Intelligent identification and targeted repair control method for pipeline defects based on deep learning

Abstract

The present application belongs to the technical field of municipal pipeline repair, and provides a pipeline defect intelligent identification and targeted repair control method based on deep learning, which comprises: collecting pipeline internal images, inputting the images into a deep learning model, identifying the category, position and contour of defects, and generating a target area to be repaired with defect types labeled; according to the spatial distribution and defect types of the target area to be repaired, a clustering algorithm is used to divide the target area to be repaired into a plurality of spraying task units; for each spraying task unit, the defect types and the motion state parameters of the spraying robot are combined; for point defects, the length of the deceleration section is dynamically calculated in advance to effectively compensate for the positioning deviation caused by motion inertia, ensuring that the robot accurately stays at the geometric center of the defect target area; at the same time, a pulse injection duty cycle subdivision control strategy is introduced, and according to the equivalent diameter size of the defects, the material is injected on demand, which not only ensures sufficient coverage of severe defects, but also avoids excessive accumulation of materials.

Description

Technical Field

[0001] This invention belongs to the field of municipal pipeline repair technology, specifically a method for intelligent identification and targeted repair control of pipeline defects based on deep learning. Background Technology

[0002] With the acceleration of urbanization, the aging and damage of municipal drainage pipe networks are becoming increasingly prominent, especially in the repair of small-diameter (DN100-500) pipes. Manual excavation and replacement methods, due to their high cost and significant impact, have been gradually replaced by trenchless repair technologies. Spray coating repair, with its advantages of fast construction speed, high lining strength, and strong adaptability, has become one of the important means of repairing small-diameter pipes. A spraying robot carries two-component polyurethane material into the pipe and uses a spray gun to evenly spray the mixed material onto the pipe wall, forming an inner lining layer to achieve structural reinforcement and leak prevention. In recent years, with the development of artificial intelligence technology, deep learning-based image recognition technology has been gradually introduced into the field of pipeline defect detection, providing a technological foundation for automated repair.

[0003] However, when applied to small-diameter spraying robots, existing technologies still face a series of unresolved challenges. First, the confined space of small-diameter pipes necessitates the robot dragging the raw material pipe, resulting in significant motion response lag due to load inertia and dragging resistance. This leads to a spatiotemporal misalignment between the identified and actual spraying positions when spraying is directly controlled based on real-time image recognition results, as the robot often overshoots the defect location due to inertia. Second, the rapid chemical reaction after mixing two-component polyurethane materials means that if point-like defects are addressed with fixed-point spraying, the material tends to solidify at the nozzle during the dwell time, causing damage to the spray gun. Blockages or uneven atomization during subsequent spraying affect repair quality and equipment reliability. In addition, the rotation of the spray gun during spraying can easily obstruct the camera's field of view, causing the robot to enter a visual blind spot. If there are unidentified obstacles or defects in the defect map ahead, collisions or missed spraying are very likely to occur. Existing methods lack a safe execution mechanism under limited perception conditions. The existence of these problems means that current deep learning-based defect recognition technology is mostly limited to the "recognition-marking" level, making it difficult to truly transform into stable and accurate targeted repair control, which restricts the promotion and application of small-diameter spraying robots in actual engineering.

[0004] To this end, the present invention provides a method for intelligent identification and targeted repair control of pipeline defects based on deep learning. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a deep learning-based intelligent identification and targeted repair control method for pipeline defects, comprising:

[0007] Images of the inside of the pipeline are collected, input into a deep learning model, and the type, location, and contour of defects are identified to generate a target area to be repaired labeled with the type of defect.

[0008] Based on the spatial distribution and defect type of the target area to be repaired, a clustering algorithm is used to divide the target area to be repaired into several spraying task units.

[0009] For each painting task unit, based on the defect type and the motion state parameters of the painting robot, a painting path is planned to generate an instruction sequence containing the actions of the painting primitives.

[0010] During the execution of the spraying unit action, the residence time of the two-component material and the occlusion status of the spray gun on the camera are monitored in real time. If the risk of material solidification and blockage of the spray gun is detected, an automatic micro-motion cleaning action is executed. The overlap between the spray gun posture and the camera's field of view is continuously monitored. If it is determined to be a perception-limited state, a memory navigation process is executed based on the perception-limited state. After the field of view is restored, the real-time image is re-acquired and compared and verified. If a deviation is found, a correction command is triggered.

[0011] The beneficial effects of this invention are as follows:

[0012] By analyzing images of the inner wall of the pipe, the type of defect can be accurately identified, the severity of the defect can be assessed based on the geometric data of the defect, and only severe defects can be repaired by subsequent spraying, thus improving repair efficiency and material utilization.

[0013] For point defects, the length of the pre-deceleration segment is dynamically calculated to effectively compensate for the positioning deviation caused by motion inertia, ensuring that the robot accurately stops at the geometric center of the defect target area. At the same time, a pulse jet duty cycle subdivision control strategy is introduced to achieve on-demand material spraying based on the equivalent diameter of the defect, which ensures sufficient coverage of severe defects while avoiding excessive material accumulation.

[0014] To address the annular defects, the robot's roll angle is detected in real time by an inertial measurement unit, and the spray gun rotation mechanism is controlled to compensate in the reverse direction during the attitude pre-adjustment section. This ensures that the center of rotation of the spray gun coincides with the axis of the pipeline, effectively solving the problem of spray trajectory deviation caused by robot tilt. In the rotating spray section, the travel speed and rotation speed are dynamically matched according to the target spray thickness and material flow rate per unit time, forming a uniform and continuous spiral spray trajectory.

[0015] The coating thickness is monitored in real time by a laser displacement sensor. When insufficient thickness is detected, the travel speed is automatically reduced and the amount of material is increased. When poor circumferential overlap is detected, the starting angle of the next week's spraying is adjusted. The uniformity of spraying in the middle of the spray gun is better than that at the edge, so as to compensate and repair the poor overlap area. The repair quality is ensured through multi-level compensation and effect verification.

[0016] Real-time monitoring of the residence time of a single pulse spray. When the residence time exceeds the preset threshold of the material gel time, a micro-cleaning action is automatically inserted. The mechanical movement breaks the static contact interface between the material and the air at the nozzle, preventing the material from solidifying and clogging at the nozzle, thus improving the operational reliability of the spraying equipment and reducing maintenance costs.

[0017] The system continuously monitors the overlap between the spray gun's posture and the camera's field of view. When it determines that the spray gun has entered the working position and may obstruct the camera, it is marked as a state of limited perception. Real-time image recognition is automatically paused, and the system switches to memory navigation mode. It then calls up a pre-stored memory map of local defects to continue spraying. Once the field of view is restored, real-time images are reacquired for effect verification and deviation correction. This effectively solves the problem of blind spots caused by the spray gun obstructing the camera during the spraying process, avoids safety accidents such as missed spraying and collisions, and ensures the continuity and safety of the spraying operation. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of the steps in Embodiment 1 of the present invention;

[0020] Figure 2 This is a module architecture diagram of Embodiment 2 of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0022] Example 1

[0023] One of the core inventive points of this invention is that, in response to the three major coupled challenges faced by small-diameter spraying robots when performing targeted repair—positioning deviation, material curing, and perception conflict—a deep learning-based intelligent pipeline defect identification and targeted repair control method is proposed by integrating deep learning-based defect identification, path planning based on motion state and material properties, and memory navigation mechanism under perception limitations. This method optimizes the closed-loop control of the entire process from defect perception to precise execution in the complex environment of small-diameter (DN100-500) pipelines, and solves the problem that traditional methods cannot achieve stable and accurate repair due to robot inertia passing over defects, curing of two-component materials at the nozzle, and visual blind spots caused by spraying obstruction.

[0024] Please see Figure 1 As shown in the embodiment of the present invention, the intelligent identification and targeted repair control method for pipeline defects based on deep learning includes the following steps:

[0025] Step S1: Acquire images of the inside of the pipeline, input them into a deep learning model, identify the type, location, and contour of defects, and generate target areas to be repaired labeled with the type of defects;

[0026] Among them, the defect types include at least point defects and ring defects;

[0027] The specific process of step S1 is as follows: control the high-definition camera on the spraying robot to continuously capture images of the inner wall of the pipe at a preset acquisition frequency (such as 5 frames per second) during the movement, and record the robot position encoder data corresponding to each frame of image to establish the mapping relationship between the image and the spatial position.

[0028] The acquired images are preprocessed, including grayscale conversion, noise reduction and enhancement, and distortion correction.

[0029] Among them, the noise reduction enhancement adopts an adaptive histogram equalization algorithm to improve the image contrast in low-light environments inside the pipe; the distortion correction is based on the camera calibration parameters to eliminate the barrel distortion caused by the wide-angle lens shooting in the narrow pipe, and ensure the measurement accuracy of the defect geometry.

[0030] The preprocessed image is input into a pre-trained deep learning semantic segmentation model, which outputs a pixel-level defect segmentation map.

[0031] Among them, the deep learning semantic segmentation model adopts an encoder-decoder architecture. The encoder part uses a lightweight convolutional neural network to adapt to the real-time requirements of the robot embedded platform, and the decoder part fuses multi-scale features through skip connections to achieve accurate segmentation of defects of different sizes.

[0032] In the output defect segmentation map, each pixel is labeled as one of three categories: normal, point defect, or ring defect. The geometric features of each connected component are extracted, including area, aspect ratio, principal axis direction, and minimum bounding rectangle.

[0033] Based on the extracted geometric features and the recorded image position mapping relationship, the actual coordinate position, geometric size and severity level of each defect in the three-dimensional space of the pipeline are calculated.

[0034] Furthermore, the calculation process for the actual coordinate position of the defect is as follows: using the robot position encoder data, the axial travel distance of the robot in the pipe during each frame image capture is obtained. Combined with the pre-calibrated camera intrinsic parameters (focal length, principal point coordinates) and the camera's installation position relative to the robot center (axial installation offset, circumferential installation angle), the defect pixel coordinates are converted into three-dimensional coordinates in a cylindrical coordinate system with the pipe's starting point as the origin. ;

[0035] The axial coordinate x is the sum of the axial travel distance, the axial installation offset of the camera relative to the robot center, and the axial offset of the defect relative to the point pointed to by the camera's optical axis.

[0036] The axial offset of the defect relative to the camera's optical axis is calculated based on the deviation between the defect's longitudinal pixel coordinates in the image and the image center, the camera's focal length parameter, and the actual distance from the defect point to the camera (using the pipe design radius or measured value from the range sensor), through the geometric relationship of similar triangles in perspective projection.

[0037] Circumferential angle The purpose is to calculate the projection angle of the defect on the pipe wall based on the lateral position of the pixel and the camera's field of view.

[0038] Radial coordinate r: Take the pipe design radius (defect located on the pipe wall surface); if a distance sensor is equipped, use the measured distance.

[0039] The calculation process for the geometric dimensions of a defect is as follows: For point defects (such as holes and cracks): Calculate the equivalent diameter. Where A is the projected area of ​​the defect on the pipe wall.

[0040] For circumferential defects (such as circumferential cracks and misalignments): calculate the circumferential length and axial width;

[0041] The size calculation is based on the principle of similar triangles in the camera imaging model, and is accurately converted by combining the depth information of the actual distance between the defect point and the camera.

[0042] The process for assessing the severity level of defects is as follows: based on the defect type and size parameters, the severity level is divided according to preset rules;

[0043] For point defects, if the equivalent diameter is greater than or equal to the diameter threshold (10mm), it is judged as a severe defect; otherwise, it is a mild defect.

[0044] For circumferential defects, if the circumferential length accounts for a proportion of the pipe circumference greater than or equal to the proportion threshold (30%) and the axial width is greater than or equal to the width threshold (3mm), it is judged as a severe defect; otherwise, it is a mild defect.

[0045] The defect areas marked as severe defects are identified as repair target areas. Each repair target area is generated with a unique identifier and stores the spatial coordinate range, defect type, actual coordinate location, and severity level.

[0046] Step S2: Based on the spatial distribution and defect type of the target area to be repaired, a clustering algorithm is used to divide the target area to be repaired into several spraying task units;

[0047] The specific process of step S2 is as follows: acquire the data of all target areas to be repaired, and take the geometric center point coordinates of each target area to be repaired as the representative point of that target area;

[0048] Based on the physical parameters and process requirements of the painting robot, two key parameters of the clustering algorithm are set: spatial distance threshold and minimum number of cluster points;

[0049] Spatial distance threshold: set according to the effective spraying width of the spray gun, that is, the product of the effective spraying width and the overlap coefficient, where the overlap coefficient is 0.8-1.2, to ensure that the target area within the distance of the threshold can be covered by a single spray.

[0050] Minimum number of cluster points: set according to process requirements, usually 1, that is, a single isolated target area can also be an independent spraying task unit;

[0051] Density-based clustering (DBSCAN) was used to cluster representative points of all target areas to be repaired. The specific steps are as follows:

[0052] A neighborhood is constructed with a representative point of each target area to be repaired as the center and a spatial distance threshold as the radius;

[0053] Calculate representative points of any two target areas to be repaired , Spatial distance between The distance calculation formula takes into account both axial distance and circumferential angle difference: ;

[0054] in, The minimum value of the circumferential angle difference (considering the periodicity of the angle, take...) and The smaller value;

[0055] The target areas to be repaired with spatial distances less than the spatial distance threshold are divided into the same cluster, forming several initial clusters;

[0056] For each initial cluster generated, examine the defect types of all target regions to be repaired within the cluster.

[0057] If all target areas to be repaired within a cluster are of the same defect type (all point-like or all ring-like), then the cluster is retained as the spraying task unit.

[0058] If a cluster contains both point-like and ring-like defect types, the cluster is split: the cluster is divided into two or more sub-clusters using the boundaries of different defect types as dividing lines, ensuring that each sub-cluster contains only a single defect type; the split sub-clusters are each used as independent spraying task units.

[0059] For each spraying task unit, perform the following operations:

[0060] Calculate the corresponding centroid coordinates as the target reference point for subsequent path planning;

[0061] Determine the outer envelope range: axial minimum and maximum values, circumferential minimum and maximum values.

[0062] Step S3: For each painting task unit, combine the defect type and the motion state parameters of the painting robot to plan the painting path and generate an instruction sequence containing the painting element actions;

[0063] The basic actions of spraying include: spraying basic actions for point defects and spraying basic actions for ring defects.

[0064] For point defects: Generate a three-stage action command consisting of an early deceleration stage, a precise stop stage, and a pulse injection stage. The length of the early deceleration stage is dynamically calculated based on the current travel speed and load inertia to compensate for positioning deviations caused by motion inertia.

[0065] For circumferential defects: Generate a three-stage action command consisting of a posture pre-adjustment segment, a rotation spraying segment, and an overlap compensation segment. The posture pre-adjustment segment adjusts the rotation center of the spray gun according to the real-time detected robot roll angle to ensure that the spraying trajectory is parallel to the tangential direction of the pipe wall.

[0066] The specific process of step S3 is as follows: acquire data for each spraying task unit, including: outer envelope range, centroid coordinates, and defect type;

[0067] Collect real-time motion parameters of the spraying robot, including: current travel speed, load inertia and drag resistance coefficient of the material pipe;

[0068] Among them, the load inertia is calculated in real time based on the robot's body mass, the length of the raw material pipe, and the amount of remaining material.

[0069] The drag resistance coefficient of the raw material pipe is obtained by looking up a table using a pre-calibrated resistance-speed curve;

[0070] Path planning based on defect type includes the following:

[0071] For path planning of point defects, a three-stage motion command is generated, consisting of an early deceleration segment, a precise stop segment, and a pulse injection segment:

[0072] Early deceleration phase: Calculate the deceleration length of the painting robot based on the robot's current travel speed and the system's load inertia. The calculation method is: divide the square of the current travel speed by twice the maximum deceleration, and then multiply by the safety factor.

[0073] The maximum deceleration is not a fixed value, but is dynamically adjusted based on the current load inertia and drag resistance coefficient. The larger the load inertia or the greater the drag resistance, the smaller the maximum deceleration and the longer the required deceleration distance. The safety factor is 1.2 to 1.5 to compensate for additional inertial deviations caused by uneven friction on the inner wall of the pipeline and fluctuations in the drag of the raw material pipe.

[0074] The starting point of the pre-deceleration phase is: the axial coordinate value of the center of mass minus the calculated deceleration length;

[0075] Precision stopping section: When the painting robot decelerates to the axial coordinate position of the geometric center of the defect target area, the traveling speed drops to zero. At this time, the center of the spray gun is directly facing the geometric center of the defect target area in the axial direction. The circumferential angle of the spray gun is adjusted by the rotation mechanism so that the spray gun points to the geometric center of the defect target area. The axial position is controlled by feedback from the axial encoder, and the circumferential angle is controlled by feedback from the rotary encoder. The stopping position and pointing angle are continuously corrected to ensure that the axial positioning error is controlled within ±2 mm and the circumferential positioning error is controlled within ±2 degrees.

[0076] Pulse spraying section: Pulse spraying is performed at a precise stopping position. The spraying sequence is: the spray gun is turned on and sprays for a period of time, then turned off, and then turned on again after a period of time, forming a pulse cycle.

[0077] The ratio of the pulse's on-time to its off-time (i.e., the duty cycle) is set according to the severity level of the defect:

[0078] Since severe defects were previously selected as the target areas for repair, this step targets the defects that have been confirmed to require repair. The duty cycle parameters can be further subdivided within the severe range based on the specific geometric dimensions of the defects (such as the equivalent diameter of point defects).

[0079] For severe point defects with a large equivalent diameter (≥15mm), the ratio of opening time to closing time is three to one to ensure sufficient material coverage.

[0080] For severe point defects with a small equivalent diameter (<15mm), the ratio of opening time to closing time is 2:1 to ensure coverage while avoiding excessive material accumulation.

[0081] The total spraying time is calculated based on the defect area and the spraying flow rate per unit time to ensure that the material can completely cover the entire defect area;

[0082] For path planning of circumferential defects, a three-stage action command is generated, comprising a posture pre-adjustment segment, a rotary spraying segment, and an overlap compensation segment:

[0083] Attitude pre-adjustment stage: Before entering the annular defect area, the roll angle of the robot body (i.e., the angle of rotation of the robot around the pipe axis) is detected in real time by the inertial measurement unit sensor.

[0084] If the detected roll angle exceeds the corresponding preset threshold (e.g., ±3 degrees), the spray gun rotation mechanism is controlled to rotate in the opposite direction by the same angle to ensure that the rotation center of the spray gun coincides with the pipeline axis.

[0085] After the attitude is pre-adjusted, the spray gun starts to rotate, and the rotation speed is linked to the robot's travel speed according to a preset relationship;

[0086] Rotary spraying section: The spray gun rotates at a constant angular velocity, while the robot moves along the axial direction of the pipe at a certain travel speed, forming a spiral spraying trajectory.

[0087] The travel speed is determined by calculating the target coating thickness and the material flow rate per unit time, and then dividing the material flow rate per unit time by the product of the target coating thickness and the effective coating width of the spray gun to obtain the required robot travel speed.

[0088] The rotation speed is determined by matching the travel speed and the effective spray width of the spray gun to ensure that the axial overlap between adjacent rotation cycles is positive. Usually, the overlap is taken as 20% to 30% of the effective spray width. That is, the relationship between the rotation speed and the travel speed should ensure that the axial movement distance of the robot is less than the effective spray width of the spray gun for each rotation of the spray gun.

[0089] Overlap compensation section: During the rotation spraying section, monitor the real-time thickness value fed back by the spraying thickness sensor;

[0090] If the current coating thickness is detected to be lower than the target thickness, it is determined that there is insufficient thickness in the current area. The control system gradually reduces the robot's axial travel speed according to the preset adjustment step size (such as 5% of the initial travel speed).

[0091] Since the material flow rate per unit time is fixed, a decrease in travel speed means an increase in the amount of material accumulated per unit area, thereby increasing the coating thickness.

[0092] When the thickness sensor feedback value returns to the target thickness, the control system gradually restores the travel speed to the original set value to avoid overcompensation that could cause the thickness to exceed the limit.

[0093] During the rotary spraying process, the starting and ending angle positions of the spraying are recorded after each complete rotation of the spray gun.

[0094] If the thickness of the edge area sprayed in the previous week is found to be less than 80% of the target thickness, and the continuous angle range of the low-thickness area in the circumferential direction exceeds 10 degrees, it is determined that there is a circumferential overlap problem in the area. When rotating and spraying in the next week, the starting angle of the spray gun will be offset by a preset offset amount (such as 15 degrees).

[0095] After the initial angle shifts, the area with poor overlap will no longer be the edge overlap area of ​​the next week's spray, but will be covered by the middle area of ​​the next week's spray. By taking advantage of the fact that the uniformity of spraying in the middle of the spray gun is better than that at the edge, the area with poor overlap can be compensated and repaired.

[0096] If poor overlap is detected for two consecutive weeks, an alternating offset strategy will be adopted in subsequent rotations, i.e., offset by +15 degrees in odd-numbered cycles and offset by -15 degrees in even-numbered cycles, to ensure that the overlap area is evenly distributed in multiple rotation cycles and to avoid repeated overlap defects at the same angle position.

[0097] After each compensation action is performed, the thickness data of the overlapping area is re-acquired in the next rotation cycle and compared with the thickness value before compensation:

[0098] If the thickness is increased to more than 90% of the target thickness, the compensation is deemed effective, and the current compensation parameters are maintained and the process continues.

[0099] If the thickness still does not meet the standard, the compensation range will be further increased (such as increasing the speed reduction from 5% to 8%, or increasing the starting angle offset from 15 degrees to 25 degrees) until the repair requirements are met.

[0100] Step S4: During the execution of the spraying unit action, monitor the residence time of the two-component material and the occlusion status of the spray gun on the camera in real time.

[0101] If the residence time of a single pulse spray exceeds the preset threshold of the material gel time, a micro-cleaning action is automatically inserted to drive the spray gun to make a slight oscillation to prevent the material at the nozzle from solidifying.

[0102] Specifically, the spraying status of the spray gun is monitored in real time, including the dwell time of a single pulse spray and the continuous working time of the spray gun.

[0103] When the residence time of a single pulse spray exceeds the preset threshold of the material gel time, it is determined that there is a risk of material solidification and blockage in the spray gun, and an automatic insertion micro-motion cleaning process is executed.

[0104] The material gel time is pre-determined based on the chemical properties of the two-component polyurethane material. The preset threshold is usually taken as 80% to 90% of the material gel time, with a safety margin reserved.

[0105] The automatic insertion micro-motion cleaning process includes: controlling the spray gun rotation mechanism or the spray gun swing mechanism to drive the spray gun to swing slightly at the current position. The swing amplitude is ±2 mm to ±5 mm, the swing frequency is 2 to 3 times per second, and the swing duration is 1 to 2 seconds.

[0106] Micro-oscillations disrupt the static contact interface between the material and air at the muzzle through mechanical movement, while utilizing the material's own fluidity to prevent excessive local accumulation and solidification.

[0107] After the micro-cleaning action is completed, the original spraying unit action continues. If the dwell time is still detected to exceed the threshold, the micro-cleaning is triggered again, forming a periodic protection mechanism.

[0108] The overlap between the spray gun posture and the camera's field of view is continuously monitored. When it is determined that the spray gun has entered the working position and may block the camera, it is marked as a perception-limited state. Based on the perception-limited state, the defect recognition of the real-time image is paused, and the memory navigation mode is switched to call the pre-stored local defect memory map to continue spraying. After the field of view is restored, the real-time image is re-acquired, and the actual spraying effect is compared and verified with the memory map. If a deviation is found, a correction command is triggered.

[0109] Specifically, real-time monitoring of the overlap between the spray gun's posture and the camera's field of view includes;

[0110] The current circumferential angle of the spray gun is obtained in real time by an angle sensor installed on the spray gun rotation mechanism, and the installation angle and field of view of the camera are also obtained.

[0111] Based on the relative position of the spray gun's geometry and the camera, calculate whether the spray gun is within the camera's effective field of view at the current angle.

[0112] When the circumferential angle of the spray gun enters the preset obstruction angle range (the spray gun is pointing towards the camera and the angle deviation is less than ±15 degrees), and the spray gun is performing a rotary spraying or pulse spraying action, it is determined that the spray gun has entered the working position and may obstruct the camera, and is marked as a perception-limited state. The current moment is the starting point of the perception-limited state.

[0113] Entering a state of limited perception, the memory navigation process is executed:

[0114] Stop the defect identification process of the current real-time image stream to avoid misidentification or missed identification due to occlusion;

[0115] Retrieve the spatial coordinates, defect types, and geometric data of all target areas to be repaired within the pipe segment where the current spraying task unit is located, as previously stored;

[0116] The robot control system continues the spraying operation based on the target area location information and the planned sequence of spraying element action instructions, no longer relying on real-time image feedback for decision-making;

[0117] After the current spraying action is completed and the spray gun is reset to an unobstructed position, the camera field of view is detected to be restored, the memory navigation mode is exited, and real-time image acquisition and recognition are restarted.

[0118] Reacquire real-time images of the sprayed area and compare the actual spraying effect with the expected repair area in the memory map for verification. Specific verification content includes:

[0119] Whether the sprayed area completely covers the target area to be repaired marked in the memory map;

[0120] Check whether there are any missed areas, uneven thickness, or spraying areas that exceed the target area boundary in the spraying area;

[0121] If a deviation is detected between the actual spraying effect and the memory map (such as the target area not being completely covered or there being missed spraying), a correction command is triggered. The specific correction method is as follows:

[0122] Based on the location and extent of the deviation area, a local respray command is generated;

[0123] Insert the supplementary spraying command into the subsequent task queue, and execute the supplementary spraying first after the current task unit is completed;

[0124] If the deviation is significant (e.g., the missed spray area exceeds 20% of the target area), then the subsequent tasks should be suspended and the team should immediately return to the deviation area to perform respraying.

[0125] If the verification result matches the stored data, then continue to execute the next painting task unit.

[0126] If, under conditions of limited perception, it is simultaneously determined that there is a risk of material solidification and clogging in the spray gun, the following actions must be taken according to priority:

[0127] Prioritize micro-motion cleaning to ensure the spray gun does not clog;

[0128] If the micro-movement cleaning action is completed and the system is still in a state of limited perception, the spraying will continue to be performed in the memory navigation mode.

[0129] If the spray gun shifts during the micro-motion cleaning process, causing the obstruction to be removed, real-time image recognition will be immediately resumed, and effect verification will be performed.

[0130] This embodiment constructs a closed-loop control framework encompassing "intelligent defect identification and target area generation—target area clustering and task unit division—defect adaptive path planning—spraying execution and adaptive control." It utilizes a deep learning model to identify point and ring-shaped defects and selects severe defects to generate target areas for repair. Spraying task units are optimized through clustering algorithms. For point defects, a pre-deceleration compensation mechanism considering motion inertia and drag resistance is introduced, along with pulse jet duty cycle subdivision control. For ring-shaped defects, roll angle attitude pre-adjustment, spiral spraying speed matching, and overlap compensation strategies based on thickness feedback and angle offset are implemented. Simultaneously, at the execution level, real-time monitoring of material gelation time triggers micro-motion cleaning to prevent nozzle solidification, and detection of visual obstruction automatically switches to memory navigation mode, while also verifying and correcting execution effects. This effectively solves technical problems such as robot inertia causing positioning deviations when passing over defects, easy solidification and clogging of the spray gun due to fixed-point dwell of two-component materials, and missed spraying or collisions caused by visual blind spots caused by spraying obstructions. This improves the positioning accuracy, spraying quality, equipment reliability, and operational safety of targeted repair in small-diameter pipes.

[0131] Example 2

[0132] Based on the same inventive concept as the deep learning-based intelligent identification and targeted repair control method for pipeline defects in the foregoing embodiments, such as Figure 2 As shown, this application provides a deep learning-based intelligent identification and targeted repair control system for pipeline defects, wherein the system specifically includes:

[0133] Defect identification and target area generation module: Acquires images of the inside of the pipeline, inputs them into a deep learning model, identifies the type, location and contour of defects, and generates target areas to be repaired labeled with the defect type;

[0134] Target area clustering and unit division module: Based on the spatial distribution and defect type of the target area to be repaired, a clustering algorithm is used to divide the target area to be repaired into several spraying task units;

[0135] Defect Adaptive Path Planning Module: For each painting task unit, the module plans the painting path based on the defect type and the motion state parameters of the painting robot, generating a sequence of instructions containing the actions of the painting primitives.

[0136] Among them, the motion state parameters include: current travel speed, load inertia and drag resistance coefficient of the raw material pipe;

[0137] The basic actions of spraying include: spraying basic actions for point defects and spraying basic actions for ring defects.

[0138] Point-shaped defect command generation unit: generates a three-stage action command including an early deceleration segment, a precise stop segment, and a pulse injection segment. The length of the early deceleration segment is dynamically calculated based on the current travel speed and load inertia to compensate for the positioning deviation caused by motion inertia.

[0139] Circular defect instruction generation unit: Generates a three-stage action instruction including a posture pre-adjustment section, a rotation spraying section, and an overlap compensation section. The posture pre-adjustment section adjusts the rotation center of the spray gun according to the real-time detected robot roll angle to ensure that the spraying trajectory is parallel to the tangential direction of the pipe wall.

[0140] Spraying execution and control module: During the execution of the spraying unit action, the residence time of the two-component material and the occlusion status of the spray gun on the camera are monitored in real time.

[0141] Material curing analysis unit: If the residence time of a single pulse spray exceeds the preset threshold of the material gel time, a micro-cleaning action is automatically inserted to drive the spray gun to make a slight oscillation to prevent the material at the nozzle from curing.

[0142] Limited Field of View Analysis Unit: Continuously detects the overlap between the spray gun posture and the camera's field of view. When it is determined that the spray gun has entered the working position and may block the camera, it is marked as a limited perception state. Based on the limited perception state, the defect recognition of the real-time image is paused, and the system switches to memory navigation mode. The pre-stored local defect memory map is called to continue spraying. After the field of view is restored, the real-time image is re-acquired, and the actual spraying effect is compared and verified with the memory map. If a deviation is found, a correction command is triggered.

[0143] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent identification and targeted repair control method for pipeline defects, characterized in that: include: Images of the inside of the pipeline are collected, input into a deep learning model, and the type, location, and contour of defects are identified to generate a target area to be repaired labeled with the type of defect. Based on the spatial distribution and defect type of the target area to be repaired, a clustering algorithm is used to divide the target area to be repaired into several spraying task units. For each painting task unit, based on the defect type and the motion state parameters of the painting robot, a painting path is planned to generate an instruction sequence containing the actions of the painting primitives. For point defects: generate a three-stage action command including an early deceleration stage, a precise stop stage, and a pulse spray stage; for ring defects: generate a three-stage action command including an attitude pre-adjustment stage, a rotating spray stage, and an overlap compensation stage. During the spraying process, the residence time of the two-component material is monitored in real time. If the risk of material solidification and blockage is detected in the spray gun, an automatic micro-cleaning process is executed. The overlap between the spray gun posture and the camera's field of view is detected. If the perception is restricted, a memory navigation process is executed. After the field of view is restored, the real-time image is reacquired and compared. If a deviation is found, a correction command is triggered.

2. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: The process of generating a target area to be repaired with labeled defect types is as follows: Images of the inner wall of the pipe are acquired, and a mapping relationship between the images and spatial locations is established. The preprocessed images are then input into a deep learning semantic segmentation model, which outputs a pixel-level defect segmentation map. Each pixel is labeled as one of three categories: normal, point defect, or ring defect, and geometric features are extracted. Based on the defect type and geometric dimensions, the severity level is determined, the defect areas marked as severe defects are extracted, and these areas are identified as the target areas for repair.

3. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 2, characterized in that: The process for obtaining defect regions marked as severe defects is as follows: Based on the extracted geometric features and the recorded image position mapping relationship, the axial travel distance when each frame of the image was captured is obtained, and the defect pixel coordinates are converted into three-dimensional coordinates in a cylindrical coordinate system with the pipe start point as the origin. The actual geometric dimensions are calculated based on the three-dimensional coordinates of the defect in the cylindrical coordinate system. For point defects, the equivalent diameter is calculated, and for annular defects, the circumferential length and axial width are calculated. For point defects, if the equivalent diameter is greater than or equal to the diameter threshold, it is judged as a severe defect. For ring defects, if the ratio of the circumferential length to the pipe circumference is greater than or equal to the ratio threshold and the axial width is greater than or equal to the width threshold, it is judged as a severe defect.

4. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: The process of acquiring the spraying task unit is as follows: The coordinates of the geometric center point of each target area to be repaired are used as the representative point of the target area to be repaired. Density clustering algorithm is used to cluster representative points of all target areas to be repaired, forming several initial clusters; For each generated initial cluster, check the defect type of all target areas to be repaired within the cluster. If all target areas to be repaired within the cluster are of the same defect type, retain the cluster as a spraying task unit. If the cluster contains both point and ring defect types, split the cluster and divide it into two or more sub-clusters using the boundaries of different defect types as dividing lines. The split sub-clusters are then used as independent spraying task units. For each painting task unit, the centroid coordinates are calculated as the target reference point for subsequent path planning, and the outer envelope range is determined.

5. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: Early deceleration phase: Based on the robot's current travel speed and the system's load inertia, calculate the deceleration length of the painting robot. The calculation method is: divide the square of the current travel speed by twice the maximum deceleration, and then multiply by the safety factor. The starting point of the early deceleration phase is: the axial coordinate value of the center of mass minus the calculated deceleration length. The duty cycle of the pulse jet segment is set according to the equivalent diameter of the defect: For severe point defects with an equivalent diameter of 15 mm or more, the ratio of opening time to closing time is 3:1; for severe point defects with an equivalent diameter of less than 15 mm, the ratio of opening time to closing time is 2:

1.

6. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: Attitude pre-adjustment stage: Before entering the annular defect area, the roll angle of the robot body is detected in real time by the inertial measurement unit sensor; if the detected roll angle exceeds the preset threshold, the spray gun rotation mechanism is controlled to rotate in the opposite direction by the same angle.

7. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: Rotary spraying section: Divide the material flow rate per unit time by the product of the target spraying thickness and the effective spraying width of the spray gun to obtain the required robot travel speed; the rotation speed is matched with the travel speed so that the axial movement distance of the robot is less than the effective spraying width of the spray gun for each rotation of the spray gun, and the axial overlap between adjacent rotation cycles is 20% to 30% of the effective spraying width.

8. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: Overlap compensation section: During the rotation spraying section, if the current spraying thickness is detected to be lower than the target thickness, the robot's axial travel speed is gradually reduced according to the preset adjustment step size until the thickness is restored to the target thickness; The spray gun records the starting angle position after each rotation and collects the thickness distribution data of the sprayed edge area in the previous week. If it is detected that the thickness of the sprayed edge area in the previous week is less than 80% of the target thickness, and the continuous angle range of the low thickness area in the circumferential direction exceeds 10 degrees, then the starting angle of the spray gun will be offset by a preset amount when rotating and spraying in the next week. If poor overlap is detected for two consecutive weeks, an alternating offset strategy is adopted, that is, offset by a positive angle in odd-numbered weeks and offset by a negative angle in even-numbered weeks. After each compensation action is performed, the thickness data of the overlapping area is re-collected in the next rotation cycle and compared with the thickness value before compensation. If the thickness still does not meet the standard, the compensation range is further increased.

9. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: If a risk of material solidification and clogging is detected in the spray gun, an automatic insertion micro-motion cleaning process is executed, including: The spray gun's spraying status is monitored in real time. If the residence time of a single pulse spray exceeds the preset threshold of the material's gel time, it is determined that the spray gun is at risk of material solidification and clogging. The automatic insertion micro-motion cleaning process includes: controlling the spray gun rotation mechanism or the spray gun swing mechanism to drive the spray gun to swing at the current position; The oscillation disrupts the static contact interface between the material and air at the muzzle through mechanical movement, while simultaneously utilizing the material's own fluidity to prevent excessive local accumulation and solidification.

10. The method for intelligent identification and targeted repair control of pipeline defects based on deep learning according to claim 1, characterized in that: If the state is determined to be perceptually limited, the memory navigation process is executed, including: Real-time acquisition of the spray gun's current circumferential angle, as well as the camera's installation angle and field of view; Based on the relative position of the spray gun's geometry and the camera, calculate whether the spray gun is within the camera's effective field of view at the current angle. When the circumferential angle of the spray gun enters the preset obstruction angle range, and the spray gun is performing rotary spraying or pulse spraying, it is marked as a perception-limited state. Stop the defect identification and processing of the current real-time image stream, and call up the stored spatial coordinates, defect types and geometric data of all target areas to be repaired in the pipe section where the current spraying task unit is located; the robot control system continues the spraying operation according to the target area position information and the planned spraying element action instruction sequence.

Images