Pipeline repairing method based on defect detection and three-dimensional reconstruction

Abstract

The invention relates to the technical field of pipeline internal defect detection and repair, and particularly discloses a pipeline repair method based on defect detection and three-dimensional reconstruction, which comprises the following steps of: 1, scanning the inner wall of a pipeline through a pipeline internal defect detection module, and confirming the position and the type of a pipeline defect; 2, judging whether a pipeline defect is a repairable type or not through a pipeline internal defect detection module; 3, determining three-dimensional information coordinates of pipeline defects relative to the pipeline through a pipeline internal defect detection module; 4, if the pipeline defects can be repaired, image information of the pipeline defects is collected through a pipeline internal defect repairing module, three-dimensional reconstruction and trajectory planning are carried out, and repairing is completed. According to the method, the pipeline defects are recognized and detected through the methods of defect detection, three-dimensional reconstruction and trajectory planning, intelligent recognition detection and automatic welding repair of the pipeline defects are achieved, and the recognition efficiency and accuracy are high.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline internal defect detection and repair, and in particular to a pipeline repair method based on defect detection and three-dimensional reconstruction. Background Art

[0002] The existing large-diameter pipeline internal defects are still repaired manually. The defects of large-diameter pipelines are mainly pipeline cracks. Since large-diameter pipelines are buried deep underground, it is necessary to dig up the land above the pipeline cracks, drain the pipeline, cut the crack position of the pipeline, and then weld a new pipeline. This repair process is time-consuming, consumes huge manpower and material resources, and is inefficient, so a new pipeline repair method is needed.

[0003] With the rapid development of artificial intelligence technology, deep neural networks (deep learning) have developed rapidly in the direction of image processing, and the functions achieved have gradually expanded from classification and positioning tasks to positioning and segmentation tasks, and are particularly prominent in unmanned driving, medical image detection, etc. Compared with traditional image detection methods, deep learning has stronger robustness and generalization capabilities. With the improvement of computer data computing power today, it has gradually become the mainstream defect detection method today. Among them, the Chinese invention patent, application number 202310900926.5, discloses a pipeline defect detection and evaluation method and device based on deep learning. The technical solution of the application includes preprocessing the internal image of the pipeline to be processed, inputting the trained neural network model, and obtaining the damage type of the corresponding position inside the pipeline; the neural network is AlexNet. The application overcomes the defects of the prior art that the detection cost is high and the automation in the field of pipeline detection and evaluation cannot be achieved, but its detection efficiency is relatively slow.

[0004] Currently, the commonly used 3D reconstruction technologies are mainly divided into active vision methods and passive vision methods. Active vision methods include laser scanning method, structured light method, shadow method, TOF technology, radar technology, Kinect technology, etc. Passive vision methods include monocular vision method, binocular vision method and multi-eye vision method. Among them, the Chinese invention patent, application number 201910486706.6, discloses a 3D reconstruction method and system based on binocular vision. The technical solution of the application includes: (1) image acquisition; (2) feature extraction; (3) feature matching; (4) 3D coordinate calculation, which can solve the problem of low matching accuracy of matching features in the 3D model reconstruction method based on binocular images in the prior art, but it cannot be applied to underwater pipeline repair.

[0005] Compared with traditional path planning methods, deep reinforcement learning is a non-model-based method, so the motion paths planned using deep reinforcement learning in various environments have better robustness. Among them, the Chinese invention patent, application number 201910612564.3 discloses a humanoid robot control system based on deep learning and its control method. The technical solution of this application is aimed at the characteristics of the humanoid robot itself, as well as the operational requirements of intelligent control, intelligent recognition, and intelligent operation, but it cannot be combined with three-dimensional reconstruction and cannot be used for repairing internal defects of pipelines. Therefore, in order to improve the deficiencies of the above-mentioned prior art, it is necessary to propose an improvement to overcome the above-mentioned deficiencies. Summary of the invention

[0006] The purpose of the present invention is to solve the problems in the prior art and provide a pipeline repair method based on defect detection and three-dimensional reconstruction.

[0007] The technical solution of the present invention is: a pipeline repair method based on defect detection and three-dimensional reconstruction, comprising the following steps: Step 1: Scan the inner wall of the pipeline through the pipeline internal defect detection module to confirm the location and type of pipeline defects; Step 2: Determine whether the pipeline defect is a repairable type through the pipeline internal defect detection module; Step 3: Determine the three-dimensional information coordinates of the pipeline defect relative to the pipeline through the pipeline internal defect detection module; Step 4: If the pipeline defect is repairable, the image information of the pipeline defect is collected through the pipeline internal defect repair module, and three-dimensional reconstruction and trajectory planning are performed to complete the repair.

[0008] In order to achieve a better repair working environment, the pipeline internal defect detection module includes a drainage hood, a vertical thruster, a horizontal thruster and an external camera. The drainage hood can be adsorbed on the pipeline to drain moisture. The vertical thruster and the horizontal thruster can transport the drainage hood to the pipeline defect in the vertical and horizontal directions. The external camera is a binocular camera used to scan the inner wall of the pipeline.

[0009] To facilitate repair, the pipeline internal defect repair module includes an internal camera and a pipeline repair robot. The internal camera is used to collect image information of pipeline defects. The pipeline repair robot includes a robotic arm and a welding gun. The robotic arm is installed inside the drainage hood, and the welding gun is installed at the lower end of the robotic arm and is used to repair pipeline defects.

[0010] Step 1 includes: Step 1: Generate spatial point cloud inside the pipeline: Use a binocular camera to collect images of the inside of the pipeline, and perform overall three-dimensional reconstruction of the inside of the pipeline to obtain spatial point cloud information; Step 2: Pipeline defect location detection: The pipeline image collected in step 1 is used to detect defects using a target detection algorithm based on a combination of prior knowledge and deep learning algorithms, and the detection results and positioning information are output.

[0011] Step 2 includes: Step 3: Pipeline defect point cloud generation: Use the YOLACT instance segmentation algorithm to perform instance segmentation on the defect location and obtain the image point cloud information of the pipeline defect location.

[0012] Step three includes: Step 4: Mapping the pipeline defect image point cloud with the pipeline internal space point cloud: Map the image point cloud information of the pipeline defect with the pipeline internal space point cloud to obtain the three-dimensional information of the defect location.

[0013] Step 4 includes: Step 5: The pipeline repair robot arrives at the repair starting position: the pipeline repair robot is controlled to arrive at the repair starting position based on the reinforcement learning algorithm; Step 6: 3-5-3 trajectory planning of the repair path: Use 3-5-3 mixed polynomial interpolation to plan the trajectory of the repair path.

[0014] Step 2 The specific steps for pipeline defect location detection are as follows: K1: The RGB three-channel data of the original input of the deep learning network is modified into RGB-F (feature) four-channel data, and the image feature value with the largest distinction in c is added to the fourth channel. The largest image feature value is used as the fourth channel in RGB-F and the original color three-channel image as the input of the network structure; K2: Output the corresponding feature maps of different sizes through 5 convolutional layers, generate different detection frames for each key point of each feature map, and obtain the detection frame; K3: Determine the category of each detection frame based on the confidence level, retain a small number of detection frames, perform non-maximum suppression algorithm on the retained detection frames, and filter out those prediction frames with large overlap to obtain the detection results; K4: Determine the type of output pipeline defect and locate the defect.

[0015] Step 5 The specific steps for the pipeline repair robot to reach the repair starting position are as follows: E1: Based on step 4, the three-dimensional coordinate information of the part to be welded is obtained, and the coordinate information of the part to be repaired is obtained: x0, y0, z0, x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1, xn, yn, zn, where (x0, y0, z0) is the starting position of welding, (x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1) is the position where welding needs to pass, (xn, yn, zn) is the end position of welding, and the coordinate position of the welding gun (xrobot ,y robot , z robot ); E2: Design a reward function for the relative distance between the welding gun at the end of the pipeline repair robot and the welding starting position. By training the network in the constructed environment, the model converges quickly and the robot arm can be controlled to approach the welding starting position. Reward function reward: (1) (2) Where d is the distance from the welding gun to the welding starting position; 5cm and 1cm are artificially set thresholds. When d<1cm, it is considered that the welding gun has successfully reached the target position; E3: Use the Actor-Critic network framework, initialize the DDPG algorithm, and set the state space St = {x robot ,y robot , z robot , x0, y0, z0}, including the welding gun position (x robot ,y robot , z robot ) and welding starting position (x0, y0, z0); E4: Initialize the experience replay pool; E5: Calculate state S t+1 The distance d between the lower welding gun and the welding starting position is used to calculate the reward value. If the welding exceeds the working boundary, the reward value is -2; if d>5cm in the current round, the reward value is 0; if d is between 1cm and 5cm, the reward value is 0.1; if d<1cm, it is considered that the end of the pipeline repair robot has successfully reached the target position, and the reward value is 1; E6: Change state S t+1 The reward value and the actual action of the pipeline repair robot are input into the Critic network, and the Actor and Critic network parameters are updated accordingly; E7: Repeat E4 until the target position is reached successfully, and the training ends; E8: Obtain the trajectory to the repair starting point through the trained reinforcement learning algorithm.

[0016] The present invention adopts the above structure and has the following advantages: 1. Through defect detection, 3D reconstruction and trajectory planning methods, pipeline defects are identified and detected, realizing intelligent identification and detection of pipeline defects and automated welding repair, with high identification efficiency and accuracy; 2. The pipeline internal defect detection module is used to scan the pipeline, confirm the location and type of the pipeline defect, and determine the three-dimensional information coordinates of the pipeline defect relative to the pipeline; the pipeline internal defect repair module is used to collect image information of the pipeline defect, perform three-dimensional reconstruction and trajectory planning, and complete the repair. The two modules work together, complement each other, and are fast and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is the overall flow chart of the present invention; Figure 2 This is a schematic diagram of the detection status of the pipeline internal defect detection module; Figure 3 This is a schematic diagram of the repair status of the pipeline internal defect repair module; Figure 4 It is a detailed flowchart of the present invention; Figure 5 Schematic diagram of target detection algorithm based on the combination of prior knowledge and deep learning; Figure 6 This is a schematic diagram of the registration between the defect image point cloud and the pipeline internal space point cloud image; Figure 7 Detailed flowchart for trajectory planning.

[0018] In the figure, 1. pipeline; 2. drainage cover; 3. vertical thruster; 4. horizontal thruster; 5. internal camera; 6. external camera; 7. robotic arm; 8. welding gun; 9. pipeline defect. DETAILED DESCRIPTION

[0019] In order to make the technical means, technical features, invention objectives and technical effects achieved by the present invention easy to understand, the present invention is further explained below with reference to specific illustrations.

[0020] Embodiment 1: like Figure 1 and Figure 4 As shown, a pipeline repair method based on defect detection and three-dimensional reconstruction includes the following steps: Step 1: guide the pipeline repair robot to move forward in the pipeline, scan the inner wall of the pipeline through the pipeline internal defect detection module, and confirm the location and type of the pipeline defect 9; Step 1 includes: Step 1: Generate spatial point cloud inside the pipeline: Use a binocular camera to collect images of the inside of the pipeline, and perform overall three-dimensional reconstruction of the inside of the pipeline to obtain spatial point cloud information; Step 2: Pipeline defect 9 position detection: The pipeline image collected in step 1 is used to detect defects using a target detection algorithm based on a combination of prior knowledge and deep learning algorithms, and the detection results and positioning information are output; Step 2 The specific steps for detecting the position of pipeline defect 9 are as follows: like Figure 5 As shown, K1: The RGB three-channel data of the original input of the deep learning network is modified into RGB-F (feature) four-channel data, the image feature value with the largest distinction in c is added to the fourth channel, and the largest image feature value is used as the fourth channel in RGB-F and the original color three-channel image as the input of the network structure; K2: Output the corresponding feature maps of different sizes through 5 convolutional layers, generate different detection frames for each key point of each feature map, and obtain the detection frame; K3: Determine the category of each detection frame based on the confidence level, retain a small number of detection frames, perform non-maximum suppression algorithm on the retained detection frames, and filter out those prediction frames with large overlap to obtain the detection results; K4: Determine 9 types of output pipeline defects and locate the defect locations; The internal image of the pipeline obtained in step 2 is analyzed, and a total of 9 grayscale co-occurrence matrix feature indicators are extracted, and their expressions are: Mean (M): (3) Standard Deviation (S): (4) Contrast (C): (5) Dissimilarity (D): (6) Homogeneity (H): (7) Angular second moment (A): (8) Energy (Ene): (9) Maximum probability (Ma), which indicates the problematic feature that appears most often in the image; Entropy (E): (10) Calculate the eigenvalues ​​of 9 gray-level co-occurrence matrix methods respectively, take the eigenvalue mean to represent the distinctiveness of the features and perform data normalization; Step 2: Determine whether the pipeline defect 9 is a repairable type through the pipeline internal defect detection module; Step 2 includes: Step 3: Pipeline defect 9 point cloud generation: Use the YOLACT instance segmentation algorithm to perform instance segmentation on the defect location and obtain the image point cloud information of the pipeline defect 9 location, such as Figure 6As shown, image registration is completed by constructing the transformation from the camera coordinate system of the depth camera to the image coordinate system, the transformation from the camera coordinate system of the color camera to the image coordinate system, the transformation from the left camera coordinate system to the right camera coordinate system, and the transformation from the left image coordinates to the right image coordinates, and the corresponding relationship between the image coordinates of the defect area and each three-dimensional point in the point cloud is obtained; According to the three-dimensional coordinate information (X, Y, Z), color information (R, G, B), and corresponding color image coordinates (u, v) contained in each three-dimensional point in the color point cloud, the color information corresponding to each three-dimensional point is judged to determine whether the three-dimensional point belongs to the defect area; the YOLACT algorithm is used to obtain the defect position and category in the color image, and the color of the defect area is set to (0, 0, 255), and the color of the other area is set to (255, 255, 255), to obtain the three-dimensional point cloud, and complete the point cloud segmentation of the pipeline defect 9 information; Step 3: Determine the three-dimensional information coordinates of the pipeline defect 9 relative to the pipeline through the pipeline internal defect detection module; Step three includes: Step 4: Mapping the image point cloud of the pipeline defect 9 with the internal space point cloud of the pipeline: Mapping the image point cloud information of the pipeline defect 9 with the internal space point cloud of the pipeline to obtain the three-dimensional information of the defect location; In order to obtain the three-dimensional coordinate information of the defective part, the image registration method is used to map the defect position graphic point cloud information of step 3 with the spatial point cloud of the pipeline in step 1, and the segmentation of the defect point cloud information is realized. The spatial point cloud coordinates of the defective part (x0, y0, z0, x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1, xn, yn, zn) are obtained, including ((x0, y0, z0) represents the starting position of the defect, (x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1) represents the middle position of the defect, and (xn,, yn, zn) is the end position of the defect; Step 4: If the pipeline defect 9 is repairable, the image information of the pipeline defect 9 is collected through the pipeline internal defect repair module, and three-dimensional reconstruction and trajectory planning are performed to complete the repair; Step 4 includes: Step 5: The pipeline repair robot arrives at the repair starting position: the pipeline repair robot is controlled to arrive at the repair starting position based on the reinforcement learning algorithm; like Figure 7 As shown, the specific steps for the pipeline repair robot to reach the repair starting position in step 5 are as follows: E1: Based on step 4, the three-dimensional coordinate information of the part to be welded is obtained, and the coordinate information of the part to be repaired is obtained: x0, y0, z0, x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1, xn, yn, zn, where (x0, y0, z0) is the starting position of welding, (x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1) is the position where welding needs to pass, (xn, yn, zn) is the end position of welding, and the coordinate position of the welding gun (x robot ,y robot , z robot ); E2: Design a reward function for the relative distance between the welding gun at the end of the pipeline repair robot and the welding starting position. By training the network in the constructed environment, the model converges quickly and the robot arm can be controlled to approach the welding starting position. Reward function reward: (1) (2) Where d is the distance from the welding gun to the welding starting position; 5cm and 1cm are artificially set thresholds. When d<1cm, it is considered that the welding gun has successfully reached the target position; E3: Adopts the Actor-Critic network framework. The Actor network is a neural network that approximates the policy function, which is divided into the Actor current network and the Actor target network. The Critic network is a neural network that approximates the value function, which is divided into the Critic current network and the Critic target network. The Actor network is the action network. The Critic network outputs the action score Q based on the input state and the current action. The Actor network is responsible for outputting the action based on the state S and the evaluation score Q. Initialize the DDPG algorithm and set the state space St = {x robot ,y robot , z robot , x0, y0, z0}, including the welding gun position (x robot ,y robot , z robot ) and welding starting position (x0, y0, z0); E4: Initialize the experience replay pool, which is responsible for storing data ; The Actor-Critic network is used to update network parameters through batch sampling, where S t represents the state of the pipeline repair robot at time t, a t Indicates that the Actor network is based on the current state S t The action taken, rt represents the reward obtained after taking the action, S t+1 Indicates that the pipeline repair robot takes action a t The status of the pipeline repair robot, done indicates whether the round task is completed; according to the current status S t ,The pipeline repair robot action group a is generated by the DDPG action strategy Actor network t , and execute to obtain the pipeline repair robot state S at the next moment t+1 ;; E5: Calculate state S t+1 The distance d between the lower welding gun and the welding starting position is used to calculate the reward value. If the welding exceeds the working boundary, the reward value is -2; if d>5cm in the current round, the reward value is 0; if d is between 1cm and 5cm, the reward value is 0.1; if d<1cm, it is considered that the end of the pipeline repair robot has successfully reached the target position, and the reward value is 1; E6: Change state S t+1 The reward value and the actual action of the pipeline repair robot are input into the Critic network to obtain the Q estimate; Stored in the experience replay pool, and used to update the Actor and Critic network parameters; E7: Repeat E4 until the target position is reached successfully, and the training ends; E8: Obtain the trajectory to the starting point of the repair through the trained reinforcement learning algorithm Step 6: 3-5-3 trajectory planning of the repair path: After the robotic arm reaches the starting position for repair, in order to ensure the smooth operation of the pipeline repair robot, the 3-5-3 mixed polynomial is used to interpolate (x0, y0, z0, x1, y1, z1, x2, y2, z2, …xn-1, yn-1, zn-1, xn, yn, zn) to control the pipeline repair robot to complete the repair work at the defective position.

[0021] Among them, Figure 2 As shown, the internal defect detection module of the pipeline 1 includes a drainage cover 2, a vertical propeller 3, a horizontal propeller 4 and an external camera 6. The drainage cover 2 can be adsorbed on the pipeline 1 to drain water. The vertical propeller 3 and the horizontal propeller 4 can transport the drainage cover 2 to the pipeline defect 9 in the vertical direction and the horizontal direction. The external camera 6 is a binocular camera for scanning the inner wall of the pipeline 1. like Figure 3 As shown, the internal defect repair module of the pipeline 1 includes an internal camera 5 and a pipeline repair robot. The internal camera 5 is used to collect image information of the pipeline defect 9. The pipeline 1 repair robot includes a robotic arm 7 and a welding gun 8. The robotic arm 7 is installed inside the drainage hood 2. The welding gun 8 is installed at the lower end of the robotic arm 7 and is used to repair the pipeline defect 9.

[0022] The present invention collects images of the interior of the pipeline 1, performs three-dimensional reconstruction of the interior of the pipeline 1 as a whole, and obtains spatial point cloud information of the interior of the pipeline 1 as a whole; inputs the collected images into a trained target detection for defect detection, and outputs the detection result and positioning information; guides the pipeline 1 repair robot to be fixed near the pipeline defect 9; performs instance segmentation on the pipeline defect 9 to obtain image point cloud information of the pipeline defect 9; performs image registration on the image point cloud information of the pipeline defect 9 and the spatial point cloud information of the interior of the pipeline 1 to obtain three-dimensional information of the pipeline defect 9; guides the pipeline 1 repair robot to be adsorbed near the pipeline defect 9; performs repair trajectory planning based on reinforcement learning, and then repairs the pipeline defect 9 by the pipeline 1 repair robot.

[0023] It should be noted that the above-mentioned robot arm 7, welding gun 8, thruster and camera are all applications of existing technologies.

[0024] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of implementation of the present invention. That is, all equivalent changes and modifications made according to the content of the patent application scope of the present invention should belong to the technical scope of the present invention.

Claims

1. A pipeline repair method based on defect detection and three-dimensional reconstruction, characterized in that: The following steps are involved: Step 1: Scan the inner wall of the pipeline using the pipeline internal defect detection module to confirm the location and type of the pipeline defect (9); Step 2: Determine whether the pipeline defect (9) is a repairable type through the pipeline internal defect detection module; Step 3: Determine the three-dimensional information coordinates of the pipeline defect (9) relative to the pipeline through the pipeline internal defect detection module; Step 4: If the pipeline defect (9) is repairable, the image information of the pipeline defect (9) is collected through the pipeline internal defect repair module, and three-dimensional reconstruction and trajectory planning are performed to complete the repair.

2. A pipeline repair method based on defect detection and three-dimensional reconstruction according to claim 1, characterized in that: The pipeline (1) internal defect detection module comprises a drainage hood (2), a vertical propeller (3), a horizontal propeller (4) and an external camera (6); the drainage hood (2) can be adsorbed on the pipeline (1) to drain water; the vertical propeller (3) and the horizontal propeller (4) can transport the drainage hood (2) to the defect of the pipeline (1) in the vertical direction and the horizontal direction; and the external camera (6) is a binocular camera for scanning the inner wall of the pipeline (1).

3. A pipeline (1) repair method based on defect detection and three-dimensional reconstruction according to claim 2, characterized in that: The pipeline (1) internal defect repair module comprises an internal camera (5) and a pipeline repair robot, wherein the internal camera (5) is used to collect image information of the pipeline (1) defect, and the pipeline repair robot comprises a mechanical arm (7) and a welding gun (8), wherein the mechanical arm (7) is installed inside the drainage cover (2), and the welding gun (8) is installed at the lower end of the mechanical arm (7) and is used to repair the pipeline defect (9).

4. A pipeline repair method based on defect detection and three-dimensional reconstruction according to claim 3, characterized in that: The step one comprises: step 1: generating a spatial point cloud inside the pipeline: using a binocular camera to collect images of the inside of the pipeline, and reconstructing the inside of the pipeline in three dimensions as a whole to obtain spatial point cloud information; Step 2: Pipeline defect (9) location detection: The pipeline image collected in step 1 is used to detect defects using a target detection algorithm based on a combination of prior knowledge and deep learning algorithms, and the detection results and positioning information are output.

5. A pipeline repair method based on defect detection and three-dimensional reconstruction according to claim 4, characterized in that: The step 2 includes: Step 3: Pipeline defect (9) point cloud generation: using the YOLACT instance segmentation algorithm to perform instance segmentation on the defect position to obtain image point cloud information of the pipeline defect (9) position.

6. A pipeline repair method based on defect detection and three-dimensional reconstruction according to claim 5, characterized in that: The step three comprises: Step 4: Mapping the image point cloud of the pipeline defect (9) with the spatial point cloud inside the pipeline: mapping the image point cloud information of the pipeline defect (9) with the spatial point cloud inside the pipeline to obtain three-dimensional information of the defect location.

7. A pipeline repair method based on defect detection and three-dimensional reconstruction according to claim 6, characterized in that: The step 4 includes: Step 5: The pipeline repair robot arrives at the repair starting position: the pipeline repair robot is controlled to arrive at the repair starting position based on the reinforcement learning algorithm; Step 6: 3-5-3 trajectory planning of the repair path: Use 3-5-3 mixed polynomial interpolation to plan the trajectory of the repair path.

8. The pipeline repair method based on defect detection and three-dimensional reconstruction according to claim 4 is characterized in that: The specific steps of step 2: pipeline defect (9) position detection are as follows: K1: Modify the RGB three-channel data of the original input of the deep learning network into RGB-F four-channel data, add the image feature value with the largest distinction in 3 to the fourth channel, and use the largest image feature value as the fourth channel in RGB-F and the original color three-channel image as the input of the network structure; K2: Output the corresponding feature maps of different sizes through 5 convolutional layers, generate different detection frames for each key point of each feature map, and obtain the detection frame; K3: Determine the category of each detection frame based on the confidence level, retain a small number of detection frames, perform non-maximum suppression algorithm on the retained detection frames, and filter out those prediction frames with large overlap to obtain the detection results; K4: Determine the type of output pipeline defect (9) and locate the defect.

9. The pipeline repair method based on defect detection and three-dimensional reconstruction according to claim 5 is characterized in that: The specific steps for the pipeline repair robot to reach the repair starting position in step 5 are as follows: E1: Based on step 4, the three-dimensional coordinate information of the part to be welded is obtained, and the coordinate information of the part to be repaired is obtained: x0, y0, z0, x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1, xn, yn, zn, where (x0, y0, z0) is the starting position of welding, (x1, y1, z1, x2, y2, z2, ... xn-1, yn-1, zn-1) is the position where welding needs to pass, (xn, yn, zn) is the end position of welding, and the coordinate position of the welding gun (x robot ,y robot , z robot ); E2: Design a reward function for the relative distance between the welding gun at the end of the pipeline repair robot and the welding starting position. By training the network in the constructed environment, the model converges quickly and the robot arm can be controlled to approach the welding starting position. The reward function reward: （1） （2） Where d is the distance from the welding gun to the welding starting position; 5cm and 1cm are artificially set thresholds. When d<1cm, it is considered that the welding gun has successfully reached the target position; E3: Use the Actor-Critic network framework, initialize the DDPG algorithm, and set the state space S t ={x robot ,y robot , z robot , x0, y0, z0}, including the welding gun position (x robot ,y robot , z robot ) and welding starting position (x0, y0, z0); E4: Initialize the experience replay pool; E5: Calculate state S t+1 The distance d between the lower welding gun and the welding starting position is used to calculate the reward value. If the welding exceeds the working boundary, the reward value is -2; if d>5cm in the current round, the reward value is 0; if d is between 1cm and 5cm, the reward value is 0.1; if d<1cm, it is considered that the end of the pipeline repair robot has successfully reached the target position, and the reward value is 1; E6: Change state S t+1 The reward value and the actual action of the pipeline repair robot are input into the Critic network, and the Actor and Critic network parameters are updated accordingly; E7: Repeat E4 until the target position is reached successfully, and the training ends; E8: Obtain the trajectory to the repair starting point through the trained reinforcement learning algorithm.

Images