A deep learning-based industrial conduit inner wall microscopic defect detection, positioning and tracking system and method

Abstract

The application discloses an AI visual detection and tracking system and method for microscopic defects of an inner wall of an industrial pipeline, which comprises an in-pipe visual acquisition device, a pose perception module, an image preprocessing module, a defect identification module, a defect tracking and positioning module and a result output module. The visual acquisition device moves in the pipeline with a carrying mechanism and acquires images of the inner wall of the pipeline. The image preprocessing module corrects distortion, balances illumination, removes noise and performs unwrapping mapping on the images. The defect identification module identifies microscopic defects such as cracks, corrosion, pinholes, welding defects and coating peeling by using a trained deep learning model. The defect tracking and positioning module combines adjacent frame image features and pose information to perform cross-frame correlation, spatial positioning and size evaluation on the same defect. The result output module generates defect positions, categories, severity and repair suggestions. The application can improve the accuracy, continuity and positioning efficiency of the detection of microscopic defects of the inner wall of the industrial pipeline and reduce manual detection errors.

Description

Technical Field

[0001] This invention relates to the fields of non-destructive testing of industrial conduits, machine vision inspection, and artificial intelligence image recognition, and particularly to an AI vision inspection and tracking system and method suitable for detecting microscopic defects in the inner walls of industrial conduits, spatially locating defects, and continuously tracking defects. Specifically, this invention can be applied to the inner wall inspection of various industrial conduits, pipelines, pipes, or similar closed cavity structures in fields such as petrochemicals, power, metallurgy, pharmaceuticals, food processing, aerospace, and equipment manufacturing. Through in-pipe visual image acquisition, image preprocessing, deep learning defect recognition, pose information fusion, and cross-frame defect tracking, it achieves automatic identification, location, dimensional assessment, and result output of microscopic defects such as cracks, corrosion, pinholes, pits, weld abnormalities, coating peeling, and foreign matter adhesion on the inner wall of the conduit. Background Technology

[0002] 1. Technical Background of Industrial Pipeline Inner Wall Inspection

[0003] Industrial conduits are widely used in petrochemical, power, metallurgy, pharmaceutical, food processing, aerospace, and equipment manufacturing industries to transport gases, liquids, slurries, steam, or other process media. During long-term service, industrial conduits are often subjected to factors such as media corrosion, temperature changes, pressure fluctuations, mechanical vibration, welding residual stress, sediment erosion, and insufficient cleaning and maintenance. As a result, their inner walls are prone to defects such as cracks, corrosion, pinholes, pits, abnormal welds, coating peeling, and foreign matter adhesion.

[0004] In their early stages, these defects typically exhibit characteristics such as small size, weak boundaries, low contrast, irregular shape, and concealed location. If not detected and located in a timely manner, these defects may expand as the conduit's service life increases, affecting its sealing performance, pressure resistance, flow efficiency, and operational safety. In industrial settings involving high temperatures, high pressures, flammable or explosive materials, toxic substances, or high cleanliness requirements, defects on the conduit's inner wall can also cause leaks, blockages, contamination, downtime for maintenance, and even safety accidents. Therefore, accurate, continuous, and traceable detection of microscopic defects on the inner wall of industrial conduits is of significant engineering importance.

[0005] Because industrial conduits are typically characterized by narrow spaces, enclosed interiors, varying diameters, numerous bends, complex lighting conditions, and difficulties in direct manual observation, traditional external inspection methods are insufficient for effectively assessing the condition of the conduit's internal walls. To obtain images or defect information about the conduit's interior, existing inspection methods usually require the use of industrial endoscopes, pipeline robots, camera inspection devices, ultrasonic testing devices, eddy current testing devices, or other non-destructive testing equipment to enter the conduit or move along its exterior to complete the inspection.

[0006] 2. Existing industrial conduit inner wall inspection technology

[0007] Existing technologies for inspecting the inner walls of industrial conduits mainly include manual endoscopic inspection, pipeline robot vision inspection, ultrasonic inspection, eddy current inspection, X-ray inspection, and machine vision-based image inspection.

[0008] Manual endoscopic inspection typically involves inspectors using an industrial endoscope to enter the conduit and observe the inner wall of the conduit via real-time video or captured images, relying on experience to determine the presence of defects. This method uses relatively simple equipment and is suitable for inspecting conduits of a certain length and diameter. However, the results depend heavily on the inspector's experience and are easily affected by the observation angle, lighting conditions, image clarity, and operator fatigue. For microscopic defects such as tiny cracks, pinholes, shallow corrosion pits, and initial coating peeling, manual observation carries the risk of missed detections and misjudgments.

[0009] Pipeline robot visual inspection typically mounts a camera, lighting assembly, drive mechanism, and control unit on a movable support structure. This structure moves along the inside of the duct to acquire continuous images or videos. While this method improves the inspection range and efficiency, in practical applications, images of the duct's inner wall often suffer from lens distortion, ring reflections, localized shadows, motion blur, dirt interference, and variations in shooting distance. Relying solely on manual video review or frame-by-frame screenshots for judgment remains insufficient for stable identification and precise localization of microscopic defects.

[0010] Non-destructive testing (NDT) methods such as ultrasonic testing, eddy current testing, and radiographic testing can be used to obtain information on defects inside or on the surface of conduit materials. Ultrasonic testing has a certain ability to detect wall thickness changes, internal cracks, or corrosion thinning; eddy current testing is suitable for detecting surface and near-surface defects in conductive materials; and radiographic testing can be used to detect some structural defects. However, these testing methods typically have high requirements for conduit material, wall thickness, surface condition, coupling conditions, sensor placement, and the testing environment, and the test results are not easily directly correlated with visual images. For scenarios requiring simultaneous acquisition of defect morphology, image evidence, axial position, circumferential position, and a visual report, a single non-visual testing method has certain limitations.

[0011] Machine vision-based image detection methods acquire images of the inner wall of ducts using camera devices, and then employ image enhancement, edge detection, threshold segmentation, texture analysis, template matching, or machine learning algorithms for defect identification. While these methods can reduce the workload of manual judgment to some extent, traditional image processing algorithms typically rely on manually designed features and fixed thresholds. When the inner wall of the duct exhibits different materials, roughnesses, levels of contamination, lighting conditions, and defect morphologies, fixed rules are difficult to adapt to complex working conditions, easily leading to false positives, false negatives, or unstable identification.

[0012] With the development of deep learning technology, convolutional neural networks, object detection models, semantic segmentation models, and instance segmentation models are increasingly being applied to industrial surface defect identification. Deep learning models can learn the texture, boundary, and morphological features of defects from sample data, exhibiting stronger feature representation capabilities compared to traditional image processing methods. However, images of the inner walls of industrial conduits differ from ordinary planar industrial images. These images typically exhibit ring-shaped, arc-shaped, or perspective distortion features, accompanied by uneven lighting, reflections, high noise, and inter-frame variations caused by movement within the conduit. Simply applying deep learning models directly to single-frame image recognition is insufficient to meet the requirements for continuous tracking, spatial localization, and dimensional assessment of microscopic defects on the inner walls of conduits.

[0013] 3. Major problems with existing technologies

[0014] While existing industrial conduit inner wall inspection technologies can acquire images or defect information about the conduit interior to some extent, they still have the following problems in terms of automatic identification, continuous tracking, spatial positioning, and report generation of micro-defects.

[0015] First, the unstable image quality inside the pipe affects the accuracy of micro-defect identification. The narrow internal space of industrial conduits causes the distance and angle between the camera and the inner wall to change with the movement of the supporting structure. When the supplementary lighting component illuminates the curved inner wall, it easily creates reflections, shadows, or localized overexposure. Simultaneously, wide-angle lenses or endoscopic imaging introduce radial distortion, causing deformation of the defect's shape, size, and boundaries in the original image. Without preprocessing steps such as distortion correction, illumination equalization, image denoising, and inner wall unfolding mapping, micro-defects such as cracks, corrosion pits, pinholes, and coating peeling are easily obscured by background textures, weld textures, stains, or noise.

[0016] Second, single-frame image recognition cannot guarantee the continuity of defect detection results. Defects on the inner wall of industrial conduits often appear repeatedly in multiple consecutive images, and their image position, scale, and shape change with the movement of the supporting mechanism. If existing systems only identify each frame independently, the same defect may be counted repeatedly, and tracking may be interrupted due to blurring, occlusion, reflection, or low confidence in a particular frame. For detection tasks requiring the generation of defect numbers, defect trajectories, defect locations, and defect sizes, single-frame recognition results are insufficient to meet the requirements of continuity and traceability in engineering applications.

[0017] Third, there is a lack of effective correspondence between image detection results and the spatial location of the conduit. While some existing visual inspection equipment can acquire images or videos, it lacks the ability to fuse these with pose data such as axial displacement, attitude angles, movement speed, and mileage information of the supporting mechanism. Although inspectors can see images of defects, it is difficult to accurately determine the axial and circumferential positions of the defects within the conduit. For long-distance conduits, curved conduits, or complex piping systems, relying solely on video time points or manually recorded positions is insufficient to guarantee defect location accuracy, hindering subsequent re-inspection, repair, and replacement operations.

[0018] Fourth, defect size assessment lacks a unified coordinate transformation and multi-frame correction mechanism. The inner wall of industrial conduits has a curved structure, and the pixel size of a defect in an image does not directly equate to its actual size. Camera calibration parameters, conduit inner diameter parameters, shooting distance, image unfolding method, and pose changes all affect the size conversion results. If only the area of ​​a single frame pixel or manual estimation is used as the basis for defect size, it easily leads to inaccurate length, width, and area assessments. For defects such as micro-cracks, pinholes, and corrosion pits, size assessment errors further affect severity classification and maintenance decisions.

[0019] Fifth, the adaptability of defect recognition models to conduit inspection scenarios is insufficient. While existing deep learning detection models can be used for defect recognition, if the training samples do not cover common textures, weld morphologies, deposits, stains, reflective properties of different materials, and various types of defects on the inner walls of industrial conduits, the models are prone to insufficient generalization ability in actual inspections. Especially when the defect boundaries are blurred, the defect area is small, the background texture is complex, or the defect and stain morphology are similar, the defect category, boundary region, and confidence level output by the model may fluctuate.

[0020] Sixth, the output format of inspection results is not conducive to closed-loop maintenance. Some existing inspection systems mainly output raw videos, screenshots, or simple alarm information, lacking a unified compilation of defect numbers, defect categories, defect images, axial positions, circumferential positions, dimensional parameters, confidence levels, severity levels, and maintenance recommendations. Inspection personnel need to manually review videos and compile reports again, which is labor-intensive and prone to information omissions. For industrial scenarios requiring long-term monitoring, periodic re-inspections, or comparisons of multiple batches of pipelines, the lack of structured inspection reports will affect defect management and maintenance decisions.

[0021] 4. Needs for improved technical solutions

[0022] Based on the above, the field of industrial conduit inner wall micro-defect detection requires a detection and tracking technology solution that can adapt to the complex imaging environment inside the conduit. This solution should be able to acquire continuous images of the conduit inner wall through an in-conduit visual acquisition device during the movement of the supporting mechanism along the conduit's interior, and simultaneously acquire axial displacement, attitude angle, movement speed, and / or mileage information in conjunction with a pose perception module.

[0023] Simultaneously, this technical solution should be able to perform distortion correction, illumination equalization, image denoising, and inner wall unfolding mapping on the duct inner wall image, reducing the impact of duct inner curved surface imaging, uneven illumination, and noise interference on microscopic defect identification. The preprocessed image should be able to be input into a deep learning defect identification model trained from industrial duct defect samples to output defect category, defect boundary region, and defect confidence level.

[0024] Furthermore, this technical solution should be able to perform cross-frame correlation tracking of the same defect based on defect features, defect boundary regions, and the pose information of the supporting mechanism in adjacent frame images, avoiding duplicate counting and tracking interruptions, and further calculate the axial position, circumferential position, and dimensional parameters of the defect within the industrial conduit. For the same defect that appears repeatedly in multiple frames, it should also be able to perform position and size fusion to improve the stability of defect localization and size assessment.

[0025] Furthermore, the technical solution should be able to generate an inspection report based on the defect type, defect location, size parameters, and defect confidence level, and display the defect location on the unfolded diagram of the conduit or the three-dimensional conduit model, thereby forming a complete inspection process from image acquisition, image preprocessing, defect identification, defect tracking, spatial positioning, size evaluation to result output, meeting the actual needs of automated inspection and maintenance management of microscopic defects in the inner wall of industrial conduits. Summary of the Invention

[0026] 1. Purpose of the invention

[0027] The purpose of this invention is to provide an AI visual inspection and tracking system and method for microscopic defects in the inner wall of industrial conduits, which solves the problems existing in the current process of inspecting the inner wall of industrial conduits, such as unstable image quality, reliance on human experience for defect identification, discontinuous single-frame identification results of microscopic defects, easy repeated statistics of the same defect, difficulty in accurately corresponding the defect image position with the actual spatial position of the conduit, lack of unified conversion basis for defect size assessment, and difficulty in forming a structured maintenance basis for the inspection results.

[0028] This invention integrates the acquisition of visual images inside the pipe, synchronous acquisition of pose information, preprocessing of images inside the pipe, deep learning defect identification, cross-frame defect tracking, pipe coordinate system positioning, defect size assessment, defect severity classification, and output of inspection reports. This enables the microscopic defects on the inner wall of industrial pipes to be transformed from the original image state into identifiable, traceable, localizable, quantifiable, and verifiable inspection results.

[0029] Another objective of this invention is to provide a computer-based method for detecting microscopic defects in the inner wall of industrial conduits, enabling the processor to complete an automated detection process in the order of image acquisition, pose synchronization, image correction, defect identification, cross-frame correlation, position conversion, size evaluation, and report generation when executing the computer program.

[0030] Another objective of this invention is to provide an electronic device and a computer-readable storage medium that enable the detection method of this invention to be deployed in field detection terminals, edge computing devices, industrial detection hosts, servers, or offline analysis platforms, thereby meeting the implementation needs of different industrial conduit detection scenarios.

[0031] 2. Technical problems to be solved

[0032] To achieve the above objectives, the present invention mainly addresses the following technical problems.

[0033] S1: Problem of unstable image quality within the tube

[0034] The confined space inside industrial conduits means that the distance, angle, and orientation between the camera and the conduit's inner wall can easily change during inspection. When supplemental lighting illuminates the curved inner wall, it can easily cause localized reflections, shadows, overexposure, or brightness reduction. Furthermore, industrial endoscopes or wide-angle image sensors can introduce radial distortion when acquiring images of the conduit's inner wall, altering the boundaries, shape, and size of microscopic defects such as cracks, corrosion, pinholes, pits, weld defects, coating peeling, and foreign matter adhesion in the original image.

[0035] The present invention aims to solve how to perform distortion correction, illumination equalization, image denoising, and inner wall unwrapping mapping on industrial duct inner wall images under the aforementioned complex imaging conditions, so as to obtain preprocessed images suitable for subsequent defect identification and cross-frame tracking.

[0036] S2: Problem of insufficient accuracy in automatic identification of micro-defects

[0037] Existing manual endoscopic inspection or traditional image processing methods are easily affected by the inspector's experience, image quality, background texture, weld texture, contaminants and noise interference, making it difficult to reliably identify micro-defects with weak boundaries, small areas and irregular shapes.

[0038] The present invention aims to solve how to train a deep learning defect recognition model based on defect samples of industrial conduits, and use the model to output defect categories, defect boundary regions and defect confidence scores, thereby improving the automatic identification capability of microscopic defects on the inner wall of industrial conduits.

[0039] S3: Problem of discontinuous recognition results in a single frame

[0040] When the carrier moves along the inside of the industrial duct, the same defect usually appears consecutively in multiple adjacent frames. If only a single frame is identified independently, the same defect is easily counted repeatedly; if a frame has reflections, blurring, occlusion, or a decrease in confidence, the tracking of the same defect is easily interrupted.

[0041] The present invention aims to solve how to perform cross-frame correlation tracking of the same defect based on the defect features, defect boundary regions, and pose information of the supporting mechanism in adjacent frame images, so that the detection results of the same defect in multiple frame images can be merged into the same defect object.

[0042] S4: Problem of inaccurate defect spatial positioning

[0043] Current visual inspection results of catheter inner walls are typically presented as images, videos, or screenshots. However, the location of defects in the images cannot be directly converted into the actual axial and circumferential positions inside the catheter. For long-distance catheters, bent catheters, or complex pipelines, if the defect location cannot be accurately recorded, it will affect subsequent re-inspection, repair, replacement, and defect management.

[0044] The present invention aims to solve how to combine the axial displacement, attitude angle, movement speed and / or mileage information of the bearing mechanism in the industrial conduit to convert the position of the defect in the image coordinate system into the axial position and circumferential position in the industrial conduit coordinate system.

[0045] S5: Issues with instability in defect size assessment and severity grading

[0046] The inner wall of industrial conduits has a curved structure, and the pixel length, pixel width, and pixel area of ​​defects in images cannot be directly equated to their actual dimensions. Camera calibration parameters, conduit inner diameter parameters, image unfolding method, and changes in the attitude of the supporting mechanism all affect the defect size conversion results.

[0047] This invention aims to solve the problem of establishing the conversion relationship between pixel coordinates and the actual size of the conduit, and to combine multi-frame detection results to correct parameters such as the estimated value and trend of defect length, width, area, depth, and change, thereby improving the stability of defect size assessment and severity classification.

[0048] S6: The problem of difficulty in forming a closed loop for maintenance based on test results.

[0049] Existing detection methods often use raw videos, image screenshots, or simple alarms as results, lacking structured information such as defect number, defect image, defect category, axial position, circumferential position, dimensional parameters, confidence level, severity level, and maintenance recommendations. This results in a large amount of manual work for subsequent processing and is not conducive to defect review and maintenance decisions.

[0050] The present invention aims to solve how to generate a test report through the result output module and display the defect location on the catheter unfolding diagram or three-dimensional catheter model, so that the test results form a closed-loop data that is traceable, verifiable, and maintainable.

[0051] 3. Technical Solution

[0052] System technical solution

[0053] To address the aforementioned technical problems, this invention provides an AI visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits. The system includes an in-conduit visual acquisition device, a pose perception module, an image preprocessing module, a defect recognition module, a defect tracking and positioning module, and a result output module.

[0054] The in-pipe visual acquisition device is mounted on a support mechanism that can move along the interior of the industrial conduit, and is used to acquire images of the inner wall of the industrial conduit at a preset frame rate. The in-pipe visual acquisition device includes an industrial endoscope, a ring-shaped illumination assembly, and an image sensor. The ring-shaped illumination assembly is arranged around the image sensor to provide uniform illumination to the inner wall of the industrial conduit. The image sensor is used to acquire continuous images or video streams of the inner wall of the industrial conduit.

[0055] The pose sensing module is used to acquire the axial displacement, attitude angle, movement speed, and / or mileage information of the supporting mechanism within the industrial conduit. The pose sensing module includes an encoder, an inertial measurement unit, an odometer, a laser rangefinder, or a combination thereof. The pose information output by the pose sensing module establishes a correspondence with image frames acquired by an image sensor, used for cross-frame defect correlation, spatial position conversion, and dimensional parameter evaluation.

[0056] The image preprocessing module performs distortion correction, illumination equalization, image denoising, and inner wall unfolding mapping on the image of the inner wall of the industrial conduit to obtain a preprocessed image. The image preprocessing module includes a distortion correction unit, an illumination equalization unit, a denoising enhancement unit, and an unfolding mapping unit. The distortion correction unit corrects radial distortion of the image inside the conduit according to camera calibration parameters; the illumination equalization unit reduces the impact of reflections, shadows, and local overexposure on defect identification; the denoising enhancement unit enhances the texture features of cracks, corrosion pits, pinholes, and weld abnormalities; and the unfolding mapping unit converts the image of the inner wall of annular or arc-shaped conduit into a planar unfolded image.

[0057] The defect recognition module is used to input the preprocessed image into a deep learning defect recognition model trained on industrial conduit defect samples, and output the defect category, defect boundary region, and defect confidence score. The deep learning defect recognition model is a convolutional neural network model, an object detection model, a semantic segmentation model, an instance segmentation model, or a combination thereof. The defect categories identified by the deep learning defect recognition model include at least one or more of the following: cracks, corrosion, pinholes, pits, weld defects, coating peeling, and foreign matter adhesion. The deep learning defect recognition model is trained using normal sample images and defect sample images of the inner wall of the industrial conduit, and the defect sample images are labeled with the defect category, defect boundary, defect location, and defect severity.

[0058] The defect tracking and localization module is used to perform cross-frame correlation tracking of the same defect based on defect features in adjacent frame images, defect boundary regions, and pose information output by the pose perception module, and to calculate the axial position, circumferential position, and size parameters of the defect within the industrial conduit. The defect tracking and localization module includes a defect feature extraction unit, a cross-frame matching unit, and a position fusion unit. The defect feature extraction unit extracts shape features, texture features, grayscale features, and deep learning features of the defect region; the cross-frame matching unit determines whether defects in adjacent frames belong to the same defect based on defect feature similarity, defect region overlap, and pose change in adjacent frames; the position fusion unit fuses the detection results of the same defect in multiple frames to obtain the final position and size parameters of the defect.

[0059] The defect tracking and positioning module converts the defect's position in the image coordinate system into its axial and circumferential positions in the industrial conduit coordinate system based on the axial displacement and attitude angle. It then establishes a conversion relationship between pixel coordinates and the actual dimensions of the conduit based on camera calibration parameters, conduit inner diameter parameters, and image pixel dimensions to calculate one or more parameters, including estimated values ​​and trends of the defect's length, width, area, depth, and changes. The defect tracking and positioning module is also used to classify the severity of the defect based on a comparison of these parameters with preset thresholds.

[0060] The result output module is used to generate an inspection report based on the defect category, defect location, size parameters, and defect confidence level. The result output module includes a visualization display unit, a defect annotation unit, and a report generation unit. The visualization display unit is used to display the defect location on the duct unfolded diagram or 3D duct model; the defect annotation unit is used to mark the defect boundaries, category, and confidence level in the image; the report generation unit is used to generate an inspection report containing the defect number, defect image, defect category, axial position, circumferential position, size parameters, confidence level, severity level, and maintenance recommendations.

[0061] Methods and technical solutions

[0062] The present invention also provides an AI visual detection and tracking method for microscopic defects in the inner wall of industrial conduits, the method comprising the following steps.

[0063] S1: Control the load-bearing mechanism to move along the inside of the industrial duct and acquire images of the inner wall of the industrial duct through the in-duct vision acquisition device.

[0064] S2: Acquire information on the axial displacement, attitude angle, movement speed, and / or mileage of the load-bearing mechanism within the industrial conduit.

[0065] S3: Perform distortion correction, illumination equalization, noise reduction and enhancement, and inner wall unwrapping mapping on the acquired images of the inner wall of the industrial conduit to obtain a preprocessed image.

[0066] S4: Input the preprocessed image into the trained deep learning defect recognition model and output the defect category, defect boundary region and defect confidence level.

[0067] S5: Based on the defect features, defect boundary regions, and pose information of the supporting mechanism in adjacent frame images, perform cross-frame correlation tracking of the same defect.

[0068] S6: Calculate the axial position, circumferential position, and dimensional parameters of the defect within the industrial conduit based on the cross-frame correlation tracking results.

[0069] S7: Generate defect detection results and inspection reports based on defect type, defect location, size parameters, and defect confidence level.

[0070] Before step S4, a step of establishing a defect sample dataset is also included. The defect sample dataset includes normal sample images and defect sample images of the inner wall of industrial ducts; the defect sample images are labeled with defect category, defect boundary, defect location and defect severity; the deep learning defect recognition model is trained, validated and optimized using the defect sample dataset.

[0071] In step S5, cross-frame correlation tracking of the same defect includes: extracting the shape features, texture features, and deep learning features of the defect region in the current frame; calculating the feature similarity between the defect region in the current frame and the defect region in the previous frame; predicting the theoretical position of the defect in the previous frame in the current frame by combining the axial displacement and attitude change of the bearing mechanism; and determining whether the defect in the current frame and the defect in the previous frame are the same defect based on the feature similarity, theoretical position deviation, and defect region overlap.

[0072] In step S6, a conversion relationship between pixel coordinates and the actual size of the conduit is established based on the camera calibration parameters, the inner diameter parameters of the conduit, and the image pixel size. The actual length, actual width, and actual area of ​​the defect are calculated based on the conversion relationship. Then, the defect size parameters are corrected based on the fusion value of the multi-frame detection results.

[0073] Electronic device technical solutions

[0074] The present invention also provides an electronic device, the electronic device including a processor, a memory, and a computer program stored in the memory and executable by the processor.

[0075] When the processor executes the computer program, it implements the AI ​​visual detection and tracking method for microscopic defects in the inner wall of industrial conduits as described in this invention.

[0076] Computer-readable storage media technology solutions

[0077] The present invention also provides a computer-readable storage medium in which a computer program is stored.

[0078] When the computer program is executed by the processor, it implements the AI ​​visual detection and tracking method for microscopic defects in the inner wall of industrial conduits as described in this invention.

[0079] 4. Beneficial effects

[0080] Compared with the prior art, the present invention has at least the following beneficial effects.

[0081] First, the present invention uses an in-pipe vision acquisition device to continuously acquire images of the inner wall of industrial conduits, transforming the detection of microscopic defects inside the conduits from manual observation of single-point images to continuous image sequence detection, thereby improving the integrity of the detection data.

[0082] Second, the present invention synchronously acquires the axial displacement, attitude angle, movement speed and / or mileage information of the bearing mechanism in the industrial conduit through the pose perception module, so as to establish a correspondence between the image frame and the actual spatial position of the conduit, providing a data basis for defect cross-frame tracking and spatial positioning.

[0083] Third, this invention reduces the impact of curved surface imaging, reflection, shadow, local overexposure, radial distortion and noise on defect identification of industrial duct inner walls by distortion correction, illumination equalization, image denoising and inner wall unfolding mapping, thereby improving the stability and recognizability of preprocessed images.

[0084] Fourth, the present invention outputs defect category, defect boundary region and defect confidence level through a deep learning defect recognition model trained on industrial conduit defect samples, which can meet the needs of identifying various types of micro-defects such as cracks, corrosion, pinholes, pits, weld defects, coating peeling and foreign matter adhesion.

[0085] Fifth, this invention performs cross-frame correlation tracking of the same defect by using defect feature similarity, defect region overlap, and pose change in adjacent frames. This can reduce repeated statistics of the same defect, reduce the risk of tracking interruption caused by anomalies in a single frame image, and improve the continuity of defect detection results.

[0086] Sixth, this invention transforms defects from the image coordinate system to the industrial conduit coordinate system by using axial displacement, attitude angle, and conduit inner diameter parameters, thereby obtaining the axial and circumferential positions of the defects and providing a clear basis for re-inspection, repair, and replacement operations.

[0087] Seventh, this invention establishes the conversion relationship between pixel coordinates and the actual size of the conduit by using camera calibration parameters, conduit inner diameter parameters and image pixel size, and corrects the size parameters by multi-frame fusion, which can improve the reliability of the estimated values ​​of defect length, width, area and depth and the assessment of change trends.

[0088] Eighth, the present invention generates an inspection report through the result output module, which includes defect number, defect image, defect category, axial position, circumferential position, dimensional parameters, confidence level, severity level and maintenance suggestions, forming a complete inspection closed loop from collection, identification, tracking, positioning, evaluation to report output.

[0089] 5. Comparison of existing authorized patents

[0090] Comparison with CN113469177A

[0091] CN113469177A discloses a method and system for detecting defects in drainage pipes based on deep learning. The disclosed text describes the technical solution of collecting images of the inside of drainage pipes, filtering and preprocessing the images, using sinGAN to expand the dataset, labeling defect regions, constructing a drainage pipe defect recognition model, and optimizing the model output. The legal status of the disclosed text of this application is Granted, and it was authorized on April 26, 2024.

[0092] The difference between this invention and CN113469177A is at least in the following aspects:

[0093] CN113469177A mainly focuses on the construction and training of a deep learning defect detection model for images inside drainage pipes. Its key features include image dataset expansion, multi-scale feature extraction, region generation, ROI alignment, and mask generation, in order to achieve the identification and localization of defects in drainage pipes.

[0094] This invention not only includes a deep learning defect recognition model 61, but also further defines the collaborative relationship between the in-pipe visual acquisition device 3, the pose perception module 4, the image preprocessing module 5, the defect tracking and localization module 7, and the result output module 8. This invention acquires the axial displacement, attitude angle, movement speed, and / or mileage information of the bearing mechanism 2 within the industrial conduit 1 through the pose perception module 4, and performs cross-frame correlation tracking based on defect features, defect boundary regions, and pose information in adjacent frames, further calculating the axial position, circumferential position, and dimensional parameters of the defect within the industrial conduit 1.

[0095] Therefore, compared with CN113469177A, the technical focus of this invention is not limited to the image defect recognition model itself, but forms a complete closed loop of "intra-pipe image acquisition - pose synchronization - image preprocessing - AI defect recognition - cross-frame tracking - duct coordinate positioning - size assessment - inspection report output".

[0096] Comparison with CN113284109A

[0097] CN113284109A discloses a method, apparatus, terminal device, and storage medium for pipeline defect identification. The disclosed text describes the process of acquiring depth images, color images, and spatial trajectory information of a pipeline, determining the three-dimensional point cloud data of the depth images, determining a first pipeline defect identification result based on the three-dimensional point cloud data and spatial trajectory information, determining a second pipeline defect identification result through a preset pipeline defect identification algorithm and the color image, and then determining the pipeline defect level based on the dual identification results. The legal status of this application is Granted, and it was authorized on August 18, 2023.

[0098] The difference between this invention and CN113284109A is at least in the following aspects:

[0099] CN113284109A focuses on using depth images, color images, and spatial trajectory information for dual pipeline defect identification, and determines the pipeline defect level based on the identification results.

[0100] This invention focuses on AI detection and tracking of microscopic defects on the inner wall of industrial conduits in a continuous visual image sequence. The invention explicitly sets up a distortion correction unit 51, an illumination equalization unit 52, a noise reduction and enhancement unit 53, and an unfolding mapping unit 54 to solve the problems of radial distortion, reflection, shadow, noise, and unfolding in the curved surface image of the inner wall of the industrial conduit 1. Furthermore, this invention performs cross-frame correlation tracking of the same defect through defect feature similarity, defect region overlap, and pose change between adjacent frames, and fuses the detection results from multiple frames through a position fusion unit 73.

[0101] Therefore, compared with CN113284109A, the technical focus of this invention is on the continuous tracking of microscopic defects in the inner wall of industrial conduits, axial and circumferential positioning, conversion of pixel coordinates to actual conduit dimensions, and structured output of inspection reports, rather than determining the defect level solely based on depth and color images.

[0102] Comparison with US20160139061A1

[0103] US20160139061A1 discloses a pipe inspection device, which is described in the publication text as a pipe inspection device. The corresponding patent application is US10697901B2. This application was published on May 19, 2016, and granted on June 30, 2020. Its current legal status is Active.

[0104] The present invention differs from US20160139061A1 at least in that:

[0105] US20160139061A1 mainly relates to the structural design of pipeline inspection devices. Its technical focus is on the housing, fixed camera, rotatable camera and lighting-related structures used for imaging inside pipelines to improve the observation capability of the internal surface of pipelines.

[0106] Although this invention can also acquire images of the inner wall of the industrial conduit 1 through the in-pipe visual acquisition device 3, the core technology of this invention is not simply a camera mounting structure or a rotating shooting structure, but rather it performs distortion correction, illumination equalization, image denoising, and inner wall unfolding mapping on the acquired images of the inner wall of the industrial conduit 1. It also identifies the defect category, defect boundary region, and defect confidence level through a deep learning defect recognition model 61, and combines the axial displacement, attitude angle, motion speed, and / or mileage information output by the pose perception module 4 to perform cross-frame correlation tracking, spatial positioning, and size parameter evaluation for the same defect.

[0107] Therefore, compared with US20160139061A1, the technical focus of this invention has been expanded from the structure of the pipeline observation device to computer vision recognition, continuous tracking, pipeline coordinate positioning, size fusion evaluation and structured report output of microscopic defects in the inner wall of industrial conduits, which can form a complete intelligent detection closed loop from image acquisition to defect management. Attached Figure Description

[0108] Figure 1 This is a block diagram of the overall structure of the AI ​​visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits according to the present invention.

[0109] Figure 2 This is a schematic diagram of the installation structure of the in-pipe visual acquisition device and the supporting mechanism of the present invention inside an industrial conduit.

[0110] Figure 3 This is a data processing flowchart for the image preprocessing module, defect identification module, defect tracking and localization module, and result output module of the present invention.

[0111] Figure 4 This is a schematic diagram illustrating the defect of the present invention in the conversion between the image coordinate system and the industrial conduit coordinate system.

[0112] Figure 5 This is a flowchart of the AI ​​visual detection and tracking method for microscopic defects in the inner wall of industrial conduits according to the present invention.

[0113] Figure 6 This is a schematic diagram of the structure of the electronic device of the present invention.

[0114] Explanation of reference numerals in the attached figures

[0115] 1. Industrial conduits

[0116] 2. Load-bearing mechanism;

[0117] 3. In-pipe visual acquisition device;

[0118] 31. Industrial endoscopes;

[0119] 32. Ring-shaped fill light assembly;

[0120] 33. Image sensor;

[0121] 4. Pose perception module;

[0122] 41. Encoder;

[0123] 42. Inertial Measurement Unit;

[0124] 43. Odometer;

[0125] 44. Laser rangefinder sensor;

[0126] 5. Image preprocessing module;

[0127] 51. Distortion correction unit;

[0128] 52. Illuminance equalization unit;

[0129] 53. Noise reduction and enhancement unit;

[0130] 54. Expand the mapping unit;

[0131] 6. Defect identification module;

[0132] 61. Deep learning defect identification model;

[0133] 7. Defect tracking and location module;

[0134] 71. Defect Feature Extraction Unit;

[0135] 72. Cross-frame matching unit;

[0136] 73. Position fusion unit;

[0137] 8. Result Output Module;

[0138] 81. Visualization display unit;

[0139] 82. Defect labeling unit;

[0140] 83. Report Generation Unit;

[0141] 9. Electronic equipment;

[0142] 91. Processor;

[0143] 92. Memory;

[0144] 93. Communication interface;

[0145] 94. Computer program. Detailed Implementation Plan

[0146] 1. General Description of Implementation Methods

[0147] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Those skilled in the art should understand that the following embodiments are only used to explain the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions, conventional modifications, or structural adjustments made to the present invention without departing from the technical concept of the present invention should all fall within the scope of protection of the present invention.

[0148] This invention provides an AI visual inspection and tracking system and method for microscopic defects in the inner wall of industrial conduits, which enables continuous image acquisition, image preprocessing, AI defect recognition, cross-frame correlation tracking, spatial positioning, size assessment, severity classification, and inspection report output for microscopic defects in the inner wall of industrial conduits.

[0149] like Figure 1 As shown, the system of the present invention includes: an industrial conduit 1; a supporting mechanism 2; an in-pipe visual acquisition device 3; a pose perception module 4; an image preprocessing module 5; a defect identification module 6; a defect tracking and positioning module 7; a result output module 8; and an electronic device 9.

[0150] Among them, the in-pipe vision acquisition device 3 is used to acquire images of the inner wall of the industrial conduit 1; the pose perception module 4 is used to acquire pose data of the bearing mechanism 2 during its movement inside the industrial conduit 1; the image preprocessing module 5 is used to improve image quality; the defect recognition module 6 is used to identify defect categories and defect boundaries; the defect tracking and positioning module 7 is used to achieve stable association and spatial positioning of the same defect in continuous images; and the result output module 8 is used to generate structured inspection results.

[0151] This invention forms the following complete data processing closed loop: image acquisition → pose synchronization → image preprocessing → AI defect identification → cross-frame defect correlation → conduit coordinate transformation → size parameter calculation → defect level assessment → inspection report output.

[0152] 2. Implementation methods for industrial conduits and load-bearing mechanisms

[0153] Industrial conduits

[0154] like Figure 2 As shown, industrial conduit 1 is the object to be tested, which can be a metal conduit, composite material conduit, plastic conduit, corrosion-resistant conduit, welded conduit, or coated conduit.

[0155] Industrial conduit 1 can be applied to industrial scenarios such as petrochemicals, power energy, metallurgical manufacturing, pharmaceutical equipment, food processing, aerospace, and industrial equipment. The interior of industrial conduit 1 may contain microscopic defects such as cracks, corrosion, pinholes, pits, abnormal welds, coating peeling, and foreign matter adhesion.

[0156] Bearing mechanism

[0157] The support mechanism 2 is located inside the industrial conduit 1 and can move along the axial direction of the industrial conduit 1.

[0158] The load-bearing mechanism 2 can adopt one or more of the following: wheeled structure; tracked structure; flexible propulsion structure; cable traction structure; pipeline robot structure; and manual propulsion structure.

[0159] In this embodiment, the supporting mechanism 2 is provided with a central support structure, so that the visual acquisition device 3 inside the tube maintains a stable observation distance inside the duct.

[0160] The moving speed of the bearing mechanism 2 is set according to: image acquisition frame rate; conduit inner diameter; defect detection accuracy; and image clarity.

[0161] 3. Implementation method of visual acquisition within the pipe

[0162] In-tube visual acquisition device

[0163] like Figure 2As shown, the in-duct visual acquisition device 3 includes: an industrial endoscope 31; a ring-shaped supplementary lighting assembly 32; and an image sensor 33. The industrial endoscope 31 is used to acquire images of the inner wall of the industrial conduit 1.

[0164] The industrial endoscope 31 can be equipped with one of the following: a wide-angle lens; a macro lens; a high-temperature resistant lens; or a waterproof and dustproof lens.

[0165] The ring-shaped lighting assembly 32 is arranged around the image sensor 33 to provide uniform illumination to the inner wall of the industrial duct 1.

[0166] The ring light supplement component 32 can be one of the following: LED ring light source; fiber optic light supplement structure; or multi-point uniform light supplement structure.

[0167] Image sensor 33 is used to acquire continuous images or video streams of the inner wall of industrial conduit 1 at a preset frame rate.

[0168] In this embodiment, the image acquisition frame rate is 10fps to 120fps.

[0169] Image synchronization mechanism

[0170] During the detection process, each frame of image acquired by image sensor 33 corresponds to: frame number; timestamp; pose number, thus forming a one-to-one correspondence between image frames and pose data.

[0171] This synchronization relationship is used for subsequent tasks: cross-frame defect correlation; duct coordinate localization; dimensional parameter calculation; and multi-frame fusion correction.

[0172] 4. Pose Awareness Implementation Method

[0173] Composition of the pose perception module

[0174] like Figure 3 As shown, the pose sensing module 4 includes: an encoder 41; an inertial measurement unit 42; an odometer 43; and a laser rangefinder 44.

[0175] The encoder 41 is used to obtain the axial movement distance of the bearing mechanism 2.

[0176] The inertial measurement unit 42 is used to obtain the attitude angle of the bearing mechanism 2.

[0177] The attitude angles include one or more of the following: pitch angle, roll angle, and yaw angle.

[0178] The odometer 43 is used to record the cumulative distance traveled.

[0179] The laser rangefinder 44 is used to obtain the distance between the bearing mechanism 2 and the reference position.

[0180] Pose synchronization processing

[0181] The pose information output by the pose perception module 4 is synchronized with the image frames acquired by the image sensor 33 in time.

[0182] Synchronized data formation: image frame-pose data set.

[0183] This dataset is used for: theoretical location prediction; defect cross-frame matching; duct coordinate transformation; and size fusion correction.

[0184] This invention uses a combination of pose information and defect region matching constraints to stably associate the same defect in consecutive images, reducing the defect number drift problem caused by misidentification in a single frame.

[0185] 5. Image Preprocessing Implementation Methods

[0186] Image preprocessing module

[0187] like Figure 3 As shown, the image preprocessing module 5 includes: a distortion correction unit 51; an illumination equalization unit 52; a noise reduction and enhancement unit 53; and an expansion mapping unit 54.

[0188] Image preprocessing module 5 processes the original image in the following order: distortion correction → illumination equalization → noise reduction and enhancement → unpacking and mapping.

[0189] Distortion correction

[0190] The distortion correction unit 51 performs radial distortion correction on the image of the inner wall of the industrial conduit 1 according to the camera calibration parameters.

[0191] The camera calibration parameters include: focal length; principal point coordinates; radial distortion coefficient; tangential distortion coefficient; and image resolution parameters.

[0192] After distortion correction, the defect boundary is more consistent with the actual boundary.

[0193] Illumination balance

[0194] The illuminance equalization unit 52 is used to reduce: reflections; shadows; local overexposure; uneven lighting; and the impact on defect identification.

[0195] The illuminance equalization unit 52 can employ one or more of the following: luminance normalization; local contrast enhancement; background luminance compensation; and adaptive luminance correction.

[0196] Noise reduction and enhancement

[0197] The noise reduction and enhancement unit 53 is used to reduce: random noise; motion blur; compression artifacts; and contaminant interference.

[0198] It also enhances the texture features of cracks, corrosion pits, pinholes, and weld abnormalities.

[0199] Expand mapping

[0200] The unfolding mapping unit 54 is used to map the curved surface image of the inner wall of the industrial duct 1 into a planar unfolded image.

[0201] After the mapping is expanded: the horizontal coordinate corresponds to the circumferential position; the vertical coordinate corresponds to the axial position.

[0202] The unfolded image is used for subsequent continuous defect tracking and size calculation.

[0203] 6. Defect Identification Implementation Methods

[0204] Defect identification module

[0205] like Figure 3 As shown, the defect identification module 6 includes a deep learning defect identification model 61.

[0206] The deep learning defect recognition model 61 receives a preprocessed image and outputs: defect category; defect boundary region; defect confidence level.

[0207] Defect identification model

[0208] The deep learning defect recognition model 61 can be one or more of the following: convolutional neural network model; object detection model; semantic segmentation model; instance segmentation model; multi-model fusion structure.

[0209] The defect categories include: cracks; corrosion; pinholes; pits; weld defects; coating peeling; foreign matter adhesion.

[0210] Model Training Implementation Methods

[0211] Establish a dataset of defects in the inner wall of an industrial conduit.

[0212] The dataset includes: normal sample images; defective sample images.

[0213] The defect sample images are labeled with: defect category; defect boundary; defect location; and defect severity.

[0214] During training, the samples are subjected to the following enhancements: rotation enhancement, brightness enhancement, noise enhancement, blur enhancement, and scale enhancement, in order to improve the model's generalization ability.

[0215] 7. Defect Tracking and Location Implementation Methods

[0216] Defect tracking and location module

[0217] like Figure 3 As shown, the defect tracking and localization module 7 includes: a defect feature extraction unit 71; a cross-frame matching unit 72; and a location fusion unit 73.

[0218] Defect Feature Extraction

[0219] The defect feature extraction unit 71 is used to extract: shape features; texture features; grayscale features; and deep learning features. Shape features include: length; width; area; aspect ratio; and boundary curvature.

[0220] Texture features include: grayscale gradient; edge density; and local texture variations.

[0221] Defect cross-frame correlation

[0222] The cross-frame matching unit 72 is used to determine whether the defects in the current frame and the defects in the historical frames belong to the same defect object based on: defect feature similarity; defect region overlap; and pose change in adjacent frames.

[0223] In one implementation:

[0224] When the feature similarity is higher than a preset threshold and the positional deviation is lower than a preset threshold, the defects in the current frame and the defects in the historical frames are associated with the same defect number.

[0225] The above methods can reduce the repeated counting of the same defect.

[0226] Multi-frame position fusion

[0227] The position fusion unit 73 is based on: defect confidence; image sharpness; pose stability;

[0228] Defect boundary integrity; fusion of multi-frame detection results for the same defect object.

[0229] After fusion, we obtain: final axial position; final circumferential position; final dimensional parameters.

[0230] 8. Implementation method for catheter coordinate transformation

[0231] coordinate transformation relationship

[0232] like Figure 4 As shown, this invention establishes the correspondence between the image coordinate system and the industrial conduit coordinate system.

[0233] Where: the horizontal coordinate of the image corresponds to the circumferential position; the vertical coordinate of the image corresponds to the axial position.

[0234] Axial position calculation

[0235] The axial position of the defect is calculated using the following formula: ,in:

[0236] Z represents the axial position of the defect; Li represents the cumulative mileage of the current frame; kz represents the vertical pixel conversion ratio. This represents the vertical pixel offset.

[0237] Circumferential position calculation

[0238] The circumferential location of the defect is calculated using the following formula: ,in: π is the circumferential angle; u is the horizontal pixel coordinate; W is the width of the unfolded image; This is the attitude correction angle.

[0239] Dimensional parameter calculation

[0240] The actual circumferential width of the defect is calculated using the following formula: ,in:

[0241] C represents the actual width of the defect; D represents the inner diameter of the conduit. W represents the horizontal pixel width of the defect; W represents the width of the expanded image.

[0242] The estimated values ​​for defect length, area, and depth are converted in the same way.

[0243] 9. Implementation Method for Defect Level Assessment

[0244] The defect tracking and location module 7 is based on: defect type; defect length; defect width; defect area; depth estimate; and trend parameters.

[0245] The severity of defects is classified.

[0246] The severity levels include: minor; moderate; severe; critical.

[0247] Repair suggestions are automatically generated based on the severity of the problem.

[0248] 10. Result Output Implementation Method

[0249] Result Output Module

[0250] like Figure 3 As shown, the result output module 8 includes: a visualization display unit 81; a defect annotation unit 82; and a report generation unit 83.

[0251] Visualization

[0252] The visualization display unit 81 is used to display the location of defects on: the catheter unfolding diagram; and the three-dimensional catheter model.

[0253] Defect labeling

[0254] Defect labeling unit 82 is used to label: defect boundary; defect category; defect confidence level; defect number; defect grade.

[0255] Test report generation

[0256] The report generation unit 83 is used to generate test reports.

[0257] The inspection report includes: defect number; defect image; defect category; axial position; circumferential position; dimensional parameters; defect confidence level; severity level; and repair recommendations.

[0258] The test report can be output as: PDF document; database record; structured table; or visual interface.

[0259] 11. Method Implementation

[0260] like Figure 5 As shown, the present invention also provides an AI visual detection and tracking method for microscopic defects in the inner wall of industrial conduits.

[0261] The method includes:

[0262] S1: Acquire images of the inner wall of industrial conduits;

[0263] S2: Obtain pose information;

[0264] S3: Perform image preprocessing;

[0265] S4: Perform AI defect identification;

[0266] S5: Perform cross-frame correlation tracking;

[0267] S6: Perform catheter coordinate transformation;

[0268] S7: Generate a test report.

[0269] 12. Implementation methods of electronic devices

[0270] like Figure 6 As shown, the electronic device 9 includes: a processor 91; a memory 92; a communication interface 93; and a computer program 94.

[0271] When processor 91 executes computer program 94 in memory 92, it implements the AI ​​visual detection and tracking method for microscopic defects in the inner wall of industrial conduits as described in this invention.

[0272] Electronic device 9 can be deployed in one or more of the following: edge computing terminals; mobile inspection robots; online inspection equipment; industrial inspection hosts; and offline analysis platforms.

[0273] 13. Implementation Results Description

[0274] Compared with existing technologies, this invention achieves automated continuous detection of microscopic defects on the inner wall of industrial conduits through: pose synchronization; surface unfolding; AI defect recognition; cross-frame defect tracking; conduit coordinate positioning; and multi-frame size fusion.

[0275] This invention can improve: defect identification accuracy; continuous defect tracking stability; defect location accuracy; defect size assessment reliability; and the degree of structure in inspection reports.

Claims

1. An AI visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits, characterized in that, include: The system comprises an in-pipe visual acquisition device, a pose perception module, an image preprocessing module, a defect recognition module, a defect tracking and positioning module, and a result output module. The in-pipe visual acquisition device is mounted on a support mechanism that can move along the interior of the industrial conduit and is used to acquire images of the inner wall of the industrial conduit at a preset frame rate. The pose perception module is used to acquire the axial displacement, attitude angle, movement speed, and / or mileage information of the support mechanism within the industrial conduit. The image preprocessing module is used to perform distortion correction, illumination equalization, image denoising, and inner wall unfolding mapping processing on the images of the inner wall of the industrial conduit to obtain a preprocessed image. The defect recognition module is used to input the preprocessed image into a deep learning defect recognition model trained with industrial conduit defect samples, and output the defect category, defect boundary region and defect confidence. The defect tracking and localization module is used to perform cross-frame correlation tracking of the same defect based on defect features, defect boundary regions, and pose information output by the pose perception module in adjacent frame images, and to calculate the axial position, circumferential position, and size parameters of the defect within the industrial conduit; wherein, the cross-frame correlation tracking includes at least defect feature extraction, theoretical position prediction, and defect region matching; the result output module is used to generate a detection report based on the defect category, defect location, size parameters, and defect confidence level.

2. The AI ​​visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits according to claim 1, characterized in that: The in-pipe visual acquisition device includes an industrial endoscope, a ring-shaped illumination assembly, and an image sensor; the ring-shaped illumination assembly is arranged around the image sensor and is used to provide uniform illumination to the inner wall of the industrial conduit; the image sensor is used to acquire continuous images or video streams of the inner wall of the industrial conduit; the pose sensing module includes one or more of an encoder, an inertial measurement unit, an odometer, and a laser rangefinder.

3. The AI ​​visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits according to claim 1, characterized in that: The image preprocessing module includes a distortion correction unit, an illumination equalization unit, a noise reduction and enhancement unit, and a unfolding and mapping unit. The distortion correction unit is used to correct the radial distortion of the image inside the pipe according to the camera calibration parameters. The illumination equalization unit is used to reduce the impact of reflection, shadows, and local overexposure on defect identification. The noise reduction and enhancement unit is used to enhance the texture features of cracks, corrosion pits, pinholes, and abnormal weld areas. The unfolding and mapping unit is used to convert the curved surface image of the inner wall of the industrial conduit into a planar unfolded image.

4. The AI ​​visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits according to claim 1, characterized in that: The deep learning defect recognition model is a convolutional neural network model, an object detection model, a semantic segmentation model, an instance segmentation model, or a combination thereof; The deep learning defect recognition model is used to identify one or more defects, including cracks, corrosion, pinholes, pits, weld defects, coating peeling, and foreign matter adhesion.

5. The AI ​​visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits according to claim 1, characterized in that: The defect tracking and localization module includes a defect feature extraction unit, a cross-frame matching unit, and a position fusion unit. The cross-frame matching unit is used to determine whether the defect in the current frame and the defect in the historical frame belong to the same defect object based on the similarity of defect features, the overlap of defect regions, and the change in pose between adjacent frames. The position fusion unit is used to fuse the detection results of the same defect object in multiple frames to obtain the final position and size parameters of the defect.

6. The AI ​​visual inspection and tracking system for microscopic defects in the inner wall of industrial conduits according to claim 1, characterized in that: The defect tracking and positioning module converts the position of the defect in the image coordinate system into the axial and circumferential positions in the industrial conduit coordinate system based on the axial displacement and attitude angle; and establishes the conversion relationship between pixel coordinates and the actual size of the conduit based on the camera calibration parameters, conduit inner diameter parameters and image pixel size, so as to calculate the estimated values ​​of the defect's length, width, area, depth and the trend parameters of change. The defect tracking and location module is also used to classify the severity of defects based on the comparison results between the parameters and preset thresholds.

7. An AI visual detection and tracking method for microscopic defects in the inner wall of industrial conduits, characterized in that, Includes the following steps: S1: Control the carrier mechanism to move along the inside of the industrial conduit and acquire images of the inner wall of the industrial conduit through the vision acquisition device inside the conduit; S2: Obtain the axial displacement, attitude angle, movement speed and / or mileage information of the carrier mechanism inside the industrial conduit; S3: Perform distortion correction, illumination equalization, noise reduction and enhancement and inner wall unfolding mapping on the acquired images of the inner wall of the industrial conduit to obtain a preprocessed image. S4: Input the preprocessed image into the trained deep learning defect recognition model and output the defect category, defect boundary region and defect confidence level; S5: Based on the defect features, defect boundary regions, and pose information of the supporting mechanism in adjacent frame images, perform cross-frame correlation tracking for the same defect; S6: Calculate the axial position, circumferential position, and dimensional parameters of the defect within the industrial conduit based on the cross-frame correlation tracking results; S7: Generate defect detection results and inspection reports based on defect type, defect location, size parameters, and defect confidence level.

8. The AI ​​visual detection and tracking method for microscopic defects in the inner wall of industrial conduits according to claim 7, characterized in that: Before step S4, the method further includes establishing a defect sample dataset; the defect sample dataset includes normal sample images and defect sample images of the inner wall of industrial ducts; The defect sample images are labeled with defect category, defect boundary, defect location and defect severity; the deep learning defect recognition model is trained, validated and optimized using the defect sample dataset.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the AI ​​visual detection and tracking method for microscopic defects in the inner wall of industrial conduits as described in claim 7 or 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program; when the computer program is executed by a processor, it implements the AI ​​visual detection and tracking method for microscopic defects in the inner wall of industrial conduits as described in claim 7 or 8.

Images