Wall-climbing robot weld tracking method, device and wall-climbing robot

Abstract

This application provides a method, apparatus, and wall-climbing robot for tracking weld seams. The method includes: when the wall-climbing robot is at the starting position of the weld seam to be tracked, acquiring a weld seam image of the weld seam using a camera on the wall-climbing robot; inputting the weld seam image into a pre-trained rotating target detection model, outputting the coordinates of the corner points of the rotating bounding box of the weld seam; determining the current pose of the wall-climbing robot relative to the weld seam based on the coordinates of the corner points of the rotating bounding box; calculating the current offset and yaw angle of the wall-climbing robot based on the current pose of the wall-climbing robot relative to the weld seam and a predefined standard pose of the wall-climbing robot relative to the weld seam, and inputting these values ​​into a pre-trained differential speed adjustment prediction model, outputting the differential speed adjustment of the wall-climbing robot; determining a set of velocities for the wall-climbing robot based on the differential speed adjustment, and controlling the wall-climbing robot's movements according to the set of velocities to track the weld seam to be tracked.

Description

Technical Field

[0001] This application relates to the field of wall-climbing robot technology, and in particular to a wall-climbing robot weld seam tracking method, apparatus and wall-climbing robot. Background Technology

[0002] Weld seam tracking, as a key technology in automated welding and intelligent manufacturing processes, is widely used in industrial welding robots, mobile robots, and special-purpose robots. In high-risk or complex environments such as special equipment manufacturing, ship hull structure welding, and steel structure engineering, weld seams are often located on irregular surfaces in space. Traditional fixed industrial robots struggle to meet the requirements for weld seam accessibility and flexible movement. Therefore, wall-climbing robots have been extensively researched and applied.

[0003] Existing methods for tracking weld seams in wall-climbing robots mainly rely on pre-set trajectories or use single-point sensors such as lasers and ultrasound to obtain offsets, making it difficult to accurately track weld seams. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus and a wall-climbing robot for tracking weld seams, so as to achieve high-precision, real-time tracking of weld seams.

[0005] Specifically, this application is implemented through the following technical solution:

[0006] The first aspect of this application provides a method for tracking weld seams in a wall-climbing robot, the method comprising:

[0007] When the wall-climbing robot is at the starting position of the weld seam to be tracked, the camera mounted on the wall-climbing robot captures an image of the weld seam to be tracked; wherein, the camera is mounted on a cantilever extending from the right side wall of the wall-climbing robot, and is used to capture an image of the weld seam located on the right side of the wall-climbing robot; when the wall-climbing robot is at the starting position of the weld seam to be tracked, the weld seam to be tracked is within the field of view of the camera;

[0008] The weld image is input into a pre-trained rotating target detection model, which performs target detection on the weld image and outputs the coordinates of the corner points of the rotating bounding box of the weld.

[0009] The current pose of the wall-climbing robot relative to the weld is determined based on the coordinates of the corner points of the rotating bounding box; wherein the current pose of the wall-climbing robot relative to the weld is characterized by the coordinates of the center point of the rotating bounding box and the rotation angle of the rotating bounding box.

[0010] Based on the current pose of the wall-climbing robot relative to the weld and the predefined standard pose of the wall-climbing robot relative to the weld, calculate the current offset and yaw angle of the wall-climbing robot.

[0011] The offset and the yaw angle are input into a pre-trained differential speed adjustment prediction model, and the differential speed adjustment prediction model outputs the differential speed adjustment of the wall-climbing robot based on the offset and the yaw angle.

[0012] A set of speeds for the wall-climbing robot is determined based on the differential speed adjustment, and the wall-climbing robot's movements are controlled according to the set of speeds to enable the wall-climbing robot to track the weld seam to be tracked; wherein, the set of speeds includes the speeds of the left wheel and the right wheel of the wall-climbing robot.

[0013] A second aspect of this application provides a tracking device for a wall-climbing robot, the device comprising a data acquisition module, a detection module, a determination module, a calculation module, and a control module, wherein...

[0014] The acquisition module is used to acquire images of the weld seam to be tracked via a camera mounted on the wall-climbing robot when the wall-climbing robot is at the starting position of the weld seam to be tracked; wherein, the camera is mounted on a cantilever extending from the right side wall of the wall-climbing robot and is used to acquire images of the weld seam located on the right side of the wall-climbing robot; when the wall-climbing robot is at the starting position of the weld seam to be tracked, the weld seam to be tracked is within the field of view of the camera;

[0015] The detection module is used to input the weld image into a pre-trained rotating target detection model, and the rotating target detection model performs target detection on the weld image and outputs the coordinates of the corner points of the rotating bounding box of the weld.

[0016] The determining module is used to determine the current pose of the wall-climbing robot relative to the weld seam based on the coordinates of the corner points of the rotating bounding box; wherein the current pose of the wall-climbing robot relative to the weld seam is characterized by the coordinates of the center point of the rotating bounding box and the rotation angle of the rotating bounding box.

[0017] The calculation module is used to calculate the current offset and yaw angle of the wall-climbing robot based on the current pose of the wall-climbing robot relative to the weld and the predefined standard pose of the wall-climbing robot relative to the weld.

[0018] The calculation module is also used to input the offset and the yaw angle into a pre-trained differential speed adjustment prediction model, and the differential speed adjustment prediction model outputs the differential speed adjustment of the wall-climbing robot based on the offset and the yaw angle.

[0019] The control module is used to determine a set of speeds for the wall-climbing robot based on the differential speed adjustment, and to control the wall-climbing robot's movements based on the set of speeds so that the wall-climbing robot tracks the weld seam to be tracked; wherein, the set of speeds includes the speeds of the left wheel and the right wheel of the wall-climbing robot.

[0020] A third aspect of this application provides a wall-climbing robot, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described in the above embodiments.

[0021] The wall-climbing robot tracking method, apparatus, and wall-climbing robot provided in this application, when the wall-climbing robot is at the starting position of the weld seam to be tracked, acquires an image of the weld seam using a camera mounted on the wall-climbing robot. The weld seam image is then input into a pre-trained rotating target detection model, which performs target detection on the weld seam image, outputs the coordinates of the corner points of the rotating bounding box of the weld seam, and determines the current pose of the wall-climbing robot relative to the weld seam based on the coordinates of the corner points of the rotating bounding box. Based on the current pose of the wall-climbing robot relative to the weld seam and a predefined standard position of the wall-climbing robot relative to the weld seam... The robot's current offset and yaw angle are calculated and input into a pre-trained differential speed adjustment prediction model. This model then outputs the differential speed adjustment amount based on these values. Finally, a set of velocities for the robot is determined based on the differential speed adjustment amount, and the robot's movements are controlled accordingly to track the weld seam. This establishes an integrated weld seam tracking process of "image perception—pose estimation—intelligent control," enabling high-precision, real-time tracking of the weld seam path by the robot under complex working conditions. First, weld seam images are acquired via a camera, providing continuous and reliable image input to the rotating target detection model. This model accurately represents the spatial pose of the weld seam, thereby accurately quantifying the robot's lateral offset and heading deviation relative to the weld seam. Based on this, the differential speed adjustment prediction model outputs wheel speed adjustment amounts, allowing the robot to smoothly and accurately approach the weld seam trajectory within the control closed loop. The aforementioned image perception, pose estimation, and intelligent control form a tightly coupled technical chain, supporting each other and driving the whole system, ensuring that the system has good robustness, real-time performance, and environmental adaptability, thereby significantly improving the tracking accuracy of the wall-climbing robot in automated welding operations. Attached Figure Description

[0022] Figure 1A flowchart of Embodiment 1 of the wall-climbing robot weld seam tracking method provided in this application;

[0023] Figure 2 This is a schematic diagram of a wall-climbing robot shown in an exemplary embodiment of this application;

[0024] Figure 3 This is a schematic diagram of an initial rotating target detection model shown in an exemplary embodiment of this application;

[0025] Figure 4 This is a schematic diagram illustrating pose deviation as shown in an exemplary embodiment of this application;

[0026] Figure 5 This is a schematic diagram of the structure of Embodiment 1 of the wall-climbing robot weld seam tracking device provided in this application;

[0027] Figure 6 This is a hardware structure diagram of a wall-climbing robot, in which the weld seam tracking device of this application is located. Detailed Implementation

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0029] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0030] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0031] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0032] Figure 1 This is a flowchart of an embodiment of the wall-climbing robot weld seam tracking method provided in this application. Please refer to... Figure 1The method provided in this embodiment may include:

[0033] S101. When the wall-climbing robot is at the starting position of the weld seam to be tracked, the weld seam image of the weld seam to be tracked is acquired by the camera mounted on the wall-climbing robot.

[0034] Specifically, Figure 2 This is a schematic diagram illustrating a wall-climbing robot as an exemplary embodiment of this application. Please refer to... Figure 2 The wall-climbing robot is equipped with a camera. Specifically, in one possible implementation, the camera is mounted on a cantilever extending from the right side of the robot, with a field of view covering 180°, used to acquire images of the weld seam to be tracked located on the right side of the robot. That is, during the tracking of the weld seam, which is located on the right side of the robot, the robot moves to the left of the weld seam, while the camera on the cantilever extending from the right side of the robot acquires images of the weld seam.

[0035] In practice, when using a wall-climbing robot to track a weld seam, the operator places the robot at the starting position of the weld seam to be tracked, which is located to the left of the starting point of the weld seam. It should be noted that the specific location of the starting point is selected by the operator according to requirements, and this application does not limit it.

[0036] Optionally, in one possible implementation, during the movement of the wall-climbing robot, the weld seam to be tracked is in a dark box environment completely shielded from external light, and the camera acquires images of the weld seam based on its own light source.

[0037] It should be noted that the installation position of the light source is set according to actual needs, and is not limited in this application. For example, in one embodiment, the light source is arranged around the camera at a side-tilted angle. In this way, the illumination range of the light source can not only accurately cover the weld area, but also effectively avoid interference from possible specular reflections around the weld to be tracked, stabilize the distinction between the weld and its surroundings, and thus improve the quality of the weld image captured by the camera.

[0038] Optionally, in one possible implementation, the method further includes:

[0039] After each acquisition of a weld seam image, the quality of the weld seam image is evaluated to obtain a quality score, and the angle of the light source is adjusted based on the quality score and the pose of the wall-climbing robot relative to the weld seam.

[0040] It should be noted that appropriate evaluation methods can be selected to assess the quality of weld seam images according to actual needs, and this application does not limit such methods. For example, in one embodiment, the quality of weld seam images can be assessed based on image statistical features, including contrast, sharpness, noise level, and illumination uniformity. Specifically, the contrast can be quantified by selecting the gray-scale mean difference or edge gradient amplitude between the weld seam region and the background, the sharpness can be measured by the high-frequency energy ratio or local gradient variance, the noise level can be assessed by the standard deviation of flat areas or frequency domain noise energy, and the illumination uniformity can be analyzed by combining the block gray-scale mean variance. Then, the image features of each weld seam image are normalized and fused into a comprehensive quality score in the range of 0 to 1 according to a preset weight. If the comprehensive quality score of the weld seam image is lower than the threshold, the light source angle is adjusted or the camera exposure is optimized to optimize the conditions for the camera to acquire weld seam images. In this way, it can be effectively ensured that the weld seam images acquired by the camera can meet the requirements for high-precision weld seam feature extraction, thereby improving the accuracy of the rotating target detection model in target detection of weld seam images.

[0041] S102. Input the weld image into a pre-trained rotating target detection model, and have the rotating target detection model perform target detection on the weld image and output the coordinates of the corner points of the rotating bounding box of the weld.

[0042] Specifically, the rotating target detection model can accurately locate the input weld seam image. In this application, the weld seam image is input into the rotating target detection model, which can output the coordinate values ​​of the four corner points of the rotating bounding box corresponding to the weld seam image. The rotating bounding box can closely fit the actual direction of the weld seam, thereby capturing the accurate geometric and angular information of the weld seam and providing a key data foundation for subsequent calculations.

[0043] Furthermore, in one possible implementation, the training process of the pre-trained rotating target detection model includes:

[0044] Step 1: Construct an initial rotating target detection model based on the YOLOv11-OBB model.

[0045] Specifically, Figure 3 This is a schematic diagram of an initial rotating target detection model shown in an exemplary embodiment of this application. Please refer to... Figure 3The initial rotating target detection model is built based on the YOLOv11-OBB model and includes a backbone network, a neck network, and a detection head. A channel attention module and a spatial attention module are sequentially connected between the SPPF module and the C2PSA module of the backbone network. In the bounding box regression branch of the detection head, ordinary convolution is replaced with a DCNv4 deformable convolution module. The channel attention module performs global average pooling and global max pooling on the input feature map to obtain a first channel description vector and a second channel description vector. Based on the first and second channel description vectors, attention weights are determined, and then the feature map is channel-weighted according to these attention weights to obtain a first processed feature map. The spatial attention module performs max pooling and average pooling on the first processed feature map along the channel dimension to obtain a first spatial feature map and a second spatial feature map. Based on the first and second spatial feature maps, spatial attention weights are determined, and then the first processed feature map is spatially weighted according to these spatial attention weights to output a second processed feature map.

[0046] Specifically, the backbone network in the initial rotated object detection model is primarily responsible for extracting features from the input image. See also Figure 3 After the input image is processed by the module before SPPF, a feature map is obtained. This feature map is then input into the SPPF module, which consists of multiple parallel max pooling layers (including 5×5, 9×9, and 13×13). These layers are used to perform multi-scale pooling on the input feature map and integrate the multi-scale pooled features through serial fusion to obtain a feature map containing information at different scales.

[0047] Furthermore, the channel attention module, connected to the SPPF module, receives feature maps containing information at different scales from the SPPF module and performs global average pooling and global max pooling on these feature maps. Understandably, the channel attention module includes a global average pooling layer and a global max pooling layer. The global average pooling layer performs global average pooling on the feature map in the spatial dimension to obtain the first channel description vector; the global max pooling layer performs global max pooling on the feature map in the spatial dimension to obtain the second channel description vector. Then, the first and second channel description vectors are processed separately through a shared multilayer perceptron. The multilayer perceptron includes two fully connected layers and one activation layer. The first fully connected layer compresses the number of channels in the first and second channel description vectors (e.g., reducing the number of channels from...). compression channel , (For compression ratio), the activation layer uses the ReLU activation function to introduce a non-linear relationship, and then passes through a second fully connected layer to reduce the number of channels in the first and second channel description vectors from... Restore to After processing by a multilayer perceptron, the first and second channel description vectors are added together to obtain a fused channel feature vector. This fused feature vector is then normalized using the sigmoid function, setting the output value range to [0,1] to obtain the attention weights. Subsequently, the input feature map can be channel-weighted according to these attention weights to obtain the first processed feature map.

[0048] Furthermore, the spatial attention module performs max pooling on the input feature map (the first processed feature map) along the channel dimension to obtain the first spatial feature map, and then performs average pooling on the input feature map (the first processed feature map) along the channel dimension to obtain the second spatial feature map. Next, the first and second spatial feature maps are concatenated along the channel dimension to obtain the concatenated feature map. Then, a convolutional layer (with a kernel size of 7×7) is used to process the concatenated feature map, followed by normalization using the Sigmoid function, with the output value set between [0,1] to obtain the spatial attention weights. Finally, the first processed feature map is spatially weighted according to the obtained spatial attention weights to output the second processed feature map.

[0049] Furthermore, the second processed feature map output by the spatial attention module is input into the C2PSA module. The C2PSA module adopts a parallel branch design, with one branch being a standard convolution and the other branch incorporating a spatial attention mechanism. The two branches extract the second processed feature map in parallel to obtain the convolution result and the spatial attention result. Then, the convolution result and the spatial attention result are fused to obtain the further enhanced features.

[0050] Furthermore, the neck network, located between the backbone network and the detection head, is responsible for generating multi-scale feature maps based on the enhanced features from the backbone network, thereby providing the detection head with rich and highly discriminative features.

[0051] Furthermore, the detection head includes a bounding box classification branch and a bounding box regression branch, wherein the bounding box classification branch is used to predict the target category; and the bounding box regression branch is used to predict the target location.

[0052] It should be noted that you should continue to refer to... Figure 3In this embodiment, the ordinary convolution is replaced by a DCNv4 deformable convolution module in the bounding box regression branch of the detection head. The DCNv4 deformable convolution module can adaptively adjust the sampling position of the convolution kernel according to the shape of the target. In this way, the initial rotating target detection model can better adapt to the shape change of the rotating target and improve the accuracy of rotating target detection.

[0053] Step 2: Train the initial rotating target detection model to obtain the pre-trained rotating target detection model.

[0054] Specifically, during the first preset number of epochs of training the initial rotating target detection model, the parameters of the backbone network are frozen, and only the parameters of the neck network and the detection head are trained.

[0055] The specific value of the preset quantity is set according to actual needs, and is not limited in this embodiment. For example, in one possible implementation, the specific value of the preset quantity is 50.

[0056] Specifically, for example, in the first 50 epochs of the training process, the initial rotating target detection model receives the input image, extracts features through the backbone network, then performs feature fusion through the neck network, and finally performs target detection and localization by the detection head. The loss is then calculated based on the prediction results and the true label, and the parameters of the neck network and the detection head are updated using the gradient descent method. In this way, it can prevent the large target features in the weld seam image captured by the wall-climbing robot from destroying the pre-trained small target knowledge when the pose of the wall-climbing robot relative to the weld seam is too different from the standard pose.

[0057] Furthermore, after completing a preset number of epochs of training, the subsequent training continues to obtain a pre-trained rotating target detection model.

[0058] Based on the preceding description, after obtaining the pre-trained rotating target detection model, the weld image is input into this model, which then outputs the coordinates of the corner points of the detected weld's rotated bounding box. Here, the weld's rotated bounding box refers to the rectangular frame containing the weld identified by the rotating target detection model in the weld image; the shape and position of this bounding box are related to the shape and position of the weld in the image. Furthermore, the coordinates of the corner points of the rotated bounding box refer to the coordinates of the four corner points of this rectangular frame, which allow for accurate location of the weld.

[0059] The wall-climbing robot weld seam tracking method provided in this embodiment is based on the YOLOv11-OBB model. It embeds channel attention and spatial attention modules into the backbone network, and combines this with multi-scale pooling feature fusion using the SPPF module. This significantly enhances the model's sensitivity to weld seam texture and geometry, enabling precise focusing on the differential features between the weld seam edge and the background. Furthermore, the detection head uses a DCNv4 deformable convolution module to adaptively adjust the convolution kernel sampling position, effectively capturing the deformation characteristics of the weld seam boundary and improving the positioning accuracy of curved or inclined weld seams. In addition, during training... In the first preset number of epochs of the initial rotating target detection model, the parameters of the backbone network are frozen, and the parameters of the neck network and the detection head are optimized first. This not only preserves the general feature extraction capability of the backbone network pre-training, but also avoids the interference of large-size target features caused by initial pose deviations with the small target knowledge in the pre-trained weights. Thus, in the joint training after unfreezing, the rotating target detection model can achieve stable detection of multi-scale and multi-angle welds. In this way, it can finally achieve highly robust and high-precision rotating bounding box output in complex industrial scenarios, providing reliable input for subsequent differential speed regulation control.

[0060] S103. Determine the current pose of the wall-climbing robot relative to the weld seam based on the coordinates of the corner points of the rotating bounding box.

[0061] The current pose of the wall-climbing robot relative to the weld seam is characterized by the coordinates of the center point of the rotating bounding box and the rotation angle of the rotating bounding box.

[0062] Specifically, by calculating the average of the coordinates of the four corner points of the rotated bounding box, the coordinates of the center point of the rotated bounding box can be obtained.

[0063] For example, suppose the coordinates of the four corner points of the rotated bounding box are corner point 1 , , corner point 2 , , corner point 3 , and corner point 4 , The coordinates of the center point of the rotated bounding box can be calculated using the following formula:

[0064] ;

[0065] in, These are the coordinates of the center point of the rotated bounding box.

[0066] Furthermore, the rotation angle of the rotated bounding box is the angle between the longer side of the rotated bounding box and the horizontal direction. In practice, the distance between each adjacent point can be calculated based on the coordinates of the four corner points to determine the two longest sides of the rotated bounding box. Then, the angle between the longest side and the horizontal direction can be calculated based on the coordinates of the two corner points of one of the two longest sides, thus obtaining the rotation angle of the rotated bounding box.

[0067] For example, in one possible implementation, the rotation angle of the rotated bounding box can be calculated using the following formula:

[0068] ;

[0069] in, This represents the rotation angle of the bounding box. ; ; .

[0070] Furthermore, based on the coordinates of the center point of the rotated bounding box and the rotation angle of the rotated bounding box, the pose of the wall-climbing robot relative to the weld can be determined. In this embodiment, since the pose of the camera relative to the wall-climbing robot is fixed, the pose of the weld captured by the current camera is used to represent the pose of the current wall-climbing robot relative to the weld, and the pose of the weld is represented by the coordinates of its center point and the rotation angle.

[0071] S104. Calculate the current offset and yaw angle of the wall-climbing robot based on the current pose of the wall-climbing robot relative to the weld and the predefined standard pose of the wall-climbing robot relative to the weld.

[0072] It should be noted that the predefined standard pose of the wall-climbing robot relative to the weld is the pose when the weld is in a horizontal position and the center of the weld coincides with the center of the image. Figure 4 This is a schematic diagram illustrating pose deviation in an exemplary embodiment of this application. Please refer to... Figure 4 The specific implementation process of this step may include:

[0073] Step 1: Determine the rotation angle as the current yaw angle of the wall-climbing robot.

[0074] Step 2: Based on the current pose of the wall-climbing robot relative to the weld seam and the coordinates of the weld seam center in the predefined standard pose of the wall-climbing robot relative to the weld seam, determine the current offset of the wall-climbing robot along the weld seam direction and the current offset of the wall-climbing robot along the vertical direction; wherein, the vertical direction is the direction perpendicular to the weld seam direction.

[0075] In practice, based on the current pose of the wall-climbing robot relative to the weld seam and the coordinates of the weld seam center in the predefined standard pose of the wall-climbing robot relative to the weld seam, the current offset of the wall-climbing robot along the weld seam direction and the current offset of the wall-climbing robot along the vertical direction can be calculated using the following homogeneous equations:

[0076] ;

[0077] in, This represents the current offset of the wall-climbing robot along the weld seam direction. This represents the current vertical offset of the wall-climbing robot. This represents the current yaw angle of the wall-climbing robot. , This represents the current pose of the wall-climbing robot relative to the weld seam. , The coordinates of the weld center relative to the standard pose of the weld seam for the predefined wall-climbing robot.

[0078] Step 3: Determine the current offset of the wall-climbing robot along the vertical direction as the current offset of the wall-climbing robot.

[0079] In this step, the current vertical offset of the wall-climbing robot is taken as the current offset of the wall-climbing robot. This ensures that the vertical alignment between the wall-climbing robot and the weld seam is guaranteed first, so that the position between the wall-climbing robot and the weld seam will not be significantly offset, thus achieving a precise weld seam tracking effect.

[0080] S105. Input the offset and the yaw angle into the pre-trained differential speed adjustment prediction model, and the differential speed adjustment prediction model outputs the differential speed adjustment of the wall-climbing robot based on the offset and the yaw angle.

[0081] It should be noted that the differential speed adjustment prediction model is pre-trained. The input to this model is the offset and yaw angle of the wall-climbing robot, and the output is the differential speed adjustment. Based on the differential speed adjustment, the rotational speeds of the left and right wheels can be controlled to adjust the robot's travel path. In this step, the offset and yaw angle are input into the pre-trained differential speed adjustment prediction model, which then outputs the differential speed adjustment of the wall-climbing robot based on these parameters.

[0082] Optionally, in one possible implementation, the training process of the pre-trained differential regulation prediction model may include:

[0083] S1051. Construct an initial differential speed adjustment prediction model and initialize the parameters of the initial differential speed adjustment prediction model; wherein, the initial differential speed adjustment prediction model includes an action network, an evaluation network, a target action network corresponding to the action network, and a target evaluation network corresponding to the evaluation network.

[0084] Specifically, the action network, evaluation network, target action network corresponding to the action network, and target evaluation network corresponding to the evaluation network all employ neural networks with only one hidden layer. Furthermore, random parameters can be used to adjust the network parameters of the action network. Evaluate the network parameters The network parameters of the target action network corresponding to the action network. And the target evaluation network corresponding to the evaluation network. Initialize the network by assigning a random value to the network parameters of each network.

[0085] Specifically, the action network uses the state vector at time t. As input, output a deterministic differential adjustment action. To ensure that the differential speed adjustment action value output by the action network is within a reasonable physical range (e.g., within [-1, 1]), a hyperbolic tangent function is used as the activation function in the output layer of the action network. This maps the original output to a smooth and bounded interval, thus avoiding system instability caused by excessively large action amplitudes in the early stages of training the initial differential speed adjustment prediction model. Furthermore, the evaluation network uses the state vector of the wall-climbing robot at time t. and action vectors Input, Output The value is used to evaluate the worth of the current wall-climbing robot's actions.

[0086] Among them, the state vector of the wall-climbing robot for:

[0087]

[0088] in, This represents the state vector of the wall-climbing robot at the current moment. This represents the current offset of the wall-climbing robot. This represents the current offset angle of the wall-climbing robot.

[0089] Furthermore, the motion vectors of the wall-climbing robot (The action that should be performed in the current state, which is guaranteed by the predicted differential adjustment amount) is:

[0090] ;

[0091] in, This represents the motion vector of the wall-climbing robot at the current moment. The speed of the wall-climbing robot's revolving wheel; The speed of the right wheel of the wall-climbing robot; A fixed baseline speed for the wall-climbing robot; This represents the predicted differential adjustment amount.

[0092] It should be noted that the wall-climbing robot has a fixed base speed. It is a fixed value, and the specific value of this reference speed is set according to actual needs; this application does not limit it. Furthermore, the predicted differential adjustment amount... This is used to adjust the speed difference between the left and right wheels of a wall-climbing robot, thereby controlling the robot's steering. For example, if the differential speed adjustment... If the value is positive, the speed of the left wheel of the wall-climbing robot increases, the speed of the right wheel decreases, and the robot turns to the right; if the differential speed adjustment is positive... If the value is negative, the speed of the left wheel of the wall-climbing robot decreases, the speed of the right wheel increases, and the robot turns to the left; if... If the speed is zero, the speeds of the left and right wheels are equal, and the robot moves in a straight line in the current direction. It should be noted that the differential speed adjustment... The specific value is the action output by the action network.

[0093] Furthermore, the Q value can be expressed as:

[0094] ;

[0095] That is, Q-value is the action vector to be executed given the current state vector of the wall-climbing robot. The evaluation value obtained at that time The Q-function takes the state vector and action vector as input and outputs the evaluation value of the evaluation network. To evaluate the network parameters.

[0096] It should be noted that the Q-value can be used to measure the state vector of a wall-climbing robot. Next, execute the action vector. The quality of an action vector is determined by its Q-value. A larger Q-value indicates a better performance when the action vector is executed under the given state vector, while a smaller Q-value indicates a worse performance when the action vector is executed under the given state vector.

[0097] S1052. During the process of the wall-climbing robot moving along the weld seam, weld seam images are periodically collected according to a preset cycle, and for each collected weld seam image, the current state of the wall-climbing robot is determined based on the weld seam image; wherein, the current state is characterized by the offset and yaw angle of the wall-climbing robot.

[0098] In practice, as the wall-climbing robot tracks and walks along the weld seam, a camera mounted on the robot captures an image of the weld seam every 0.2 seconds. For each weld seam image, based on that image, referring to the previous method (i.e., inputting the weld seam image into a pre-trained rotating target detection model, where the rotating target detection model performs target detection on the weld seam image, outputs the coordinates of the corner points of the rotating bounding box of the weld seam, then determines the current pose of the wall-climbing robot relative to the weld seam based on the coordinates of the corner points of the rotating bounding box, and finally calculates the current offset and yaw angle of the wall-climbing robot based on the current pose of the wall-climbing robot relative to the weld seam and the predefined standard pose of the wall-climbing robot relative to the weld seam), the offset and yaw angle of the wall-climbing robot can be calculated, that is, the state of the wall-climbing robot under that weld seam image can be calculated.

[0099] S1053. The current state is input into the action network, and the action network outputs the action that the wall-climbing robot should perform in the current state; wherein the action is characterized by the predicted differential speed adjustment amount.

[0100] In this step, the current state of the wall-climbing robot is captured in each weld seam image. The input is fed into the action network, which outputs the original action vector that the wall-climbing robot should perform in its current state. The original motion is then mapped to the differential adjustment amount (in practice, the mapping is performed according to the following formula): ,in, (This is the maximum differential speed adjustment that the active wheel motor of the wall-climbing robot can achieve).

[0101] S1054. Based on the predicted differential speed adjustment, determine the speed of the left wheel and the speed of the right wheel of the wall-climbing robot, and control the wall-climbing robot's movement according to the speed of the left wheel and the speed of the right wheel.

[0102] In this step, based on the predicted differential speed adjustment, the speeds of the left and right wheels of the wall-climbing robot can be determined using the following formula:

[0103] ;

[0104] ;

[0105] in, The speed of the wall-climbing robot's revolving wheel; The speed of the right wheel of the wall-climbing robot; A fixed baseline speed for the wall-climbing robot; This refers to the differential adjustment amount.

[0106] Furthermore, after determining the speeds of the left and right wheels of the wall-climbing robot, the robot's movements are controlled according to these determined speeds.

[0107] S1055. Calculate the reward value of the action based on the offset, the yaw angle, and the differential adjustment.

[0108] Specifically, in one possible implementation, a reward value can be calculated based on the offset, the yaw angle, and the differential adjustment amount, using a preset reward function; wherein the preset reward function is:

[0109] ;

[0110] in, This is the reward value for the action; This is the offset at the current moment; This is the offset penalty coefficient; This represents the offset angle at the current moment; This is the offset angle penalty coefficient; This represents the differential adjustment amount at the current moment; This is the differential adjustment penalty coefficient.

[0111] Specifically, the reward function aims to maximize tracking accuracy and maintain motion smoothness, penalizing offset and angular deviation and suppressing sharp turning movements.

[0112] It should be noted that, , , All are positive numbers, used to control the penalty for different errors. , , The larger the value, the greater the penalty and the greater the effect of suppressing changes in the values ​​of the variables they control; , , The smaller the value, the weaker the penalty and the less effective the suppression of changes in the variables they control; furthermore, the reward values ​​calculated by this reward function are all negative, indicating that the actions of the wall-climbing robot need to be penalized.

[0113] The wall-climbing robot weld seam tracking method provided in this embodiment effectively suppresses the deviation of the wall-climbing robot relative to the weld seam path by setting a reward function and penalty coefficients for the robot's offset and offset angle. The larger the offset and yaw angle of the wall-climbing robot, the stronger the penalty will be applied according to the set penalty coefficient, forcing the robot to minimize these errors. Regarding the differential speed adjustment of the wall-climbing robot, the larger the differential speed adjustment, the more violent the robot's turning will be, which may lead to instability. By setting corresponding penalty coefficients, when the differential speed adjustment of the wall-climbing robot is large, the robot will be subject to stronger penalties, suppressing violent turning movements. In summary, by setting penalty coefficients for offset, offset angle, and differential speed adjustment, the wall-climbing robot can reduce deviations and maintain stability in real-time weld seam tracking, thereby achieving more accurate weld seam tracking.

[0114] S1056. After the wall-climbing robot performs the action, obtain the new state of the wall-climbing robot after performing the action, and store the current state, the action, the reward value and the new state as a set of experience values ​​in the experience replay buffer.

[0115] In this step, after each action the wall-climbing robot performs, the current state before the action, the action itself, the reward value for that action, and the new state obtained after completing the action are collectively used as a set of experience values. The experience values ​​obtained after the wall-climbing robot performs actions are stored in the experience playback buffer.

[0116] It should be noted that the new state after this action is performed is determined based on the weld image acquired after the action is performed.

[0117] S1057. Randomly collect multiple sets of data from the experience playback buffer to calculate the target Q value, update the parameters of the action network and the evaluation network based on the calculated target Q value, and pass the updated parameters to the target action network and the target evaluation network through soft update.

[0118] In this step, multiple sets of data are collected from the experience playback buffer to calculate the target Q value, which can be calculated using the following formula:

[0119] ;

[0120] in, The target Q value; This is the reward value for the action; Discount factor; Let be the state vector of the wall-climbing robot given at the next moment. The evaluation value that can be obtained by executing the action vector at the next moment is output by the target evaluation network based on the state vector at the next moment; Let be the state vector of the wall-climbing robot at the next moment; For the network parameters of the target action network; The network parameters of the target network are evaluated.

[0121] Specifically, the parameters of the evaluation network can be updated by minimizing the variance loss:

[0122] ;

[0123] Where L is the loss function; N is the batch size, representing the number of sets of empirical data randomly collected from the empirical replay buffer; and y_t is the target Q-value. To evaluate the Q-value of the network prediction; To evaluate the parameters of the network; The squared difference between the network's predicted Q-value and the target Q-value is used to measure the accuracy of the network's predicted Q-value.

[0124] The parameters of the action network can be updated by calculating the policy gradient:

[0125] ;

[0126] in, For the objective function About Action Networks The gradient; The objective function is... Batch size; To evaluate the Q-value of the network prediction with respect to actions The gradient; For the action network, based on the state vector at the current time step The output action vector; The action vector output by the action network with respect to the action network parameters The gradient; These are the parameters of the action network.

[0127] Furthermore, the updated parameters are passed to the target action network and the target evaluation network through soft updates, as shown below:

[0128] ;

[0129] ;

[0130] in, This is the soft update coefficient. ; These are the parameters for the action network; For the target action network parameters; To evaluate network parameters; To evaluate the network parameters for the target.

[0131] It should be noted that the soft update coefficient Used to control the update speed of the target network. The larger the value, the faster the target network updates. The smaller the value, the slower the target network updates; you can choose according to your actual needs. The specific size is not limited in this application.

[0132] S1058. After training, the target action network is used as a pre-trained differential speed adjustment prediction model.

[0133] In this step, after training, the target motion network is used as a pre-trained differential speed adjustment prediction model, which can output the differential speed adjustment of the wall-climbing robot based on the offset and yaw angle.

[0134] The wall-climbing robot weld seam tracking method provided in this application enables the wall-climbing robot to autonomously adjust its path during movement by constructing and optimizing a differential speed adjustment prediction model. Furthermore, the robot periodically acquires weld seam images to obtain its current state in real time, and predicts appropriate differential speed adjustments through an action network, thereby adjusting the speeds of the left and right wheels to control the movement. In addition, after the wall-climbing robot performs an action, its new state is acquired, and the current state, action, reward value, and new state are stored as a set of experience values ​​in an experience replay buffer. Multiple sets of data are randomly collected from the experience replay buffer to calculate the target Q-value. Based on the calculated target Q-value, the parameters of the action network and evaluation network are updated, and the updated parameters are passed to the target action network and target evaluation network through soft updates. This improves the stability and efficiency of the differential speed adjustment prediction model training, allowing the wall-climbing robot to gradually optimize its control capabilities as it continuously executes actions and updates its strategy. This improves the tracking accuracy, motion smoothness, and adaptability of the wall-climbing robot's weld seam path, forming an efficient and stable differential speed adjustment prediction model.

[0135] S106. Determine a set of speeds for the wall-climbing robot based on the differential speed adjustment, and control the wall-climbing robot's movements based on the set of speeds, so that the wall-climbing robot tracks the weld seam to be tracked.

[0136] It should be noted that the set of speeds includes the speeds of the left and right wheels of the wall-climbing robot.

[0137] Specifically, based on the differential speed adjustment, the speeds of the left and right wheels of the wall-climbing robot can be determined using the following formula:

[0138] ;

[0139] ;

[0140] in, The speed of the wall-climbing robot's revolving wheel; The speed of the right wheel of the wall-climbing robot; A fixed baseline speed for the wall-climbing robot; This refers to the differential adjustment amount.

[0141] Furthermore, to verify the effectiveness of the method provided in this embodiment, an experimental platform was built at the welding site of the hull closure weld of a bulk carrier in a shipyard dock for testing. The wall-climbing robot moved vertically upwards along the bottom of the midship hull closure weld. During the tracking of the 20m long weld, after an initial adjustment period of approximately 0.8m, the lateral offset between the actual pose and the standard pose of the wall-climbing robot remained within 2cm, and the heading offset angle remained within 5°, ensuring that the camera could acquire high-quality weld images and achieve high-precision weld tracking.

[0142] The wall-climbing robot weld seam tracking method provided in this embodiment involves acquiring a weld seam image using a camera mounted on the robot when the robot is at the starting position of the weld seam to be tracked. This image is then input into a pre-trained rotating target detection model. The model performs target detection on the weld seam image, outputting the coordinates of the corner points of the weld seam's rotation bounding box. Based on these coordinates, the current pose of the wall-climbing robot relative to the weld seam is determined. Finally, based on the current pose of the wall-climbing robot relative to the weld seam and a predefined standard pose of the robot relative to the weld seam, a calculation is performed. The current offset and yaw angle of the wall-climbing robot are calculated and input into a pre-trained differential speed adjustment prediction model. The differential speed adjustment prediction model outputs the differential speed adjustment of the wall-climbing robot based on the offset and yaw angle. Finally, a set of velocities for the wall-climbing robot is determined based on the differential speed adjustment, and the robot's movements are controlled according to this set of velocities to enable the wall-climbing robot to track the weld seam. This constructs an integrated weld seam tracking process of "image perception—pose estimation—intelligent control," achieving high-precision, real-time tracking of the weld seam path by the wall-climbing robot under complex working conditions. First, weld seam images are acquired by a camera, providing continuous and reliable image input for the rotating target detection model. The rotating target detection model accurately expresses the spatial pose of the weld seam, thereby accurately quantifying the robot's lateral offset and heading deviation relative to the weld seam. Based on this, the differential speed adjustment prediction model outputs wheel speed adjustment, enabling the robot to smoothly and accurately approach the weld seam trajectory in the control closed loop. The aforementioned image perception, pose estimation, and intelligent control form a tightly coupled technical chain, supporting each other and driving the whole system, ensuring that the system has good robustness, real-time performance, and environmental adaptability, thereby significantly improving the tracking accuracy of the wall-climbing robot in automated welding operations.

[0143] Corresponding to the aforementioned embodiment of a wall-climbing robot weld seam tracking method, this application also provides an embodiment of a wall-climbing robot weld seam tracking device.

[0144] An embodiment of the weld seam tracking device for a wall-climbing robot disclosed in this application can be applied to a wall-climbing robot. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the wall-climbing robot reading the corresponding computer program instructions from non-volatile memory into memory and executing them. From a hardware perspective, such as... Figure 6 The diagram shown is a hardware structure diagram of a wall-climbing robot, in which the weld seam tracking device of this application is located. (Except for...) Figure 6In addition to the processor, memory, network interface, and non-volatile memory shown, the wall-climbing robot in which the device is located in the embodiment may also include other hardware depending on the actual function of the wall-climbing robot's weld seam tracking device, which will not be described in detail here.

[0145] Figure 5 This is a schematic diagram of the structure of Embodiment 1 of the wall-climbing robot weld seam tracking device provided in this application. Please refer to... Figure 5 The device provided in this embodiment includes a data acquisition module 510, a detection module 520, a determination module 530, a calculation module 540, and a control module 550, wherein...

[0146] The acquisition module 510 is used to acquire an image of the weld seam to be tracked by a camera mounted on the wall-climbing robot when the wall-climbing robot is at the starting position of the weld seam to be tracked; wherein, the camera is mounted on a cantilever extending from the right side wall of the wall-climbing robot and is used to acquire an image of the weld seam located on the right side of the wall-climbing robot; when the wall-climbing robot is at the starting position of the weld seam to be tracked, the weld seam to be tracked is within the field of view of the camera;

[0147] The detection module 520 is used to input the weld image into a pre-trained rotating target detection model, and the rotating target detection model performs target detection on the weld image and outputs the coordinates of the corner points of the rotating bounding box of the weld.

[0148] The determining module 530 is used to determine the current pose of the wall-climbing robot relative to the weld seam based on the coordinates of the corner points of the rotating bounding box; wherein the current pose of the wall-climbing robot relative to the weld seam is characterized by the coordinates of the center point of the rotating bounding box and the rotation angle of the rotating bounding box.

[0149] The calculation module 540 is used to calculate the current offset and yaw angle of the wall-climbing robot based on the current pose of the wall-climbing robot relative to the weld and the predefined standard pose of the wall-climbing robot relative to the weld.

[0150] The calculation module 540 is also used to input the offset and the yaw angle into a pre-trained differential speed adjustment prediction model, and the differential speed adjustment prediction model outputs the differential speed adjustment of the wall-climbing robot based on the offset and the yaw angle.

[0151] The control module 550 is used to determine a set of speeds for the wall-climbing robot based on the differential speed adjustment, and to control the wall-climbing robot's movements based on the set of speeds so that the wall-climbing robot tracks the weld seam to be tracked; wherein, the set of speeds includes the speeds of the left wheel and the right wheel of the wall-climbing robot.

[0152] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0153] Please continue to refer to Figure 6 This application also provides a wall-climbing robot, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods provided in the first aspect of this application.

[0154] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods provided in this application.

[0155] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0156] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0157] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for tracking weld seams in a wall-climbing robot, characterized in that, The method includes: When the wall-climbing robot is at the starting position of the weld seam to be tracked, the camera mounted on the wall-climbing robot captures an image of the weld seam to be tracked; wherein, the camera is mounted on a cantilever extending from the right side wall of the wall-climbing robot, and is used to capture an image of the weld seam located on the right side of the wall-climbing robot; when the wall-climbing robot is at the starting position of the weld seam to be tracked, the weld seam to be tracked is within the field of view of the camera; The weld image is input into a pre-trained rotating target detection model, which performs target detection on the weld image and outputs the coordinates of the corner points of the rotating bounding box of the weld. The current pose of the wall-climbing robot relative to the weld is determined based on the coordinates of the corner points of the rotating bounding box; wherein the current pose of the wall-climbing robot relative to the weld is characterized by the coordinates of the center point of the rotating bounding box and the rotation angle of the rotating bounding box. Based on the current pose of the wall-climbing robot relative to the weld and the predefined standard pose of the wall-climbing robot relative to the weld, calculate the current offset and yaw angle of the wall-climbing robot. The offset and the yaw angle are input into a pre-trained differential speed adjustment prediction model, and the differential speed adjustment prediction model outputs the differential speed adjustment of the wall-climbing robot based on the offset and the yaw angle. A set of speeds for the wall-climbing robot is determined based on the differential speed adjustment, and the wall-climbing robot's movements are controlled according to the set of speeds to enable the wall-climbing robot to track the weld seam to be tracked; wherein, the set of speeds includes the speeds of the left wheel and the right wheel of the wall-climbing robot.

2. The method according to claim 1, characterized in that, The training process of the pre-trained rotating target detection model includes: An initial rotating target detection model is constructed based on the YOLOv11-OBB model. This model includes a backbone network, a neck network, and a detection head. The backbone network's SPPF and C2PSA modules are sequentially connected by a channel attention module and a spatial attention module. The convolutional module of the bounding box regression branch in the detection head is a DCNv4 deformable convolutional module. The channel attention module performs global average pooling and global max pooling on the input feature map to obtain a first channel description vector and a second channel description vector. Based on these vectors, attention weights are determined, and the feature map is then channel-weighted according to these weights to obtain a first processed feature map. The spatial attention module performs max pooling and average pooling on the first processed feature map along the channel dimension to obtain a first spatial feature map and a second spatial feature map. Based on these vectors, spatial attention weights are determined, and the first processed feature map is then spatially weighted according to these weights to output a second processed feature map. The initial rotating target detection model is trained to obtain the pre-trained rotating target detection model; wherein, in the first preset number of epochs of training the initial rotating target detection model, the parameters of the backbone network are frozen, and only the parameters of the neck network and the detection head are trained.

3. The method according to claim 1, characterized in that, The training process of the pre-trained differential regulation prediction model includes: Construct an initial differential speed adjustment prediction model and initialize the parameters of the initial differential speed adjustment prediction model; wherein, the initial differential speed adjustment prediction model includes an action network, an evaluation network, a target action network corresponding to the action network, and a target evaluation network corresponding to the evaluation network; As the wall-climbing robot moves along the weld seam, images of the weld seam are periodically acquired according to a preset cycle, and the current state of the wall-climbing robot is determined for each acquired weld seam image; wherein, the current state is characterized by the offset and yaw angle of the wall-climbing robot. The current state is input into the action network, which outputs the action that the wall-climbing robot should perform in the current state; wherein, the action is characterized by the predicted differential speed adjustment amount; Based on the predicted differential speed adjustment, the speeds of the left and right wheels of the wall-climbing robot are determined, and the robot's movements are controlled according to the speeds of the left and right wheels. The reward value for the action is calculated based on the offset, the yaw angle, and the differential adjustment. After the wall-climbing robot performs the action, the new state of the wall-climbing robot after performing the action is obtained, and the current state, the action, the reward value and the new state are stored as a set of experience values ​​in the experience replay buffer; Multiple sets of data are randomly collected from the experience replay buffer to calculate the target Q value. The parameters of the action network and the evaluation network are updated based on the calculated target Q value. The updated parameters are then passed to the target action network and the target evaluation network through soft updates. After training, the target action network is used as a pre-trained differential speed regulation prediction model.

4. The method according to claim 3, characterized in that, The calculation of the reward value for the action based on the offset, the yaw angle, and the differential adjustment includes: A reward value is calculated based on the offset, the yaw angle, and the differential adjustment amount, using a preset reward function; wherein the preset reward function is: ； in, This is the reward value for the action; This is the offset at the current moment; This is the offset penalty coefficient; This represents the offset angle at the current moment; This is the offset angle penalty coefficient; This represents the differential adjustment amount at the current moment; This is the differential adjustment penalty coefficient.

5. The method according to claim 1, characterized in that, The predefined standard pose of the wall-climbing robot relative to the weld seam is the pose when the weld seam is in a horizontal position and the center of the weld seam coincides with the center of the image; the calculation of the current offset and yaw angle of the wall-climbing robot based on the current pose of the wall-climbing robot relative to the weld seam and the predefined standard pose of the wall-climbing robot relative to the weld seam includes: The rotation angle is determined as the current yaw angle of the wall-climbing robot; Based on the current pose of the wall-climbing robot relative to the weld seam and the coordinates of the weld seam center in the predefined standard pose of the wall-climbing robot relative to the weld seam, the current offset of the wall-climbing robot along the weld seam direction and the current offset of the wall-climbing robot along the vertical direction are determined; wherein, the vertical direction is the direction perpendicular to the weld seam direction. The current offset of the wall-climbing robot along the vertical direction is determined as the current offset of the wall-climbing robot.

6. The method according to claim 5, characterized in that, Based on the current pose of the wall-climbing robot relative to the weld seam, and the coordinates of the weld seam center in a predefined standard pose of the wall-climbing robot relative to the weld seam, determine the current offset of the wall-climbing robot along the weld seam direction and the current offset of the wall-climbing robot along the vertical direction, including: The current offset of the wall-climbing robot along the weld seam direction and the current offset of the wall-climbing robot along the vertical direction are calculated using the following homogeneous equations: ； in, This represents the current offset of the wall-climbing robot along the weld seam direction. This represents the current vertical offset of the wall-climbing robot. This represents the current yaw angle of the wall-climbing robot. , This represents the current pose of the wall-climbing robot relative to the weld seam. , The coordinates of the weld center relative to the standard pose of the weld seam for the predefined wall-climbing robot.

7. The method according to claim 1, characterized in that, During the movement of the wall-climbing robot, the weld seam to be tracked is in a dark box environment that completely shields it from external light, and the camera acquires images of the weld seam based on its own light source.

8. The method according to claim 7, characterized in that, The method further includes: After each acquisition of a weld image, the quality of the weld image is evaluated to obtain a quality score; The angle of the light source is adjusted based on the quality score and the pose of the wall-climbing robot relative to the weld.

9. A weld seam tracking device for a wall-climbing robot, characterized in that, The device includes a data acquisition module, a detection module, a determination module, a calculation module, and a control module, wherein... The acquisition module is used to acquire images of the weld seam to be tracked via a camera mounted on the wall-climbing robot when the wall-climbing robot is at the starting position of the weld seam to be tracked; wherein, the camera is mounted on a cantilever extending from the right side wall of the wall-climbing robot and is used to acquire images of the weld seam located on the right side of the wall-climbing robot; when the wall-climbing robot is at the starting position of the weld seam to be tracked, the weld seam to be tracked is within the field of view of the camera; The detection module is used to input the weld image into a pre-trained rotating target detection model, and the rotating target detection model performs target detection on the weld image and outputs the coordinates of the corner points of the rotating bounding box of the weld. The determining module is used to determine the current pose of the wall-climbing robot relative to the weld seam based on the coordinates of the corner points of the rotating bounding box; wherein the current pose of the wall-climbing robot relative to the weld seam is characterized by the coordinates of the center point of the rotating bounding box and the rotation angle of the rotating bounding box. The calculation module is used to calculate the current offset and yaw angle of the wall-climbing robot based on the current pose of the wall-climbing robot relative to the weld and the predefined standard pose of the wall-climbing robot relative to the weld. The calculation module is also used to input the offset and the yaw angle into a pre-trained differential speed adjustment prediction model, and the differential speed adjustment prediction model outputs the differential speed adjustment of the wall-climbing robot based on the offset and the yaw angle. The control module is used to determine a set of speeds for the wall-climbing robot based on the differential speed adjustment, and to control the wall-climbing robot's movements based on the set of speeds so that the wall-climbing robot tracks the weld seam to be tracked; wherein, the set of speeds includes the speeds of the left wheel and the right wheel of the wall-climbing robot.

10. A wall-climbing robot, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the steps of the method according to any one of claims 1-8.

Images