment bracket of the binocular inspection device;

[0011] The 3D-DIC binocular inspection device is moved to the front of the structure to be tested by the bionic quadruped robot. The posture of the bionic quadruped robot is adjusted by the overall level of the quadruped robot to ensure that the area to be tested of the structure is in the field of view.

[0012] A sequence of speckle images of the surface of the structure under test is acquired using a binocular camera with the aid of a supplementary light.

[0013] Based on a preset strain threshold, the speckle image sequence is analyzed using a crack identification algorithm to automatically identify crack regions and quantify crack parameters in the crack regions.

[0014] Furthermore, the distance between the two cameras of the binocular camera can be adjusted on the slide rail support plate via a camera sliding bracket.

[0015] Furthermore, based on a preset strain threshold, the specific steps for automatically identifying crack regions by performing 3D-DIC analysis on the speckle image sequence using a crack identification algorithm include:

[0016] Image preprocessing is performed on the speckle image sequence of the surface of the structure under test;

[0017] The grayscale correlation of pixels in the speckle images of the left and right views obtained by the two cameras of the binocular camera is calculated by normalized cross-correlation algorithm, and the corresponding pixels are found and the disparity is calculated.

[0018] Based on the calibration parameters of the binocular camera, the two-dimensional image coordinates in the speckle images of the left and right views are converted into three-dimensional coordinates in the camera coordinate system. Using the parallax, combined with the focal length of the binocular camera and the horizontal distance between the two cameras, the depth information of each pixel is calculated.

[0019] By combining the depth information of each pixel with the image coordinates, three-dimensional point cloud data is generated, and the three-dimensional morphology of the surface of the structure under test is obtained.

[0020] The three-dimensional topography is projected onto the speckle image sequence after image preprocessing. Taking the first frame speckle image in the undeformed state in the speckle image sequence as a reference, the sub-regions of the speckle images in subsequent frames of the speckle image sequence are tracked and matched. The three-dimensional displacement vector of each sub-region is calculated by the inverse combination Gauss-Newton algorithm to obtain the displacement field.

[0021] The strain field is obtained by calculating the Green-Lagrange strain tensor based on the displacement field data;

[0022] The regions in the full-field strain distribution of the strain field where the principal strain value exceeds the strain threshold are initially identified as potential abnormal regions, and the crack edge data in the potential abnormal regions are obtained by the Canny edge detection algorithm.

[0023] The crack edge data is input into a pre-trained deep learning model, and combined with image morphology processing, a continuous crack region is obtained.

[0024] Furthermore, the image preprocessing includes distortion correction and image enhancement.

[0025] Furthermore, the deep learning model is a crack region segmentation model based on a convolutional neural network. It generates a crack probability map by inputting the crack edge data, and combines image morphological processing to perform continuity correction on the judgment results to obtain continuous crack regions. The crack probability map satisfies the following formula:

[0026]

[0027] In the formula, P(x,y,z) is the probability value of pixel (x,y,z) in the crack probability map belonging to the crack region, (x,y,z) is the three-dimensional coordinate of the pixel on the crack edge, σ is the Sigmoid function, N is the number of feature maps output by the convolutional layer, and w i f represents the weight parameters corresponding to the i-th convolutional feature map. i (x,y,z) represents the activation value of the i-th convolutional feature map at pixel (x,y,z), and b is the bias term.

[0028] Furthermore, the crack parameters include crack width, crack depth, and crack length.

[0029] Furthermore, the specific steps for quantifying the crack parameters in the crack region include:

[0030] Based on the three-dimensional topography of the surface of the structure under test, the three-dimensional point cloud data of the crack edge is projected onto a section perpendicular to the crack direction, and a continuous crack profile is extracted. Linear fitting is performed on the two edge points of the crack profile, and the average distance between the two fitted lines in the normal direction is calculated as the crack pixel width. Then, the crack pixel width is converted into the actual crack width according to the calibration coefficient.

[0031] By combining prior knowledge base and strain field distribution characteristics, an empirical model for crack width and depth is established. Based on the crack width, strain characteristic vector and geological characteristic vector of the current crack area, the crack depth is estimated to obtain the crack depth estimate.

[0032] Cracks identified in multiple consecutive speckle images are temporally matched and connected. The central axis of the crack is extracted and the crack pixel length is calculated. Then, the crack pixel length is converted into the actual crack length according to the calibration coefficient.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This invention sets up a 3D-DIC binocular inspection device on a bionic quadruped robot platform. By coordinating the horizontal adjustment bracket of the binocular inspection device and the overall level adjustment of the quadruped robot, the posture of the bionic quadruped robot is adjusted to ensure that the measured area of ​​the structure under test is located in the field of view. This enables unmanned and automated crack detection in complex environments, further expanding the applicability of 3D-DIC crack detection, greatly expanding the scope of inspection operations and application scenarios, while avoiding the safety hazards of personnel working in high-risk environments.

[0035] 2. This invention identifies potential abnormal regions by analyzing the strain field, providing a diagnosis from a mechanical mechanism perspective. This approach is more reliable than relying solely on image texture features, helping to detect latent damage and developing defects earlier. By combining a preset strain threshold, Canny edge detection, and a pre-trained deep learning model, it can more accurately distinguish real cracks from interfering information such as stains and scratches on the structural surface, effectively reducing the false alarm rate and improving the reliability of crack detection results. At the same time, it precisely quantifies the width, depth, and length of cracks, achieving accurate crack detection. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the structure of an automated engineering crack detection device based on a biomimetic quadruped robot proposed in this invention;

[0037] Figure 2 This is a schematic diagram of the structure of a 3D-DIC binocular inspection device;

[0038] Figure 3This is a schematic diagram of the process for an automated engineering crack detection method based on a biomimetic quadruped robot proposed in this invention.

[0039] Figure 4 A flowchart illustrating the process of automatic identification of crack areas.

[0040] Legend: 1. 3D-DIC binocular inspection device; 11. Camera; 12. Fill light; 13. Slide rail support plate; 14. Camera sliding support; 15. Binocular inspection device horizontal adjustment support; 16. Adjustment rod; 21. Bionic quadruped robot; 22. Quadruped robot overall level adjustment instrument. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0042] The following English abbreviations are involved:

[0043] Three-Dimensional Digital Image Correlation (3D-DIC)

[0044] Normalized Cross-Correlation (NCC) algorithm

[0045] Inverse Compositional Gauss-Newton Algorithm (ICGN)

[0046] Field of View (FOV)

[0047] Example 1

[0048] This embodiment provides an automated engineering crack detection device based on a biomimetic quadruped robot, such as... Figure 1 As shown, the system includes a 3D-DIC binocular inspection device 1 and a bionic quadruped robot platform. The 3D-DIC binocular inspection device 1 includes a slide rail support plate 13, a binocular camera 11 mounted on the slide rail support plate 13, and a supplementary light 12. The two cameras of the binocular camera 11 are respectively mounted on the slide rail support plate 13 via camera sliding supports 14. Figure 2As shown. The bionic quadruped robot platform includes a bionic quadruped robot 21 and a quadruped robot overall level adjustment device 22. The 3D-DIC binocular inspection device 1 is connected to the bionic quadruped robot 21 through a binocular inspection device level adjustment bracket 15. An adjustment rod 16 is provided on the binocular inspection device level adjustment bracket 15. The quadruped robot overall level adjustment device 22 is located at the connection between the binocular inspection device level adjustment bracket 15 and the bionic quadruped robot platform.

[0049] Example 2

[0050] This embodiment provides an automated engineering crack detection method based on a biomimetic quadruped robot-based automated engineering crack detection device, such as... Figure 3 As shown, it includes the following steps:

[0051] S1. Set the distance between the two cameras of the binocular camera 11 and the lens parameters according to the size of the structure to be measured.

[0052] The dimensions of the structure under test include length, width, height, and the surface area of ​​key inspection areas, such as bridge piers, tunnel linings, and dam panels. Lens parameters include focal length, aperture, and field of view (FOV).

[0053] The detection distance is determined as the vertical distance from the binocular camera 11 to the surface of the structure to be tested. This distance is limited by the reachable space of the bionic quadruped robot 21, the safe operating range, and the depth of field of the camera. It is usually set to 0.5-3.0m, depending on the accuracy requirements for crack identification of the structure.

[0054] The distance between the two cameras of the binocular camera 11 directly affects the depth measurement accuracy and point cloud density of stereo vision.

[0055] The spacing between the two cameras of the binocular camera 11 adopts the "baseline-distance ratio" principle. For fine crack detection (such as 0.05-0.2mm in width), it is recommended that B = 0.2d. For rapid screening of large-sized structures, it can be relaxed to B = 0.1d. B is the spacing between the two cameras of the binocular camera 11, and d is the detection distance.

[0056] The camera sliding support 14 slides precisely on the slide rail support plate 13, enabling stepless adjustment of the distance between the two cameras of the binocular camera 11 with an adjustment accuracy of ≤1mm. It is also equipped with a scale and locking device to ensure stable position after adjustment.

[0057] Lens parameters directly determine spatial resolution, depth of field, and image quality, and need to be optimized according to crack detection requirements. The focal length is chosen based on the following objectives: at a set detection distance, ensure that the field of view (FOV) of a single camera can cover the predetermined detection area, while guaranteeing that the crack occupies sufficient pixels in the image. Aperture settings must balance light intake and depth of field.

[0058] For example, when measuring the accuracy of crack depth, the scale interval of 3D-DIC in the depth direction is related to the distance between the two cameras of the binocular camera 11, the detection distance, the camera pixel size, and the focal length. To improve the measurement accuracy of crack depth, the scale interval in the depth direction should be reduced and the distance between the two cameras of the binocular camera 11 should be increased.

[0059] S2. Adjust the 3D-DIC binocular inspection device 1 to a horizontal state by adjusting the horizontal adjustment bracket 15 of the binocular inspection device.

[0060] In 3D-DIC measurement, the absolute horizontal orientation of the binocular camera is a key prerequisite for ensuring stereo matching accuracy and avoiding systematic errors introduced into 3D reconstruction. The 3D-DIC binocular inspection device 1 is adjusted to a horizontal state by adjusting the horizontal adjustment bracket 15 of the binocular inspection device.

[0061] S3. Move the 3D-DIC binocular inspection device 1 to the front of the structure to be tested using the bionic quadruped robot 21, and adjust the posture of the bionic quadruped robot 21 using the overall level adjustment instrument 22 of the quadruped robot to ensure that the measured area of ​​the structure to be tested is in the field of view.

[0062] The quadruped robot level 22 is fixed at the connection between the binocular detection equipment level adjustment bracket 15 and the top of the bionic quadruped robot 21. Through the real-time feedback of the quadruped robot level 22, the body posture of the bionic quadruped robot 21 is controlled so that the center of the common field of view of the binocular camera 11 is facing the area being measured.

[0063] S4. Acquire a sequence of speckle images of the surface of the structure under test using a binocular camera 11 with the assistance of a supplementary light 12.

[0064] S5. Based on the preset strain threshold, the speckle image sequence is analyzed by 3D-DIC using a crack identification algorithm to automatically identify the crack region and quantify the crack parameters in the crack region.

[0065] Based on a preset strain threshold, the crack identification algorithm performs 3D-DIC analysis on the speckle image sequence to automatically identify the crack region. The specific steps are as follows: Figure 4 As shown, it includes the following steps:

[0066] S501. Perform image preprocessing on the speckle image sequence of the surface of the structure to be tested.

[0067] Image preprocessing includes distortion correction and image enhancement. Distortion correction eliminates radial and tangential distortions caused by the camera lens, ensuring geometric fidelity of the image; the corrected speckle image sequence maintains a linear correspondence between pixel coordinates and physical space. Image enhancement improves speckle contrast and texture sharpness.

[0068] S502. Calculate the grayscale correlation of pixels in the left and right view speckle images obtained by the two cameras of the binocular camera 11 respectively through the normalized cross-correlation algorithm, find the same pixel and calculate the disparity.

[0069] The left and right view speckle images, after image preprocessing in step S501, are converted into grayscale images. Gaussian filtering is applied to the images to suppress high-frequency noise and improve matching stability. A window centered on the target pixel is selected in the left and right view speckle images, and the similarity of their grayscale distribution is calculated. For each pixel in the left view speckle image, a sliding window is used in the same row of the right view speckle image to perform pixel matching, finding the most similar pixel pair. Based on finding the most similar pixel pair, a quadratic polynomial fitting method is used to estimate the sub-pixel level matching position, find the corresponding pixel, and calculate the disparity. Disparity is the horizontal displacement of the corresponding pixel in the left and right view speckle images. Disparity is inversely proportional to depth information; the larger the disparity, the smaller the depth; the smaller the disparity, the larger the depth.

[0070] S503. Based on the calibration parameters of the binocular camera 11, the two-dimensional image coordinates in the speckle images of the left and right views are converted into three-dimensional coordinates in the camera coordinate system. Using parallax, combined with the focal length of the binocular camera 11 and the horizontal distance between the two cameras, the depth information of each pixel is calculated.

[0071] Before performing 3D reconstruction, the stereo camera 11 needs to be calibrated to obtain its intrinsic and extrinsic parameters. The intrinsic parameters include focal length and principal point coordinates, while the extrinsic parameters include rotation matrix and translation vector. Camera parameters are crucial for converting 2D image coordinates into 3D coordinates in the camera coordinate system.

[0072] Using the disparity values ​​in the disparity map, the focal length of the binocular camera 11, and the horizontal distance between the two cameras, the depth information of each pixel is calculated using the following formula:

[0073]

[0074] In the formula, Z is the pixel depth, f is the camera focal length, B is the horizontal distance between the two cameras, and D is the parallax.

[0075] S504. Combine the depth information of each pixel with the image coordinates to generate three-dimensional point cloud data, and obtain the three-dimensional morphology of the surface of the structure under test.

[0076] The depth information of each pixel is combined with the image coordinates to generate 3D point cloud data. Each point cloud data contains three coordinate values ​​(x, y, z), representing the position of the point in 3D space. Through the processing and analysis of the point cloud data, the 3D morphology of the structure's surface is reconstructed.

[0077] S505. Project the three-dimensional topography onto the speckle image sequence after image preprocessing. Using the first frame speckle image in the undeformed state in the speckle image sequence as a reference, track and match the sub-regions of the speckle images in subsequent frames of the speckle image sequence. Calculate the three-dimensional displacement vector of each sub-region using the ICGN algorithm to obtain the displacement field.

[0078] The 3D topography is mapped and aligned with the preprocessed speckle image sequence to provide a geometric basis for subsequent displacement field calculations. The first frame of the speckle image in the undeformed state is defined as the reference frame, and all subsequent measurements in the deformed state are compared with this frame. On the reference frame image, a virtual computational grid is divided according to the region to be measured. The grid nodes typically correspond to physical points or pixel centers of the displacement field to be calculated. A sub-region is defined with each grid node as its center. The size of the sub-region needs to be adjusted according to the speckle characteristics and the expected degree of deformation.

[0079] The displacement is calculated iteratively using the ICGN algorithm, and the specific steps include:

[0080] First-order shape functions are used to describe the displacement changes of points within a subregion. In the three-dimensional case, the shape function is expressed as follows:

[0081]

[0082] In the formula, W(x; p) is a shape function, x is a vector, and p = [u, u] x ,u y ,v,v x ,v y ,w,w x ,w y ] T Let u be the displacement parameter vector to be determined. x Let u be the first-order partial derivative of the displacement u in the x-direction. y Let v be the first partial derivative of displacement u in the y-direction, where u is the translational displacement of the sub-region center point in the x-direction. x Let v be the first partial derivative of displacement v in the x-direction. y Let w be the first partial derivative of displacement v in the y-direction, where v is the translational displacement of the sub-region center point in the y-direction. x Let W be the first-order partial derivative of the displacement w in the x-direction. yLet Δx be the first-order partial derivative of displacement w in the y-direction, Δx be the x-coordinate offset of any point within the sub-region relative to the center point of the sub-region, and Δy be the y-coordinate offset of any point within the sub-region relative to the center point of the sub-region.

[0083] The IC-GN algorithm uses the zero-mean normalized minimum squared distance standard as its objective function:

[0084]

[0085] In the formula, C ZNSSD (p) represents the objective function, and F(x) represents the gray value of the reference frame sub-region at position x. Let S be the average gray value of the reference frame, S be the sub-region, and G(W(x;p)) be the gray value of the corresponding position in the target frame sub-region after shape function transformation. The average gray value of the target frame sub-region.

[0086] The IC-GN algorithm optimizes the displacement parameter vector p to make the objective function C ZNSSD (p) Minimize and solve for the optimal displacement parameter vector. Repeat the above IC-GN algorithm process for each sub-region in the computational grid to solve for the displacement parameter vector of the center point of each sub-region, and extract the three-dimensional displacement vector of each grid node from it. The displacement field is formed by the set of three-dimensional displacement vectors of all grid nodes.

[0087] S506. Calculate the Green-Lagrange strain tensor based on the displacement field data to obtain the strain field;

[0088] The displacement gradient tensor and deformation gradient tensor are calculated based on the displacement field data, and the Green-Lagrange strain tensor is further calculated. Eigenvalue decomposition is performed on the Green-Lagrange strain tensor at each point in the displacement field to obtain the principal strains. The above calculation process is repeated for all points within the computational region to obtain the strain distribution across the entire field, thus yielding the strain field.

[0089] S507. Regions in the full-field strain distribution of the strain field where the principal strain value exceeds the strain threshold are initially identified as potential abnormal regions, and crack edge data in the potential abnormal regions are obtained through the Canny edge detection algorithm.

[0090] The strain threshold is determined based on prior knowledge of the material's mechanical properties (such as yield strain), experimental experience, or by statistically analyzing the strain distribution across the entire field. By iterating through each calculation point across the entire field and comparing its principal strain value with the strain threshold, all points whose principal strain values ​​exceed the strain threshold are marked; the region formed by these points is the potential anomaly region.

[0091] The Canny edge detection algorithm extracts clear, continuous, and single-pixel-wide boundaries from potential anomaly regions, resulting in a binary image. The white pixels (edges) constitute the crack edge data, which are continuous, single-pixel-wide curves that clearly outline the shape and direction of the potential crack.

[0092] S508. Input the crack edge data into a pre-trained deep learning model and combine it with image morphology processing to obtain a continuous crack region.

[0093] The deep learning model is a crack region segmentation model based on convolutional neural networks. It generates a crack probability map by inputting crack edge data, and combines image morphological processing to perform continuity correction on the judgment results to obtain continuous crack regions. The crack probability map satisfies the following formula:

[0094]

[0095] In the formula, P(x,y,z) is the probability value of pixel (x,y,z) in the crack probability map belonging to the crack region, (x,y,z) is the three-dimensional coordinate of the pixel on the crack edge, σ is the Sigmoid function, N is the number of feature maps output by the convolutional layer, and w i f represents the weight parameters corresponding to the i-th convolutional feature map. i (x,y,z) represents the activation value of the i-th convolutional feature map at pixel (x,y,z), and b is the bias term.

[0096] Crack parameters include crack width, crack depth, and crack length. The specific steps for quantifying crack parameters within a crack region include:

[0097] When quantifying the crack width, based on the three-dimensional topography of the surface of the structure under test, the three-dimensional point cloud data of the crack edge is projected onto a section perpendicular to the crack direction, and a continuous crack profile is extracted. Linear fitting is performed on the two edge points of the crack profile, and the average distance between the fitted lines on both sides in the normal direction is calculated as the crack pixel width. Then, the crack pixel width is converted into the actual crack width according to the calibration coefficient.

[0098] When quantifying fracture depth, an empirical model of fracture width and depth is established by combining a prior knowledge base with strain field distribution characteristics. Based on the fracture width, strain characteristic vector, and geological characteristic vector of the current fracture area, the fracture depth is estimated to obtain the estimated value. The prior knowledge base typically originates from historical exploration data, core experimental records, geomechanical research reports, expert experience, and relevant literature. This knowledge is transformed into usable parameters and constraints for the model, mainly including: rock mechanics parameters, such as the elastic modulus, Poisson's ratio, tensile strength, and fracture toughness of rock strata; strain field distribution, stress magnitude and direction; geological structural information, such as the distribution and occurrence of weak structural planes like bedding planes, faults, and joints; and a database of typical fracture morphologies, storing fracture patterns generated under different stress conditions based on historical data or physical simulations.

[0099] The specific steps for establishing an empirical model for crack width and depth include: extracting strain characteristics, strain gradients, and the area exceeding the critical strain threshold from strain field distribution data, and estimating the crack width; querying the mechanical parameters, critical strain threshold, and geological structure information of the current rock mass from a prior knowledge base, and fusing these two types of information to form a comprehensive strain feature vector. A machine learning model is then used to train the empirical model, with prior knowledge and the comprehensive strain feature vector as input and crack depth as output.

[0100] For the new fracture depth estimation scenario, the fracture width is calculated, and the strain field distribution in the area is obtained, extracting the strain feature vector. The current rock mass conditions and geological environment are matched with the prior knowledge base to obtain the geological feature vector. The fracture width, strain feature vector, and geological feature vector are then substituted into the calibrated empirical model to calculate the estimated fracture depth.

[0101] When quantifying the crack length, cracks identified in multiple consecutive speckle images are temporally matched and connected, the central axis of the crack is extracted and the crack pixel length is calculated, and then the crack pixel length is converted into the actual crack length according to the calibration coefficient.

[0102] The rest is the same as in Example 1.

[0103] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An automated engineering crack detection device based on a biomimetic quadruped robot, characterized in that, The system includes a 3D-DIC binocular inspection device (1) and a bionic quadruped robot platform. The 3D-DIC binocular inspection device (1) includes a slide rail support plate (13) and a binocular camera (11) and a fill light (12) mounted on the slide rail support plate (13). The bionic quadruped robot platform includes a bionic quadruped robot (21) and a quadruped robot overall level adjustment device (22). The 3D-DIC binocular inspection device (1) is connected to the bionic quadruped robot (21) through a binocular inspection device level adjustment bracket (15). The quadruped robot overall level adjustment device (22) is located at the connection between the binocular inspection device level adjustment bracket (15) and the bionic quadruped robot (21).

2. The automated engineering crack detection equipment based on a bionic quadruped robot according to claim 1, characterized in that, The two cameras of the binocular camera (11) are respectively mounted on the slide rail support plate (13) via camera sliding supports (14).

3. The automated engineering crack detection device based on a bionic quadruped robot according to claim 1, characterized in that, An adjustment rod (16) is provided on the horizontal adjustment bracket (15) of the binocular detection equipment.

4. An automated engineering crack detection method using an automated engineering crack detection device based on a bionic quadruped robot as described in any one of claims 1-3, comprising the following steps: The spacing and lens parameters between the two cameras of the binocular camera (11) are set according to the size of the structure to be measured; Adjust the 3D-DIC binocular inspection device (1) to a horizontal position by adjusting the horizontal adjustment bracket (15) of the binocular inspection device; The 3D-DIC binocular inspection device (1) is moved to the front of the structure to be tested by the bionic quadruped robot (21), and the posture of the bionic quadruped robot (21) is adjusted by the overall adjustment level of the quadruped robot (22) to ensure that the measured area of ​​the structure to be tested is in the field of view. A sequence of speckle images of the surface of the structure under test is acquired by a binocular camera (11) with the assistance of a supplementary light (12); Based on a preset strain threshold, the speckle image sequence is analyzed using a crack identification algorithm to automatically identify crack regions and quantify crack parameters in the crack regions.

5. The automated engineering crack detection method according to claim 4, characterized in that, The distance between the two cameras of the binocular camera (11) can be adjusted on the slide rail support plate (13) by means of the camera sliding support (14).

6. The automated engineering crack detection method according to claim 4, characterized in that, Based on a preset strain threshold, the specific steps for automatically identifying crack regions by performing 3D-DIC analysis on the speckle image sequence using a crack identification algorithm include: Image preprocessing is performed on the speckle image sequence of the surface of the structure under test; The gray-level correlation of pixels in the speckle images of the left and right views obtained by the two cameras of the binocular camera (11) is calculated by normalized cross-correlation algorithm, and the same-named pixels are found and the disparity is calculated. According to the calibration parameters of the binocular camera (11), the two-dimensional image coordinates in the speckle images of the left and right views are converted into three-dimensional coordinates in the camera coordinate system. Using the parallax, combined with the focal length of the binocular camera (11) and the horizontal distance between the two cameras, the depth information of each pixel is calculated. By combining the depth information of each pixel with the image coordinates, three-dimensional point cloud data is generated, and the three-dimensional morphology of the surface of the structure under test is obtained. The three-dimensional topography is projected onto the speckle image sequence after image preprocessing. Taking the first frame speckle image in the undeformed state in the speckle image sequence as a reference, the sub-regions of the speckle images in subsequent frames of the speckle image sequence are tracked and matched. The three-dimensional displacement vector of each sub-region is calculated by the inverse combination Gauss-Newton algorithm to obtain the displacement field. The strain field is obtained by calculating the Green-Lagrange strain tensor based on the displacement field data; The regions in the full-field strain distribution of the strain field where the principal strain value exceeds the strain threshold are initially identified as potential abnormal regions, and the crack edge data in the potential abnormal regions are obtained by the Canny edge detection algorithm. The crack edge data is input into a pre-trained deep learning model, and combined with image morphology processing, a continuous crack region is obtained.

7. The automated engineering crack detection method according to claim 6, characterized in that, The image preprocessing includes distortion correction and image enhancement.

8. The automated engineering crack detection method according to claim 6, characterized in that, The deep learning model is a crack region segmentation model based on a convolutional neural network. It generates a crack probability map by inputting the crack edge data, and combines image morphological processing to perform continuity correction on the judgment result to obtain continuous crack regions. The crack probability map satisfies the following formula: In the formula, P(x,y,z) is the probability value of pixel (x,y,z) in the crack probability map belonging to the crack region, (x,y,z) is the three-dimensional coordinate of the pixel on the crack edge, σ is the Sigmoid function, N is the number of feature maps output by the convolutional layer, and w i f represents the weight parameters corresponding to the i-th convolutional feature map. i (x,y,z) represents the activation value of the i-th convolutional feature map at pixel (x,y,z), and b is the bias term.

9. The automated engineering crack detection method according to claim 4, characterized in that, The crack parameters include crack width, crack depth, and crack length.

10. The automated engineering crack detection method according to claim 9, characterized in that, The specific steps for quantifying crack parameters in the crack region include: Based on the three-dimensional topography of the surface of the structure under test, the three-dimensional point cloud data of the crack edge is projected onto a section perpendicular to the crack direction, and a continuous crack profile is extracted. Linear fitting is performed on the two edge points of the crack profile, and the average distance between the two fitted lines in the normal direction is calculated as the crack pixel width. Then, the crack pixel width is converted into the actual crack width according to the calibration coefficient. By combining prior knowledge base and strain field distribution characteristics, an empirical model for crack width and depth is established. Based on the crack width, strain characteristic vector and geological characteristic vector of the current crack area, the crack depth is estimated to obtain the crack depth estimate. Cracks identified in multiple consecutive speckle images are temporally matched and connected. The central axis of the crack is extracted and the crack pixel length is calculated. Then, the crack pixel length is converted into the actual crack length according to the calibration coefficient.

Images